Data-Driven Decision Making for Enhancing Small-Signal Stability in Hybrid AC/DC Grids Through Converter Control Role Assignment
Extracted source text · arXiv 2503.05386
\documentclass[lettersize,journal]{IEEEtran}
% \usepackage{amsmath,amsfonts}
% \usepackage{algorithmic}
% \usepackage{algorithm}
% \usepackage{array}
% \usepackage{textcomp}
% \usepackage{stfloats}
% \usepackage{url}
% \usepackage{verbatim}
% \usepackage{graphicx}
% \usepackage{cite}
% \hyphenation{op-tical net-works semi-conduc-tor IEEE-Xplore}
% % updated with editorial comments 8/9/2021
\pdfoutput=1
\usepackage{graphicx}
\usepackage{subfiles}
\usepackage[caption=false,font=footnotesize,labelfont=sf,textfont=sf]{subfig}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage[thmmarks,amsmath]{ntheorem} % for 'aligned' env.
\usepackage{blkarray}
\usepackage{mathtools}
\usepackage{xcolor}
\usepackage{colortbl}
\usepackage{tikz}
\usepackage{tabularx}
\usepackage[caption=false]{subfig}
%\usepackage{subcaption} % For subfigures
%\usepackage[table,xcdraw]{xcolor}
%\usepackage{algorithm2e}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{multirow}
\usepackage{booktabs}
\usepackage[ruled,vlined]{algorithm2e}
% \usepackage[linesnumbered, ruled]{algorithm2e}
\SetAlgoNlRelativeSize{-1} % Adjust line number size if needed
\setlength{\algomargin}{1em} % Adjust margin within the algorithm
\usepackage{algpseudocode}
\newtheorem{definition}{Definition}
\newcommand\Algphase[1]{%
\vspace*{-1.4\baselineskip}\Statex\hspace*{0.5\dimexpr-\algorithmicindent}\rule{\linewidth}{0.1pt}
\Statex\hspace*{-0.5\dimexpr-\algorithmicindent}\textbf{#1}%
\vspace*{-1.4\baselineskip}\Statex\hspace*{0.5\dimexpr-\algorithmicindent}\rule{\linewidth}{0.1pt}%
}
\usepackage{nomencl}
\makenomenclature
%% This code creates the groups
% -----------------------------------------
\usepackage{etoolbox}
\renewcommand\nomgroup[1]{%
\item[\bfseries
\ifstrequal{#1}{A}{Acronyms}{%
\ifstrequal{#1}{S}{Symbols}}%
]}
\begin{document}
\title{\vspace{-0.6cm}
Data-Driven Decision Making for Enhancing Small-Signal Stability in Hybrid AC/DC Grids Through Converter Control Role Assignment}
\author{Francesca Rossi, Sergi Costa Dilmé, Josep Arévalo-Soler, Eduardo Prieto-Araujo,~\IEEEmembership{Senior Member,~IEEE},
and Oriol Gomis-Bellmunt,~\IEEEmembership{Fellow,~IEEE}
% <-this % stops a space
% \thanks{This paper was produced by the IEEE Publication Technology Group. They are in Piscataway, NJ.}% <-this % stops a space
% \thanks{Manuscript received April 19, 2021; revised August 16, 2021.}
\vspace{-1.1cm}}
% The paper headers
% \markboth{Journal of \LaTeX\ Class Files,~Vol.~14, No.~8, August~2021}%
% {Shell \MakeLowercase{\textit{et al.}}: A Sample Article Using IEEEtran.cls for IEEE Journals}
% \IEEEpubid{0000--0000/00\$00.00~\copyright~2021 IEEE}
% Remember, if you use this you must call \IEEEpubidadjcol in the second
% column for its text to clear the IEEEpubid mark.
\maketitle
\begin{abstract}
Hybrid AC/DC transmission grids incorporate Modular Multilevel Converters functioning as Interconnecting Power Converters (IPCs). The control role assigned to each converter significantly influences grid dynamics. Traditionally, these converters operate with static control roles, but recent studies have proposed scheduling their roles based on day-ahead forecasts to enhance stability performance. However, in systems with high renewable energy penetration, forecast deviations can render scheduled control assignments suboptimal or even lead to instability. To address this challenge, this work proposes an online scheduling recalculation algorithm that dynamically adapts IPC control roles during system operation. The approach leverages a data-driven multi-criteria decision-making framework, integrating surrogate models of conventional small-signal stability analysis tools to enable a fast computation of system stability and stability performance indicators.
\end{abstract}
\begin{IEEEkeywords}
Small-signal stability, Machine Learning, HVDC, Control Role Assignment
\vspace{-0.6cm}
\end{IEEEkeywords}
\section{Introduction}
\vspace{-0.1cm}
\input{Intro_red}
\vspace{-0.5cm}
\section{Small-signal Stability and Small-signal Stability Performance Indicators}
\label{sec: ss_ind}
\input{SS_ind_red}
%\section{Data-driven Multi-criteria Decision-making Algorithm for Converters Control Role Assignment}
\vspace{-0.7cm}
\section{Data-driven MCDM for CCR Assignment}
\label{sec: MODM}
\vspace{-0.1cm}
\input{Section3}
\vspace{-0.5cm}
\section{Training The Data-driven Surrogate Models}
%\section{Data-driven Surrogate Models for Small-signal Stability and Stability Performances Estimation}
\vspace{-0.1cm}
\input{Section4}
\vspace{-0.8cm}
\section{Case Study}
\input{Case_study}
\vspace{-0.2cm}
\section{Conclusion}
This paper presents a tool for supporting System Operators in enhancing small-signal stability performance of hybrid AC/DC grids, by performing an online IPC control role assignment. The tool is based on a data-driven MCDM algorithm. A comprehensive methodology for training accurate data-driven surrogate models for stability assessment and stability indicators calculation is presented, which does not require sharing the IPC control algorithms but only electrical schemes. A strategy to enhance the scalability of the data-driven MCDM is introduced, able to minimize the number of CRCs required to operate the system with good dynamic performance. The proposed models demonstrate high accuracy and the implementation of the data-driven MCDM shows high computational efficiency.
\vspace{-0.2cm}
\section*{Acknowledgments}
This work was supported by FEDER/Ministerio de Ciencia, Innovación y Universidades - Agencia Estatal de Investigación, Project PDC2022-133226-I00 Total ACDC
%{\appendices
%\section*{Proof of the First Zonklar Equation}
%Appendix one text goes here.
% You can choose not to have a title for an appendix if you want by leaving the argument blank
%\section*{Proof of the Second Zonklar Equation}
%Appendix two text goes here.}
\bibliographystyle{IEEEtran}
%\bibliography{bibliography}
\input{main.bbl}
\newpage
% \section{Biography Section}
% If you have an EPS/PDF photo (graphicx package needed), extra braces are
% needed around the contents of the optional argument to biography to prevent
% the LaTeX parser from getting confused when it sees the complicated
% $\backslash${\tt{includegraphics}} command within an optional argument. (You can create
% your own custom macro containing the $\backslash${\tt{includegraphics}} command to make things
% simpler here.)
% \vspace{11pt}
% \bf{If you include a photo:}\vspace{-33pt}
% \begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{fig1}}]{Michael Shell}
% Use $\backslash${\tt{begin\{IEEEbiography\}}} and then for the 1st argument use $\backslash${\tt{includegraphics}} to declare and link the author photo.
% Use the author name as the 3rd argument followed by the biography text.
% \end{IEEEbiography}
% \vspace{11pt}
% \bf{If you will not include a photo:}\vspace{-33pt}
% \begin{IEEEbiographynophoto}{John Doe}
% Use $\backslash${\tt{begin\{IEEEbiographynophoto\}}} and the author name as the argument followed by the biography text.
% \end{IEEEbiographynophoto}
% \vfill
\end{document}
%\IEEEPARstart{T}{ransmission}
Transmission
systems are increasingly incorporating High-Voltage Direct-Current (HVDC) technologies, which play a crucial role in integrating offshore wind power plants and enabling the efficient transmission of electricity over long distances~\cite{ramosready4dc}. Numerous projects %for new installations
are currently being developed, encompassing both point-to-point and multi-terminal connections %configurations
~\cite{ramosready4dc}. %A key technology enabling these interconnections is Interconnecting Power Converters (IPCs).
Nowadays, %Voltage Source Converters based on
Modular Multilevel Converter (MMC) technology has become the dominant solution for the Interconnecting Power Converters (IPCs). A key feature of MMCs is the ability to control voltage at both the AC and DC terminals. Therefore, in addition to the conventional grid-following (GFL) control, %which relies on Phase-Locked Loop (PLL) synchronization with the AC grid,
MMCs can also operate in grid-forming (GFM) mode on either the AC or DC side~\cite{9729620}. %This enables them to establish a stable voltage at one terminal while remaining in GFL mode at the other~\cite{9729620}.
GFL control has been observed to experience instability in weak grid conditions, whereas GFM control can improve stability and provide support for grid operation in such scenarios~\cite{9714816}. However, high penetration of GFM converters may introduce oscillations in strong grids or lead to undesirable interactions when multiple GFM converters are located too close electrically~\cite{9714816}. The emergence of these converter control-driven issues affecting system dynamics has increased interest in studying how to assign appropriate control roles to ensure and enhance grid stability. In~\cite{10508461}, the required level of GFM penetration to ensure small-signal stability and strengthen a weak grid with multiple wind farms is analyzed. The study assumes that wind turbines are connected via GFL converters, while a storage system interfaced through a GFM converter compensates for wind power fluctuations and enhances grid strength. However, a static control assignment approach is adopted, without exploring different operating conditions. %The study concludes that assigning all converters as GFM — a condition that may induce large transients under significant disturbances — is not necessary to maintain small-signal stability.
In~\cite{cui2024control}, the use of inverter-based resources (IBRs) for stability services is explored. The need for dedicated converters is bypassed by dynamically assigning control roles to the existing IBRs during operation. The optimal operating point (OP) and GFM allocation are obtained as solution of a Mixed Integer Linear Programming (MILP) Unit Commitment problem, considering frequency and small-signal stability constraints.\\
Focusing on hybrid AC/DC grids, the authors in~\cite{10017135,10443561} investigated the feasibility of applying control role scheduling to IPCs. As defined in~\cite{10443561}, the assignment of specific control roles to each IPC forms a Converter Control Role Configuration (CCRC). For a given system OP, %determined as the solution of a technical or economic optimization problem,
the system dynamics varies depending on the applied CCRC, ranging from instability to good dynamic performance. Therefore, the need for dynamic CCRC scheduling arises primarily from the fact that no single CCRC guarantees stability across all operating conditions~\cite{10443561}. Additionally, among stable CCRCs, it is possible to select the one that provides the best dynamic performance. In~\cite{10443561}, numerical indicators are formulated to evaluate both steady-state and dynamic small-signal stability performance. Based on these indicators, an MILP optimization problem is developed to dynamically schedule IPC control role assignments, ensuring that, for each system operating point, the selected CCRC maximizes stability performance.
The scheduling approaches presented in~\cite{cui2024control,10443561} have been designed for day-ahead operation and thus rely on accurate day-ahead forecasts of demand and generation. However, significant discrepancies between actual renewable generation or demand and the corresponding forecasts can make a predefined CCRC schedule suboptimal or even unstable for the actual OP. This highlights the need for a fast recalculation tool, capable of determining the most suitable CCRC assignment during online operation. To accelerate computation, this work proposes two key strategies: (i) replacing the MILP-based approach with a Multi-Criteria Decision-Making (MCDM) algorithm and (ii) substituting the mathematically exact small-signal stability models used for stability assessment and performance indicator calculation with their data-driven surrogate models. MILP is not considered a viable approach due to its significant computational requirements, which render it impractical for real-size problems~\cite{floudas2008encyclopedia}. In contrast, MCDM exhibits polynomial computational complexity, making it a more scalable and suitable alternative for solving large-size problems~\cite{floudas2008encyclopedia}. The use of surrogate models to reduce computational time is explored~\cite{BHOSEKAR2018250}. The stability assessment is performed by a classification algorithm and regressions are trained to estimate the stability indicators.
The main contributions of this paper are:
\begin{itemize}
\item A data-driven MCDM algorithm for online CCRC assignment to IPCs in hybrid AC/DC grids, designed to enhance small-signal stability performance.
\item A comprehensive methodology for training data-driven surrogate models for small-signal stability assessment and the estimation of stability performance indicators.
\item A methodology to enhance the scalability of the MCDM approach by reducing the number of CCRCs alternatives, retaining only those necessary for optimal stability performance. This selection process leverages data mining techniques to cluster CCRCs with similar stability behavior within the same operating region and applies set intersection methods to integrate information from different stability indicators to support the selection.
\end{itemize}
\subsection{Conventional Exact Models}
Small-signal stability assessment is conventionally based on the eigenvalues analysis of the linearized state-space of the system, evaluated around an equilibrium point \cite{kundur1994power}. Such equilibrium point can be obtained by power flow (PF) calculation or time-domain simulation. Following Lyapunov's first method for small-signal stability assessment, the system results asymptotically stable if all the eigenvalues ($\lambda$) have negative real parts \cite{kundur1994power}. A binary label $\Upsilon \in \{0,1\}$ can be used to indicate the outcome of the small-signal stability analysis:
\vspace{-0.2cm}
\begin{align}
\left\{\begin{aligned}
\Upsilon&=1 \quad if \max\{\Re(\lambda_1),...,\Re(\lambda_N)\}<0\\
\Upsilon&=0 \quad otherwise
\end{aligned}\right.
\label{eq: stab_lab}
\end{align}
\vspace{-0.5cm}
\begin{definition}{\bf{(Exact small-signal stability assessment)}}\label{def: exact_ss}
%For a certain OP and CCRC,
The \textit{exact} computation of the small-signal stability is indicated as
\vspace{-0.5cm}
\begin{equation}
\Psi: (x,u,y) \longrightarrow \Upsilon%^{OP,C}
\vspace{-0.2cm}
\end{equation}
\noindent where $x$, $u$, $y$ are the states, the inputs, and the outputs of the system state-space, respectively. It is computed for a certain OP and CCRC.
\end{definition}
\vspace{-0.2cm}
When the system is stable, its small-signal stability performance can be evaluated by analyzing the behavior of the system's transfer function matrix, \( G(s) \). This matrix, of dimension \( \mathbb{R}^{|y| \times |u|} \), is defined as $G(s) = \frac{y(s)}{u(s)}$, where \( y(s) \) and \( u(s) \) represent the system's outputs and inputs in the Laplace domain, respectively, and \( |y| \) and \( |u| \) denote the number of outputs and inputs. To capture the specific dynamics that it is intended to enhance, $G(s)$ has to be formulated by properly setting the inputs and outputs signals. Then, to evaluate the dynamic performances, one possibility %for the estimation of such performances
is to compute the $\mathcal{H}_2$-norm of the transfer functions matrix, $||G(s)||_2$. %signals. %relevant to the description of the dynamic behavior of the system.
%Given the selected MIMO system transfer functions matrix, obtained by considering the proper inputs and output signals,
The $\mathcal{H}_2$-norm provides a unique scalar value that is %of a system measures the area under the outputs evolution curves, which is
an estimation of the total output signals energy~\cite{skogestad2005multivariable}. The higher the norm value, the larger the sum of the signals energy content and the larger their oscillations before converging to the steady state. %Large oscillations are synonymous with poor performances.
%Given the proper matrix of transfer function $\mathbf{G(s)}$, formulated according to the selected inputs and outputs, %Therefore, the equations in (\ref{eq: linsyst}) are transformed into
% \begin{equation}
% \left\{ \begin{aligned}
% \Delta {x}(s) &= (s\textbf{I}-\textbf{A})^{-1}[\Delta x(0) + \textbf{B}\Delta u (s)] \\
% \Delta y(s) &= \textbf{C}(s\textbf{I}-\textbf{A})^{-1}[\Delta x(0)+ \textbf{B}\Delta u (s)] + \textbf{D}\Delta u(s) %\tag{4}
% \label{eq: linsyst_lap}
% \end{aligned} \right.
% \end{equation}
% \noindent
%Then, according to the signal that it is intended to analyze, i.e.
% by properly setting the input and output variables, the evolution of the selected signals can be studied. The corresponding transfer function $G(s)$ is obtained. as
% \begin{equation}
% G(s)=\frac{\Delta y(s)}{\Delta u(s)}=\mathbf{C}(s\textbf{I}-\textbf{A})^{-1}\mathbf{B}+\mathbf{D}
% \label{eq: trans_fun}
% \end{equation}
%the $\mathcal{H}_2$-norm is calculated as
% \begin{equation}
% \mathcal{H}_2=\sqrt{\frac{1}{2\pi}\int_{-\infty}^{\infty}tr[\mathbf{G(s)}^{T}\mathbf{G(s)}]ds}
% \label{eq: h2_norm}
% \end{equation}
\vspace{-0.2cm}
\begin{definition}{\bf{(Exact $\mathcal{H}_2$-norm calculation)}}\label{def: exact_h2}
% %For a certain OP and CCRC, for which the system small-signal response is stable,
The \textit{exact} computation of the $\mathcal{H}_2$-norm of the system output signals $y$ is indicated as
\vspace{-0.2cm}
\begin{equation} \label{eq: exact_H2}
\Gamma_{\mathcal{H}_{2,y}}: (x,u,y) \longrightarrow \mathcal{H}_{2,y}%^{OP,C}
\vspace{-0.2cm}
\end{equation}
%noindent obtained by solving equation considering the transfer function obtained from the system. % state-space. %The sub-index $y$ indicates that the norm is related to the output $y$, set in the system state-space representation.
%The norm can be computed for a certain OP and CCRC, for which the system small-signal response is stable.
\end{definition}
Another approach for formulating a small-signal stability performance indicator is to compute the \textit{DC Gain} (\(\mathcal{K}\)) of the transfer function matrix \(G(s)\). For each individual transfer function \( g(s) \in G(s) \), the DC Gain can be obtained from its canonical formulation:
% \vspace{-0.3cm}
% \begin{equation}
$g(s) = \frac{k \left(1+\sum_{i=1}^{m} b_i s^i \right)}{s^r \left(1+\sum_{j=1}^{l} a_j s^j \right)}$.
% \label{eq: canonic_g}
% \end{equation}
In this formulation, the coefficient \( k \) represents the \textit{DC Gain} of \( g(s) \). It can be computed as $k = \lim_{t \to \infty} y(t) = \lim_{s \to 0} s^r g(s).$
Therefore, this indicator measures the deviation of the output signal from its reference value once the system has reached a steady state. A higher \textit{DC Gain} value corresponds to a larger signal deviation, indicating poorer stability performance.
The \textit{DC Gain} of \( G(s) \) is then defined as the maximum \( k \) value among all transfer functions in \( G(s) \), representing the maximum deviation experienced by any output signal \( y \).
\vspace{-0.2cm}
\begin{definition}{\bf{(Exact $\mathcal{K}$ calculation)}}\label{def: exact_dcgain}
%For a certain OP and CCRC, for which the system small-signal response is stable,
The \textit{exact} computation of the DC gain $\mathcal{K}$ of the system output signals $y$ is indicated as
\vspace{-0.2cm}
\begin{equation} \label{eq: exact_DCgain}
\Gamma_{\mathcal{K}_{y}}: (x,u,y) \longrightarrow \mathcal{K}_{y}%^{OP,C}
\end{equation}
\end{definition}
\vspace{-0.2cm}
In summary, if the small-signal stability assessment states that the system response, for a certain OP and CCRC, is stable, the $\mathcal{H}_2$-norms and the \textit{DC Gain} of relevant signals can be calculated. Such quantities indicate stability performances. %represent indicators of stability performances.
\vspace{-0.7cm}
\subsection{Data-driven Surrogate Models}% for Small-signal Stability}
\vspace{-0.1cm}
The %conventional mathematical
tools %employed for the small-signal stability analysis. They
of Definitions \ref{def: exact_ss}-\ref{def: exact_dcgain} are referred to as \textit{exact} calculations as they are based on the mathematical %modeling of the system dynamics and the exact mathematical
formulation of the linear system state-space and system transfer functions. %small-signal stability analysis.
In this paper, it is proposed to train machine learning (ML) algorithms to serve as data-driven \textit{surrogate} for these \textit{exact} models. These \textit{surrogate} models aim to replicate the output of the \textit{exact} models, specifically the stability response and values of stability indicators, by utilizing classification and regression algorithms trained on suitable data. These data are generated by computing the \textit{exact} models across various OPs and CCRCs, as detailed in Section \ref{sec: datagen}. The computed instances and their corresponding outcomes are then organized into data sets, each tailored to a specific model for training. All data sets have the same structure: a single data set can be generically described as $\mathcal{D}=(X|z)$, where $X$ and $z$ are %, respectively,
the input and target variables of the model to be trained.
Concerning the data-driven \textit{surrogate} for the small-signal stability assessment, a unique model is trained over a single data set, collecting quantities and outcomes from all the computed instances, encompassing the application of all the CCRCs.
The output variable is $\Upsilon$. As it is a binary variable, the \textit{surrogate} model to be trained has to perform a classification.
The input quantities, $X$, relate to the OP and CCRC of the computed instances, as in the \textit{exact} models' calculations, though they do not match the \textit{exact} models' inputs precisely. In the \textit{exact} model calculation, comprehensive knowledge of the system is required, including a detailed formulation of the control schemes implemented and the precise tuning values of the control parameters. On the contrary, \textit{surrogate} models do not require information on control schemes or parameter tunning, only the electrical characteristics of the system. A detailed description of the input quantities is provided next.
\vspace{-0.1cm}
\begin{itemize}
\item The PF input quantities, $X_{PF}$. These quantities can be obtained through PF calculations based on the system's OP. Consider a generic power system and define $\mathcal{N}$ as the set of buses, $\mathcal{G}$ as the set of generators, $\mathcal{L}$ as the set of loads, and $\mathcal{C}$ as the set of IPCs. Moreover, consider $\mathcal{T}$ as the set collecting Thevenin's equivalents that might be used for modeling portions of the AC grids for simulation and computation purposes. Then, the PF quantities are
% \vspace{-0.2cm}
% \begin{multline}
$X_{PF}=[(V_i,\theta_i)_{\forall i\in\mathcal{N}},(P_i,Q_i)_{\forall i\in\mathcal{G}},(P_i,Q_i)_{\forall i\in\mathcal{C}},
(P_i,Q_i)_{\forall i\in\mathcal{L}},\\
(P_i,Q_i)_{\forall i\in\mathcal{T}}]$,
% \label{eq: X_PF}
%\end{multline}
%\noindent
where $V$ and $\theta$ are the voltage module and phase angle, $P$ and $Q$ are the active and reactive power.
\item The IPC internal measures input quantities, $X_{IPC}$. As identified in \cite{PRIETOARAUJO2017}, they are %$ \in \mathbb{R}^{1\times10(|\mathcal{C}|)}$
%\begin{multline}
$X_{IPC}=[(V_i^{AC},\theta_{V,i}^{AC},V_i^{diff},\theta_{V,i}^{diff},V_i^{sum},\theta_{V,i}^{sum}, I_i^{diff},\theta_{I,i}^{diff},\\I_i^{sum},\theta_{I,i}^{sum})_{\forall i\in\mathcal{C}}]$,
% \label{eq: X_IPC}
% \end{multline}
% Presentes en el dataset pero no mencionados, ya que ya los hemos puesto en PF: $e\theta$ (angulo lado AC), $V_{DC}$, P y Q
% which includes the $qd$ elements obtained from applying the DQZ transformation to the AC side $v$ and $i$,
which includes, for the $i$-th converter, the module of the three-phase voltage on the AC side ($V_i^{AC}$), the modules of the applied three-phase differential ($V_i^{diff}$) and additive ($V_i^{sum}$) voltages, the modules of the differential ($I_i^{diff}$) and additive ($I_i^{sum}$) currents. The quantities $\theta$s are the corresponding phase angles. These quantities can be calculated once the PF is solved and known the electrical %scheme and
modeling of the IPCs.
% the AC side $v$ and $i$ represented in the $qd0$-frame (obtained by the application of the product of the Park and Clarke transformations), and the differential and additive $v$ and inner $i$ through the IPC's upper and lower arms, as defined in \cite[Sec. II]{PRIETOARAUJO2017}.
\item The input quantities related to the CCRC assignment, $X_{C}$. Since variations in control parameter tuning for converters are not considered in this study — only changes in control roles are — CCRC is accounted for by encoding the control roles as categorical variables. The CCRC assignment is then expressed by a sequence of discrete variables, each identifying a converter control role (CCR), as follows %$ \in \mathbb{R}^{1\times|\mathcal{C}|}$
%\begin{equation}
$ X_{C}=[(CCR_i)_{\forall i\in\mathcal{C}}]$.
% \label{eq: X_C}
% \end{equation}
%\noindent where the CCR is expressed by a discrete variable:
%CCR=1 for AC-GFM, CCR=2 for DC-GFM, and CCR=3 for GFL.
%\textcolor{red}{CCR is a discrete value representing the operating mode of the converter.} %Therefore, each $X_{C}$ indicates a CCRC.
\end{itemize}
Summarizing the above mentioned consideration, the data-driven \textit{surrogate} model for a small-signal stability assessment is defined as follows.
\vspace{-0.1cm}
\begin{definition}\textbf{(Data-driven surrogate of the small-signal stability assessment)}\label{def: dd_stab} The data-driven \textit{surrogate} model performing the small-signal stability assessment is a classification algorithm trained over a data set $\mathcal{D}_\Upsilon=(X_{PF},X_{IPC},X_{C}|\Upsilon)$. It maps the small-signal stability response of the system for a range of OPs and CCRCs, as expressed by %. For a certain OP and CCRC, it can predict if the system response is stable or not as
\vspace{-0.2cm}
\begin{equation}
\hat{\Psi}: (X_{PF},X_{IPC},X_{C}) \longrightarrow \hat{\Upsilon}%^{OP,C}
\vspace{-0.2cm}
\end{equation}
Set a CCRC and for a certain OP, not necessarily belonging to the training data set $\mathcal{D}_\Upsilon$, $\hat{\Psi}$ calculates an estimation of the system stability, $\hat{\Upsilon}$.
\end{definition}
\vspace{-0.1cm}
Concerning the \textit{surrogate} models for the calculation of the stability performances indicators, a different approach is used. These indicators vary widely in value across different CCRCs, making a single model for all CCRCs less accurate. Therefore, for each CCRC, a separate data set is built and a separate model is trained. Hence, the variables describing the CCRCs, $X_C$, do not need to be included in the training data sets and as input of these models. In summary, for the data-driven \textit{surrogate} models of the stability performances indicators, the following definitions are provided.
\vspace{-0.2cm}
\begin{definition}\textbf{(Data-driven \textit{surrogate} of the $\mathcal{H}_2$-norm calculation)}\label{def: dd_h2} The data-driven \textit{surrogate} model performing the $\mathcal{H}_2$-norm calculation of system outputs $y$, for the $i$-th CCRC, is a regression algorithm trained over a data set $\mathcal{D}_{\mathcal{H}_{2,y}}^{i}=(X_{PF},X_{IPC}|\mathcal{H}_{2,y})$.
It fits, for the $i$-th CCRC, the value of the indicator for a range of OPs.
\vspace{-0.2cm}
\begin{equation}
\hat{\Gamma}_{\mathcal{H}_{2,y}}^{i}: (X_{PF},X_{IPC}) \longrightarrow \hat{\mathcal{H}}_{2,y}%^{OP,C}
\vspace{-0.2cm}
\end{equation}
Given an OP not necessarily belonging to %the training data set
$\mathcal{D}_{\mathcal{H}_{2,y}}^{i}$, $\hat{\Gamma}_{\mathcal{H}_{2,y}}^{i}$ calculates an estimation of the stability performance indicator $\hat{\mathcal{H}}_{2,y}$, considering the application of the $i$-th CCRC.
\end{definition}
\vspace{-0.3cm}
\begin{definition}\textbf{(Data-driven \textit{surrogate} of the $\mathcal{K}$ calculation)}
\label{def: dd_dcgain}
The data-driven \textit{surrogate} model performing the $\mathcal{K}$ calculation of the output signals $y$, for the $i$-th CCRC, is a regression algorithm trained over a data set $\mathcal{D}_{\mathcal{K}_y}^{i}=(X_{PF},X_{IPC}|\mathcal{K}_{y})$.
It fits, for the $i$-th CCRC, the value of the indicator for a range of OPs.
\vspace{-0.3cm}
\begin{equation}
\hat{\Gamma}_{\mathcal{K}_{y}}^{i}: (X_{PF},X_{IPC}) \longrightarrow \hat{\mathcal{K}}_{y}%^{OP,C}
\vspace{-0.2cm}
\end{equation}
Given an OP not necessarily belonging to %the training data set
$\mathcal{D}_{\mathcal{K}_y}^{i}$, $\hat{\Gamma}_{\mathcal{K}_{y}}$ calculates an estimation of the value of the stability performance indicator $\hat{\mathcal{K}}_{y}$, considering the application of the $i$-th CCRC.
\end{definition}
% This section introduces the proposed data-driven MCDM algorithm for assigning CCRCs during online power system operation. First, a general and conventional formulation of the MCDM problem is presented. Then, the adaptation of this formulation for the specific context of this work is described.
% \subsection{Conventional MCDM Algorithm}
% An MCDM problem can be expressed by the use of a \textit{performance} matrix \cite{triantaphyllou_2000}. Consider that the problem has $m$ alternative solutions (\textit{alternatives}, $\alpha$) and that the \textit{performances} ($\rho$) of such \textit{alternatives} are evaluated according to $n$ \textit{criteria} ($\zeta$). Then, the \textit{performance} matrix $\mathbf{\Pi} \in \mathbb{R}^{m \times n}$ can be built as
% \begin{equation}
% \mathbf{\Pi} =
% \begin{blockarray}{cccc}
% & \zeta_1 & \dots & \zeta_n \\
% %& (w_{\zeta_1} & \dots & w_{\zeta_n}) \\
% \begin{block}{c[ccc]}
% \alpha_1 & \rho_{11} & \dots & \rho_{1n} \\
% \vdots & \vdots & \ddots & \\
% \alpha_m & \rho_{m1} & & \rho_{mn} \\
% \end{block}
% \end{blockarray}
% \label{eq: perf_matrix_gen}
% \end{equation}
% \noindent The score of each $i$-th \textit{alternative} can be calculated, applying the weighted sum model, as
% \begin{equation}
% \Pi_{i,\text{score}} = \sum_{j=1}^n{w_{\zeta_j}\, \rho_{ij}} \quad \text{for } i=1,...,m
% %F = \min_\alpha \sum_{j=1}^n{w_{\zeta_j}\cdot p_{ij}} \quad i=1,...,m
% \label{eq: obj_fun_MCDM}
% \end{equation}
% \noindent where $w_{\zeta}$s are the weights associated to each criterion.\\
% The objective of the MCDM is to find the \textit{alternative} ($\alpha_{i^*}$) that returns, in the context of maximization, the highest score. Therefore, the solution of the MCDM is the $i^*$-th \textit{alternative}, where $i^*$ is the index of the alternative with the maximum score, obtained as: %the \textit{performances}. However, according to the problem stability indicators, the lower the indicators' value, the better performances are achieved. Therefore, the objective function is reformulated as the minimum weighted sum
% \begin{equation}
% i^*=\arg \max _i \Pi_{i,\text{score}} = \quad \text{for } i=1,...,m
% %F = \min_\alpha \sum_{j=1}^n{w_{\zeta_j}\cdot p_{ij}} \quad i=1,...,m
% \label{eq: obj_fun_MCDM}
% \end{equation}
% %Such a general formulation is applied to the context of the present study as follows.
% %Applying such notation to the context of the present study, the formulation shown next can be obtained. %The decision-making \textit{criteria} correspond to the small-signal stability performance indicators. If a set $\mathcal{Y}$ of system signals is considered in the analysis, such indicators are expressed as $\mathcal{H}_{2,y\in\mathcal{Y}}$ and $\mathcal{K}_{y\in\mathcal{Y}}$. Concerning
%\subsection{Data-driven MCDM}
The conventional MCDM problem formulation \cite{triantaphyllou_2000} is adapted for this study, specifically for assigning CCRCs to achieve optimal stability performance within the online timeframe of power system operation. Fig.~\ref{fig: dm_alg} illustrates the MCDM algorithm application, summarizing its main steps along with the input and output quantities. Consider that the system has to transit from one OP, the one marked as a full circle ($\bullet$) in Fig.~\ref{fig: dm_alg}, to another, marked as an empty circle ($\circ$). The MCDM problem consists of finding, among all the possible alternatives, the CCRC to be assigned at $OP^\circ$ (i.e., $X_C^\circ$) that makes the system stable and with the best stability performances%. What makes such a decision-making algorithm a multi-criteria problem is that the stability performances are measured through
, according to multiple indicators. Concerning the online implementation of the algorithm, this goal can be achieved thanks to the use of data-driven \textit{surrogate} models for the computation of the stability ($\hat{\Upsilon}$) and the stability performance indicators ($\hat{\mathcal{H}}_{2,y}$ and $\hat{\mathcal{K}}_{y}$). However, the speed-up in computation %provided by the data-driven models
is paid for in terms of accuracy. Therefore, before operating the system with the proposed $X_C^\circ$, a verification of the stability response through the \textit{exact} models is carried out. Hence, as depicted in Fig.~\ref{fig: dm_alg}, the main steps of the algorithm are the data-driven MCDM and the verification of the data-driven solution. Next, the problem formulation is described. %and in Algorithm~\ref{alg: decision_making}.
\begin{figure}[b]
\vspace{-0.8cm}
\centering \includegraphics[width=0.95\linewidth]{figures/decision_alg.png}
\caption{Workflow of the data-driven MCDM algorithm}
\label{fig: dm_alg}
\end{figure}
\vspace{-0.4cm}
\subsubsection{Problem Alternatives}
The decision-making problem \textit{alternatives} are the CCRCs. %However, some constraints have to be considered.
In principle, all the possible CCRCs are the \textit{variations without repetitions} of the CCRs. Their number is therefore equal to the number of different CCRs, raised to the power of the number of IPCs in the system. Many of these CCRCs are not feasible, as does not fulfill the operating principles identified in \cite{gomis2020principles}. However, the number of feasible CCRCs is still high, especially for power systems with high penetration of IPCs with multiterminal configuration. Considering such a large number of CCRCs in stability studies, especially if it is intended to employ data-driven tools, adds complexity. Therefore, it is proposed to select a reduced number of feasible CCRCs, according to a selection criterion explained in Section \ref{sec: CCRCs_selection}.\\
%\input{algorithm_sinend}
Define this selection of CCRCs as $\mathcal{R}$, representing the full set of \textit{alternatives}. However, the actual set of \textit{alternatives} considered each time the MCDM algorithm is run changes according to two constraints. %Such constraints both deal with how to carry out the OP and CCRC transitions.
First, it has to be considered that it is possible to assign a CCRC to one OP only if the CCRC is stable during the transition. That is, the proposed CCRC has to be stable both in $OP^\bullet$ and $OP^\circ$. Therefore, if $\mathcal{R}_\Upsilon^{\bullet}$ and $\mathcal{R}_\Upsilon^{\circ}$ collect the CCRCs that result stable at the corresponding OP, the following condition has to be true:
\vspace{-0.3cm}
\begin{equation}
X_C^{\circ}\in \mathcal{R}_\cap \quad \text{with} \quad \mathcal{R}_\cap=\mathcal{R}_\Upsilon^{\bullet} \cap \mathcal{R}_\Upsilon^{\circ}
\label{eq: constr1}
\vspace{-0.2cm}
\end{equation}
% \noindent where
% \begin{equation}
% \mathcal{R}_\cap=\{\mathcal{R}_s^{OP^\circ} \cap \mathcal{R}_s^{OP^\bullet}\}
% \end{equation}
\noindent The second constraint is related to how many simultaneous CCRs changes are allowed when transiting from $X_C^{\bullet}$ to $X_C^{\circ}$.
Consider that a different CCRC assignment can involve the change of a single CCR or multiple CCRs. In case multiple CCR changes are involved, their switch has to happen simultaneously. As large transients might be caused by the difficult coordination, it might be preferred to constrain the decision-making algorithm to look for the best CCRC which implies a limited number of CCRs changes. Indicating $\gamma$ as the number of CCR changes that the assignation of $X_C$ at $OP^\circ$ involves, such a constraint can be expressed as $\gamma \leq \gamma^*$,
% \begin{equation}
% \gamma \leq \gamma^*
% \label{eq: constr2}
% \end{equation}
%\noindent
where $\gamma^*$ is the maximum number of allowed CCRs changes. The value of $\gamma^*$ can be defined by the System Operator (SO). The two mentioned constraints can be summarized as
\vspace{-0.2cm}
\begin{equation}
X_C^{\circ}\in\mathcal{R}'_\cap \quad \text{with} \quad \mathcal{R}'_\cap=\mathcal{R}_\Upsilon^{\bullet}\cap\{\mathcal{R}_\Upsilon^{\circ}|\gamma\leq\gamma^*\}
\label{eq: merge_constr}
\vspace{-0.2cm}
\end{equation}
\noindent According to \eqref{eq: merge_constr}, the %\textit{alternative}
solution of the MCDM problem $X_C^{\circ}$ belongs to the set of \textit{alternatives} $\mathcal{R}'_\cap$. This set collects all the CCRCs that are stable at both the current and successive OPs and involve a number of CCRs changes lower than $\gamma^*$. The value set for $\gamma^*$ might lead to $\mathcal{R}'_\cap=\emptyset$. In this case, it is proposed to relax the constraint %in (\ref{eq: constr2})
by increasing $\gamma^*$ until at least one CCRC stable at both OPs is found (i.e. $\mathcal{R}'_\cap\neq\emptyset$).\\
%\textcolor{red}{A particular case is when $\gamma^*$=0 is set. This corresponds to limiting the decision-making algorithm to look for an improvement of the small-signal stability performances only if $X_C^{\bullet}$ is unstable at $OP^\circ$.}
%All the considerations mentioned above lead to the selection of the MCDM problem \textit{alternatives} $\mathcal{R}'_\cap$, by following the instructions summarized in \textit{Phase 1} of Algorithm \ref{alg: decision_making}.
\vspace{-0.3cm}
\subsubsection{Problem Criteria and Performances}
The \textit{criteria} ($\zeta$) used to compare the alternatives are the enhancement in the stability performances, measured as the improvement of the stability performances indicator for each considered signal. % considered in the analysis. %As specified in \ref{sec: ss_ind}, the stability performances indicators are the $\mathcal{H}_2$-norm and the DC Gain of signals such as the frequency of an AC grid, the voltage of a system node, etc. Therefore, indicating as $y$ a generic system signal belonging to $\mathcal{Y}$, the set collecting all the signals considered in the analysis, the stability performances indicators are expressed as $\mathcal{H}_{2,y}$ and $\mathcal{K}_y$ $\forall y \in\mathcal{Y}$
Denoting by $\mathcal{Y}$ the set of all signals included in the analysis, the number of criteria in the MCDM algorithm is equal to 2$|\mathcal{Y}|$, as both the improvement of the $\mathcal{H}_2$-norm and of the DC Gain are considered for each signal. Therefore, for a signal $y$, the value of the performance ($\rho$) of the $i$-th \textit{alternative}, according to the $j$-th \textit{criterion} is calculated as in (\ref{eq: perf}a) or (\ref{eq: perf}b), whether it refers to the $\mathcal{H}_2$-norm or the DC Gain.
\vspace{-0.2cm}
\begin{equation}
\rho_{i,j}=\hat{\mathcal{H}}_{2,y}^{\circ}-\mathcal{H}_{2,y}^{\bullet}
\quad \text{(a)}\quad;\quad
%\label{eq: perf_h2}
% \end{equation}
% \begin{equation}
\rho_{i,j}=\hat{\mathcal{K}}_{y}^{\circ}-\mathcal{K}_{y}^{\bullet}\quad \text{(b)}
%\label{eq: perf_K}
\label{eq: perf}
\vspace{-0.2cm}
\end{equation}
\noindent The decision to adopt as criteria the improvement of the indicators values rather than their absolute values stems from the fact that these indicators can sometimes approach zero (i.e. their minimum). In such cases, using weights in a weighted sum becomes ineffective. Conversely, an indicator close to zero indicates a favorable stability condition for the system and should be rewarded accordingly. Therefore, by adopting the formulations in (\ref{eq: perf}), an improvement (of the $i$-th \textit{alternative}, according to the $j$-th \textit{criterion}) is achieved when the value of $\rho_{i,j} < 0$. Hence, the best \textit{alternative} of the problem is the one with the lowest weighted sum of performance terms.\\
In expressions (\ref{eq: perf}), the values of the indicators with the operating conditions in $\bullet$ are calculated with the \textit{exact} models, while the performance indicators at $\circ$ are calculated using the data-driven \textit{surrogate} models. This approach is chosen because the \textit{exact} information about the stability and stability performances for the current OP ($\bullet$) is assumed to be known. Whereas, the stability performance at the subsequent OP ($\circ$), considering the assignments of all the possible CCRC in $\mathcal{R}'_{\cap}$, is estimated via the \textit{surrogate} models to accelerate computation compared to using the \textit{exact} model.
\subsubsection{Complete Performance Matrix and Solution}
For the problem under consideration, the full MCDM performance matrix takes the form shown in (\ref{eq: perf_matrix_ddmodm}).
\vspace{-0.3cm}
\begin{equation}
\resizebox{0.89\linewidth}{!}{$
\mathbf{\Pi} =\!
\begin{blockarray}{ccc}
& \mathcal{H}_{2,y\in|\mathcal{Y}|} & \mathcal{K}_{y\in|\mathcal{Y}|}\\
& \overbrace{\zeta_1 \quad \dots \quad \zeta_{|\mathcal{Y}|}} & \overbrace{\zeta_{|\mathcal{Y}|+1} \quad \dots \quad \zeta_{2|\mathcal{Y}|}} \\
\begin{block}{c@{\!\;}[c|c]}
X_{C_1} & \rho_{1,1} \quad \dots \quad \rho_{1,|\mathcal{Y}|} & \rho_{1,|\mathcal{Y}|+1} \quad \dots \quad \rho_{1,2|\mathcal{Y}|} \\
\vdots & \vdots \qquad \ddots \qquad \qquad & \vdots \qquad \ddots \qquad \qquad \\
X_{C_{|\mathcal{R}'_\cap|}} & \rho_{|\mathcal{R}'_\cap|,1} \qquad \quad \rho_{|\mathcal{R}'_\cap|,|\mathcal{Y}|} & \rho_{|\mathcal{R}'_\cap|,|\mathcal{Y}|+1} \quad \rho_{|\mathcal{R}'_\cap|,2|\mathcal{Y}|} \\
\end{block}
\end{blockarray}
$}
\label{eq: perf_matrix_ddmodm}
\vspace{-0.3cm}
\end{equation}
% \begin{equation}
% \mathcal{M} =
% \begin{blockarray}{ccc}
% & \mathcal{H}_{2,y\in|\mathcal{Y}|} & \mathcal{K}_{y\in|\mathcal{Y}|}\\
% & \overbrace{\zeta_1 \quad \dots \quad \zeta_{|\mathcal{Y}|}} & \overbrace{\zeta_1 \quad \dots \quad \zeta_{|\mathcal{Y}|}} \\
% \begin{block}{c[c|c]}
% X_{C_1} & p_{11_{\mathcal{H}}} \quad \dots \quad p_{1|\mathcal{Y}|_{\mathcal{H}}} & p_{11_{\mathcal{K}}} \quad \dots \quad p_{1|\mathcal{Y}|_{\mathcal{K}}} \\
% \vdots & \vdots \qquad \ddots \qquad \qquad & \vdots \qquad \ddots \qquad \qquad \\
% X_{C_{|\mathcal{R}'_\cap|}} & p_{|\mathcal{R}'_\cap|1_{\mathcal{H}}} \qquad \qquad \qquad & p_{|\mathcal{R}'_\cap|1_{\mathcal{K}}} \qquad \qquad \qquad\\
% \end{block}
% \end{blockarray}
% \label{eq: perf_matrix}
% \end{equation}
%\noindent The process describing the computation to build such matrix is illustrated in \textit{Phase 2} of Algorithm \ref{alg: decision_making}.\\
The $X_C^\circ$ to be assigned at $OP^\circ$, to provide the best small-signal stability performances, is the solution of the data-driven MCDM problem. It is the $i^*$-th CCRC, where $i^*$ is
% \begin{align}
% X_C^{\circ}&=\arg\min_{X_C} \Pi_{X_C,\text{score}} = \nonumber\\
% & =\arg\min_{X_C} \sum_{j}^{2|\mathcal{Y}|}{w_{\zeta_j} \rho_{i,j}} \quad \forall X_C \in \mathcal{R}'_\cap
% \label{eq: objfun}
% \end{align}
% \begin{equation}
% %\resizebox{0.89\linewidth}{!}{$
% \textstyle X_C^{\circ} = %\arg\min_{X_C} \Pi_{X_C,\text{score}} =
% \arg\min_{X_C} \sum_{j}^{2|\mathcal{Y}|}{w_{\zeta_j} \rho_{i,j}} \quad \forall X_C \in \mathcal{R}'_\cap
% \label{eq: objfun}
% % $}
% \end{equation}
\vspace{-0.1cm}
\begin{equation}
\textstyle i^* = %\arg\min_{X_C} \Pi_{X_C,\text{score}} =
\arg\min_{i} \sum_{j}^{2|\mathcal{Y}|}{w_{\zeta_j} \rho_{i,j}} \quad \text{with} \quad i =1,..., |\mathcal{R}'_\cap|
\label{eq: objfun}
\vspace{-0.1cm}
\end{equation}
\noindent where $w_{\zeta_j}$ is the weight of the $j$-th criterion.
%\textit{Phase 3} of Algorithm \ref{alg: decision_making} describes this operation.
\subsubsection{Solution Verification}
%As the data-driven computations are accompanied by an error, operating the system with the proposed MCDM solution must be verified through the \textit{exact} models. %\textit{Phase 4} of Algorithm \ref{alg: decision_making} shows the proposed verification process.
%First, the
The stability of $X_C^{\circ}$ is assessed through the \textit{exact} model $\Psi$. If it is confirmed as a stable solution it is assigned for operating the system. Otherwise, the stability of the others $X_C \in \mathcal{R}'_\cap$ is assessed, starting from the one with better performances. The first CCRC providing a stable system response is considered the MCDM final solution.
The MCDM algorithm illustrated in Section \ref{sec: MODM} bases the computation of the \textit{performance} matrix on the data-driven \textit{surrogate} models defined in Section \ref{sec: ss_ind}. This Section proposes a methodology for obtaining such \textit{surrogate} models. %The main steps are summarized in Figure \ref{fig: methodology}.
First, the set of CCRCs \textit{alternatives} is selected. Then, for each selected CCRC a training data set is generated by computation, employing the \textit{exact} models. The obtained data sets are properly arranged. Finally, the classification and regression \textit{surrogate} models are trained. Next, the mentioned steps are described.
% \begin{figure}
% \centering
% \includegraphics[width=0.8\linewidth]{figures/methodology3.pdf}
% \caption{Methodology scheme.}
% \label{fig: methodology}
% \end{figure}
\vspace{-0.5cm}
\subsection{Select a Reduced Set of CCRCs Feasible Alternatives}
\label{sec: CCRCs_selection}
\vspace{-0.1cm}
One possibility is considering as the set of possible CCRCs \textit{alternatives} ($\mathcal{R}$) all the feasible CCRCs. That is, the CCRCs fulfilling the operating principle stated in \cite{gomis2020principles}. According to the size and topology of the system, i.e. the number of IPCs in the grid and whether they are involved in multi-terminal connections, the number of feasible CCRCs can rise a lot. This presents an added challenge for data-driven analysis, as data must be generated for each CCRC. To address this, this study proposes a methodology for identifying a reduced set of CCRCs that maintains good stability performance across the entire operating space based on multiple performance indicators. This approach extends the methodology introduced in \cite{10202947}, which was initially developed for a single indicator.\\
The methodology introduced in \cite{10202947} starts with generating a data set of random OPs, considering all feasible CCRCs. For each OP the small-signal stability is assessed and the performance indicators are obtained. Then, a clustering process is applied to identify and group the CCRCs that have similar stability performance behavior in the same sub-regions of the operating space. Finally, a Decision Tree (DT) regression algorithm extracts the behavior of the clusters in the form of conditional rules, which are used to select the final set of CCRCs that cover the entire operating space with the best performances.
% The methodology introduced in \cite{10202947} is developed through a series of steps detailed as follows:
% \begin{enumerate}
% \item Data set generation: $N_{OP}$ random operating points (OPs) are sampled within the system’s operating space. For each OP, the small-signal stability analysis is performed using all feasible CCRCs. The analysis involves: (i) PF calculation to obtain $X_{PF}$ and $X_{IPC}$, (ii) small-signal stability assessment (Definition~\ref{def: exact_ss}) to obtain $ \Upsilon $, and (iii) calculation of the stability performance indicator, set to infinity if the system is unstable or calculated based on Definition~\ref{def: exact_h2} or Definition~\ref{def: exact_dcgain} otherwise, depending on whether the \( \mathcal{H}_2 \)-norm or the DC gain is used. This results in $N_{OP} \cdot |\mathcal{R}|$ instances, organized into a dataset $\mathcal{D} = (X | Z)$, where $X$ contains all $X_{PF}$ and $X_{IPC}$ values, and $Z$ holds the stability indicator values for each CCRC across all $N_{OP}$ instances.
% \item Attribute matrix creation: The data in the generated dataset $\mathcal{D}$ is processed to construct the attribute matrix, which serves as the input for the clustering algorithm. Principal Component Analysis (PCA) \cite{wold1987principal} is applied to the quantities in $X$ to reduce the number of dimensions describing the system’s operating space. Thus, a 3D space is obtained which is then divided into subregions. Based on the indicator values, a stability performance level is assigned to each subregion. This enables the identification of a way to describe the stability performance behavior of each CCRC — specifically, how good the stability performance is for each CCRC in each subregion of the operating space.
% \item CCRCs clustering: The k-Means clustering algorithm \cite{arthur2006k} is applied to group the CCRCs that have similar stability performance behavior (i.e. which show similar levels of stability performances according to the indicator used) in the same subregions of the operating space.
% \item Knowledge Extraction: The clusters group CCRCs with similar stability performance behavior. By applying a Decision Tree (DT) regression algorithm \cite{franklin2005elements}, an explicit description of this behavior is derived in the form of conditional rules. These rules identify regions of the operating space and provide the average value of the stability performance indicator within each region. Therefore, after applying this DT-based rule extraction process to all the clusters, a list is generated that includes all the mined rules, along with their associated average indicator value and the cluster of CCRCs from which they were derived.
% \item CCRCs selection: Starting from the top of the sorted list of rules, each rule is analyzed in sequence. If a rule covers a portion of the operating space not yet addressed by previous rules, the CCRCs associated with that rule’s cluster are selected. Conversely, if the operating space covered by a rule overlaps with a previously analyzed rule, it indicates that other CCRCs offer better performance in that region, so the CCRCs in that cluster are not selected. This process continues until the system’s entire operating space is fully covered or until all rules in the list have been examined.
% \end{enumerate}
The process described above
%This process
can be repeated for multiple stability performance indicators. %\textcolor{red}{As noted in \cite{10443561}, a CCRC may perform well for one indicator but poorly for another, potentially leading to the selection of different CCRCs for each indicator.}
To achieve the final selection that accounts for multiple indicators, intersecting set visualization techniques are employed. Specifically, the UpSet method proposed in \cite{6876017}, designed to visualize interactions among more than three sets, is adapted to this context. %Specifically, as outlined in \cite{6876017}, a combination matrix is created where the elements represent the CCRCs, and the sets correspond to the performance indicators used in the selection process. Each element (i.e. CCRC) is characterized by attributes, resulting from the single indicator selection process. In particular, these attributes are:
Fig.~\ref{fig: set_inters} illustrates this visualization method. The CCRCs to be grouped are analyzed based on their attributes, defined as follows:
\begin{itemize}
\item \textit{Attribute 1}: Cluster membership. Each CCRC is associated with a cluster, indicated in Fig.~\ref{fig: set_inters} by a number (e.g., 1 = Cluster 1, 2 = Cluster 2, etc.).
\item \textit{Attribute 2}: Whether the cluster to which a CCRC belongs is selected. This is shown in Fig.~\ref{fig: set_inters} using a rhomboidal mark (\raisebox{-0.2em}{\includegraphics[height=1em]{figures/rombo_grigio.pdf}} for selection based on $\mathcal{H}_{2,y}$, and \raisebox{-0.2em}{\includegraphics[height=1em]{figures/rombo_nero.pdf}} for selection based on $\mathcal{K}_{y}$).
\end{itemize}
Since cluster membership is irrelevant for CCRCs in unselected clusters, these CCRCs are first grouped into a generic Cluster 0. Next, a combined attribute is created to describe CCRCs, capturing the combination of cluster memberships for selected clusters. In the example shown in Fig.~\ref{fig: set_inters_a}, seven such combinations are identified (\raisebox{-0.2em}{\includegraphics[height=1em]{figures/attr1.pdf}}, \raisebox{-0.2em}{\includegraphics[height=1em]{figures/attr2.pdf}}, \raisebox{-0.2em}{\includegraphics[height=1em]{figures/attr3.pdf}}, etc.). CCRCs with the same combined attribute are then grouped via a set intersection process. Finally, one CCRC is selected from each group (e.g., the first one in the list) and added to the set of selected CCRCs, $\mathcal{R}$. In this way, $\mathcal{R}$ collects the CCRCs required to operate in the entire system operating space with enhanced small-signal stability performances, according to multiple indicators. For completeness, Fig.~\ref{fig: set_inters_b} offers an equivalent representation of Fig.~\ref{fig: set_inters_a} using Venn diagrams. Although this familiar technique effectively illustrates the same concept, it becomes less readable when dealing with many elements and intersections across multiple indicators. In such cases, the visualization method proposed in Fig.~\ref{fig: set_inters_a} is more suitable.
%CCRCs $X_{C_1}$-$X_{C_2}$ and $X_{C_2}$-$X_{C_3}$ are dealt with as CCRCs with different properties as they show a different stability behavior depending if analyzed according to different performance indicators.
\vspace{-0.3cm}
\begin{figure}[h!]
\centering
\subfloat[Proposed sets intersection visualization.]{
\includegraphics[width=0.5\linewidth]{figures/select_intersection_a.pdf}
\label{fig: set_inters_a}}
%\hfill
\subfloat[Equivalent Venn diagrams representation.]{
\includegraphics[width=0.5
\linewidth]{figures/select_intersection_b.pdf}
\label{fig: set_inters_b}
}
\caption{Sets intersection visualization and CCRCs selection.}
\label{fig: set_inters}
\vspace{-1cm}
\end{figure}
\subsection{Data Generation}
\label{sec: datagen}
To train and validate the data-driven \textit{surrogate} models, data are required. In particular, it is necessary to collect information about the small-signal stability and stability performances in the operating space, considering the assignment of each CCRC in $\mathcal{R}$. Since these scenarios are not available during the system’s real operation, it is necessary to generate such data by computation. The computing tools involved are the PF calculation and the \textit{exact} models $\Psi$, $\Gamma_{\mathcal{H}_{2,y}}$, and $\Gamma_{\mathcal{K}_{y}}$ $\forall y \in \mathcal{Y}$. For each CCRC $\in \mathcal{R}$ the data generation process is carried out twice, once for generating the training data and once for generating data for final validation. Two different sampling strategies are applied. To generate training data, the method introduced in \cite{rossi2022data} is utilized. This approach focuses on efficiently producing high-quality data for training stability predictive models. It enables comprehensive exploration of the system's stability across its entire operating space while ensuring higher granularity within the stability margin. The process employs Latin Hypercube Sampling (LHS) to randomly select OPs and uses the Entropy function to identify the stability margin and guide the sampling process toward it. To generate validation data, $N_{OP}$ random OPs are sampled in the entire system operating space by the use of LHS. After the data are generated, they are organized into training and test data sets as indicated in Definitions \ref{def: dd_stab}-\ref{def: dd_dcgain}.
\vspace{-0.5cm}
\subsection{Features Engineering}
%\vspace{-0.2cm}
Feature engineering is applied to preprocess data in the training and test data sets. This process consists of creating, transforming, and selecting relevant features from raw data to improve the performance of ML models. It involves techniques such as data cleaning, scaling, encoding categorical data, handling outliers, and creating new features that better represent the underlying patterns in the data.
%\begin{enumerate}
\subsubsection{Creating New Features}
Additional input features are created starting from the variables collected in $X_{PF}$ and $X_{IPC}$. Specifically, the added features are:
\vspace{-0.1cm}
\begin{itemize}
\item The apparent power injected by generators, at both sides of the IPCs, absorbed by the loads, and injected or absorbed by the Thevenin's equivalents.
% \begin{equation}
% S_i=\sqrt{P_i^2 + Q_i^2}, \forall i\in\{\mathcal{G}\cup\mathcal{C}\cup\mathcal{L}\cup\mathcal{T}\}
% \end{equation}
\item Direction of the PF: This applies to system assets such as IPCs and Thevenin equivalents, which can experience PF in both directions. %This is accounted for in power flow calculations by assigning positive or negative values to the power. For IPCs, the sign convention is as follows: active and reactive power are assigned a positive sign if the power flows from the DC to the AC side of the converter, and a negative sign for the opposite flow direction. For Thevenin equivalents, the sign of active and reactive power depends on the power's interaction with the system. A positive sign indicates that power is injected into the system, while a negative sign indicates that it is absorbed. Based on this convention,
Boolean features indicating the sign of the power flow can be %derived and
added to the datasets.
% \begin{gather}
% \begin{cases}
% \mbox{sign}(P_i)\!=\!1 \quad if P_i > 0\\
% \mbox{sign}(P_i)=0 \quad otherwise\\
% \forall i \in \{\mathcal{C}\cup\mathcal{T}\}
% \end{cases}
% \label{eq: signP}
% \end{gather}
% \begin{gather}
% \begin{cases}
% \mbox{sign}(Q_i)=1 \quad if Q_i > 0\\
% \mbox{sign}(Q_i)=0 \quad otherwise\\
% \forall i \in \{\mathcal{T}\cup\mathcal{C}\}
% \end{cases}
% \label{eq: signQ}
% \end{gather}
\vspace{-0.1cm}
\end{itemize}
\subsubsection{Data Cleaning}
The data cleaning process includes removing features containing a single value, duplicated features, and correlated features. To identify correlated features, the Pearson correlation coefficient is used. The coefficient is computed for each pair of variables, and pairs with a high coefficient %exceeding 0.999
are flagged for review to ensure critical system information is preserved. For highly correlated pairs, one variable is considered for removal based on the following criteria:
\begin{itemize}
\item Power-related variables ($P$, $Q$, $S$): Both %features are
retained.
\item Features associated with different nodes: Both %features are
retained.
\item Other cases: The feature with the lower priority is removed, following this hierarchy: currents, AC voltages, DC voltages, and angles.
\end{itemize}
\vspace{-0.1cm}
\subsubsection{Data Scaling}
The input quantities and, for the regression case, the output data are scaled. % through a standard scaler and a min/max scaler, respectively.
% process \eqref{eq:stdscaler}. The standard scaler utilizes the mean $\mu_{\bf{x}}$ and standard deviation $\sigma_{\bf{x}}$ of a certain data vector $\bf{x}$ which centers and scales it into $\bf{z}$.
% % \begin{equation}
% % \label{eq:stdscaler}
% % \bf{z} = \dfrac{\bf{x}-\mu_{\bf{x}}}{\sigma_{\bf{x}}}
% % \end{equation}
% % %\end{enumerate}
\subsubsection{Handling Outliers} %When training regressions, the presence of outliers can worse the accuracy of the model.
Winsorizing is considered, which replaces the outliers with the values of a specified percentile.
\vspace{-0.5cm}
\subsection{Models Training}
As outlined in Definitions \ref{def: dd_stab}-\ref{def: dd_dcgain}, the data-driven \textit{surrogate} models designed to substitute the \textit{exact} models consist of a classifier to predict system stability and regression models to estimate the values of the stability performance indicators. Various classification and regression algorithms could effectively perform these tasks. To determine the most suitable ML technique, a comparison of their performance is conducted by training and testing multiple algorithms on the generated data. First, appropriate metrics are chosen to evaluate the performance of the algorithms. Next, several algorithms are trained, and their accuracy is assessed. Based on this comparison, either the best-performing algorithm is selected or the top-performing algorithms are shortlisted as candidates. Subsequently, the training of the selected model(s) is refined through feature selection and hyperparameter tuning. A detailed explanation of these steps is provided next.
\subsubsection{Set Metrics}
The classification task involves predicting the system's stability. Misclassifying an unstable case as stable poses a critical risk to system operation, whereas misclassifying a stable case as unstable, while inconvenient, is less severe. Thus, a conservative approach prioritizes accurately predicting true stable cases. Precision and Recall are key metrics for evaluating the model's performance in this context. Precision %~\eqref{eq: precision}
measures the proportion of true stable cases out of all cases predicted as stable (i.e., true stable and false stable) \cite{sammut2011encyclopedia}. Recall %~\eqref{eq: recall}
evaluates the proportion of true stable cases out of all cases that are actually stable (i.e., true stable and false unstable) \cite{sammut2011encyclopedia}.
% \begin{equation}
% \text{Precision} %\frac{\text{true positive}}{\text{true positive}+\text{false positive}}=\\
% =\frac{\text{true stable}}{\text{true stable}+\text{false stable}}%\nonumber
% \label{eq: precision}
% \end{equation}
% \begin{equation}
% \text{Recall}=
% %\frac{\text{true positive}}{\text{true positive}+\text{false negative}}=\\
% \frac{\text{true stable}}{\text{true stable}+\text{false unstable}}%\nonumber
% \label{eq: recall}
% \end{equation}
A metric that combines Precision and Recall can be useful for this task. The $F_{\beta}$ score \cite{sammut2011encyclopedia} %(\ref{eq: fbeta_score})
is defined as the harmonic mean of Precision and Recall, weighted by a factor $\beta$. Typically, $\beta=0.5$ or $\beta=2.0$ is chosen, depending on whether Precision or Recall should be given more weight, respectively. The balance of instances in the training data significantly affects model training, as the model is more likely to predict the dominant class. Hence, this work proposes adopting $\beta=2.0$ when less than 50\% of the input data represents stable cases, prioritizing Recall to minimize false unstable. Conversely, if more than 50\% of the data is stable, $\beta=0.5$ is used to prioritize Precision and reduce false stable.
% \begin{equation}
% F_{\beta} = (1+\beta^2) \cdot \dfrac{\textnormal{Precision}\cdot\textnormal{Recall}}{\beta^2\cdot\textnormal{Precision}+\textnormal{Recall}}
% \label{eq: fbeta_score}
% \end{equation}
Concerning the regression models fitting the stability performance indicators, the metrics proposed is the $R^2$ score.
\subsubsection{Compare Models Performances}
To determine the most suitable ML model for each \textit{surrogate} model, a performance-based comparison is conducted across all options. %Since the final application of these models (i.e., the data-driven MODM algorithm) imposes no specific constraints on the choice of ML models, the selection is driven solely by their performance on the available data.
A k-Fold Cross-Validation (kFCV) approach is employed, and the comparison is based on the selected evaluation metrics.
%For the classification task, the models compared include a Dummy classifier, DT Classifier (DTC), Neural Network (NN) based on Multi-Layer Perceptron (MLP), Extreme Gradient Boosting DTC (XGB-DTC), and Logistic Regression.
%For regression tasks, the evaluated models include a Dummy regressor, linear regression, ridge regression, DT Regressor (DTR), Neural Network (NN) based on MLP, Extreme Gradient Boosting DTR (XGB-DTR).%, and Multivariate Adaptive Regression Splines (MARS).
\subsubsection{Permutation Feature Importance}
Once the best-performing algorithm or the top-performing algorithms are identified, %to improve the algorithms performances
a feature selection can be performed, based on permutation feature importance (PFI).
\subsubsection{Hyperparameters Tuning}
A Grid-Search kFCV (GSkFCV) process is utilized to tune the hyperparameters of the selected models.
The proposed data-driven MCDM algorithm is demonstrated using a hybrid AC/DC test system. This section presents a detailed description of the test system, outlines the implementation of the methodology for training the \textit{surrogate} models, provides further insights into the behavior of the stability performance indicators, and showcases the results of the MCDM algorithm implementation. %The obtained results are also validated by comparing them with those derived from exact models and dynamic time-domain simulations.
% The application of the proposed methodology and of the developed data-driven MCDM algorithm is exemplified through the hybrid AC/DC test system represented in Fig.~\ref{fig: power_syst} \cite{arevalo2022}. Concerning the small-signal stability, among all the possibilities, the voltage in the DC grids and the frequency in the AC grids are selected as the system signals for quantifying the stability performances. Therefore, $\mathcal{Y}=\{V_{DC},f\}$ and the small-signal stability performances indicators are $\mathcal{H}_{2,V_{DC}}$, $\mathcal{H}_{2,f}$, $\mathcal{K}_{V_{DC}}$ and $\mathcal{K}_{f}$.
\vspace{-0.6cm}
\subsection{Test System Description}
\label{sec: test_system}
\vspace{-0.2cm}
The test system used throughout this study is the hybrid AC/DC grid represented in Fig.~\ref{fig: power_syst} \cite{10017135}.
This grid consists of three AC and two HVDC sub-grids, interconnected by six IPCs. %As shown in Fig.~\ref{fig: power_syst},
In both AC-1 and AC-2 sub-grids there are:
\begin{itemize}
\item Two loads, representing the aggregated power demand.
\item One renewable-based generator, representing, in aggregated form, wind or photovoltaic-based power generation.
\item A Thevenin equivalent for the rest of the grid.
\end{itemize}
The sub-grid AC-3 has one wind-based power plant, which might represent a large off-shore wind farm.
\vspace{-0.2cm}
\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{figures/power_syst_BigModel_bus.pdf}
\caption{Hybrid AC/DC grid test system \cite{10017135}.}
\label{fig: power_syst}
\vspace{-0.3cm}
\end{figure}
The system's operating space is defined by the operating ranges of the generators and the minimum and maximum power demand. Table \ref{tab: gen_range} summarizes the operating ranges of the generators, including the minimum and maximum active power they can inject, set to 5\% and 95\% of their nominal power, respectively. Regarding reactive power, all generators are assumed to support power factors ranging between 0.8 and 0.95. %, allowing them to inject or absorb reactive power.
Table \ref{tab: dem_range} outlines the load consumption ranges, where the total power demand varies between 200 MW and 700 MW. Individual loads are assumed to contribute to the total demand within a range of $\pm$30\% of their base load, defined as a percentage of the total demand.
\vspace{-0.3cm}
\begin{table}[h]
\caption{Generators operating range.}
\label{tab: gen_range}
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ccc|ccc}
\toprule
& Minimum & Maximum & & Minimum & Maximum \\ \midrule
$P_{G_1}$ {[}MW{]} & 15 & 285 & $\cos \phi_{G_1}$ & 0.8 & 0.95 \\
$P_{G_2}$ {[}MW{]} & 5 & 95 & $\cos \phi_{G_2}$ & 0.8 & 0.95 \\
$P_{G_3}$ {[}MW{]} & 7.5 & 142.5 & $\cos \phi_{G_3}$ & 0.8 & 0.95 \\ \bottomrule
\end{tabular}%
}
\vspace{-0.3cm}
\end{table}
\begin{table}[t]
\centering
\caption{Power demand range}
\label{tab: dem_range}
\begin{tabular}{cccc}
\toprule
& Base Load & Minimum & Maximum \\ \midrule
$P_D$ {[}MW{]} & {[}-{]} & 200 & 700 \\
$P_{L_1}$ & 30\% $P_D$ & 21\% $P_D$ & 39\% $P_D$ \\
$P_{L_2}$ & 20\% $P_D$ & 14\% $P_D$ & 26\% $P_D$ \\
$P_{L_3}$ & 20\% $P_D$ & 14\% $P_D$ & 26\% $P_D$ \\
$P_{L_4}$ & 30\% $P_D$ & 21\% $P_D$ & 39\% $P_D$\\
\bottomrule
\end{tabular}
\vspace{-0.7cm}
\end{table}
The IPCs interconnecting the AC and DC grids are MMCs. Each IPC is provided with three control roles. Following the definitions provided in \cite{10443561}, such control roles are:
\begin{itemize}
\item AC-GFM control that controls the voltage on the AC side
\item DC-GFM control that controls the voltage on the DC side through a droop control.
\item GFL control that doesn't control the voltage of any terminal and synchronizes with the AC grid by a PLL.
\end{itemize}
\vspace{-0.1cm}
\noindent Given the possibility of implementing such CCRs in each IPC, the number of possible CCRCs is theoretically $N_{CCRCs}=(N_{CCRs})^{|\mathcal{C}|}=3^6=729$, where $N_{CCRs}$ is the number of CCRs per IPC, and $|\mathcal{C}|$ the number of IPCs in the system. However, a large number of these CCRCs result unfeasible and have to be discarded, following the operating principles identified in \cite{gomis2020principles}. Such principles of operation state that, in a hybrid AC/DC grid, each sub-grid needs at least one grid-forming unit. This means that during system operation: (i) In each AC sub-grid, at least one node has to be an AC-GFM unit, either an IPC or another type of unit; (ii) In each DC sub-grid, at least one node has to be an IPC operating with DC-GFM control. %For the test system, this means that the
%Hence, the number of feasible CCRCs decreases to 95.
For the test system, it follows that IPC-E can only be an AC-GFM unit and also many other CCRCs must be discarded. Therefore, the number of feasible CCRCs is reduced to $N_{CCRCs}=95$. %feasible CCRCs.
\vspace{-0.6cm}
\subsection{Small-signal Stability Performances}
\vspace{-0.1cm}
The assessment of small-signal stability performance necessitates the analysis of signals capable of capturing the dynamic behavior of both AC and DC grids. Accordingly, this study evaluates frequency signals to characterize AC grid dynamics and voltage signals to capture DC grid dynamics. For the system case study, three frequency signals are analyzed — one per AC grid — along with six voltage signals, related to each DC grid bus. Although stability performance indicators can be computed for each signal individually, it is proposed to evaluate them in a way that provides an overall assessment of the dynamic behavior across all AC and DC grids. Therefore, the transfer functions matrix required for the calculation of the $\mathcal{H}_2$ norms and DC gain (as defined in Definitions \ref{def: exact_h2} and \ref{def: exact_dcgain}) are determined by setting the active and reactive power at all buses as inputs. Regarding the outputs, the transfer functions matrix for frequency uses the frequencies of the three AC grids, while the transfer functions matrix for voltage considers the voltages at all six DC buses. Then, the calculation of the $\mathcal{H}_2$-norm directly provides a single-value indicator that accounts for the total energy of all the signals. In contrast, the DC gain calculation returns a matrix of dimensions \(\mathbb{R}^{n_{\text{outputs}} \times n_{\text{inputs}}}\), which collects the DC gains for each input-output combination. To derive a single-value indicator, the maximum DC gain is considered, representing the highest deviation across all signals. As a result, four stability performance indicators are considered: \( \mathcal{H}_{2,f} \), \( \mathcal{H}_{2,V_{DC}} \), \( \mathcal{K}_f \), and \( \mathcal{K}_{V_{DC}} \).
\vspace{-0.5cm}
\subsection{Reduced Set of CCRCs}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth]{figures/venn_indicators_mod.pdf}
\caption{CCRCs selection trough sets intersection visualization.}
\label{fig: venn}
\vspace{-0.6cm}
\end{figure*}
Through the application of the methodology described in Section~\ref{sec: CCRCs_selection}, the minimum set of CCRCs required to operate the system with the best stability performance is identified. The stability behavior of the system across its entire operating space is evaluated for each stability indicator, considering all 95 feasible CCRCs. A clustering algorithm is applied to group CCRCs based on their ability to provide similar stability performance within specific subregions of the operating space. The clustering results are illustrated using stability performance maps, as proposed in \cite{10202947}. Fig.~\ref{fig: stab_maps} displays the map for the $\mathcal{H}_{2,V_{DC}}$ indicator. Maps corresponding to other indicators are omitted for brevity. In this map, each row corresponds to a CCRC, while each column represents a subregion of the %transformed 3-D
operating space. %, obtained through Principal Component Analysis.
The color in each cell indicates the stability performance level of a given CCRC in a specific subregion. CCRCs are grouped according to their clusters and ordered from top to bottom, reflecting their average performance across the entire operating space, from worst to best. The top rows in the maps represent CCRCs which consistently exhibit instability across all subregions (performance level = 5). These CCRCs are excluded in successive analyses. The subregions are arranged from left to right, starting with those where CCRCs demonstrate the best average performance and ending with those where performance is poorest.
%\vspace{-0.cm}
\begin{figure}[htbp]
\centering
%\subfloat[Subfigure 1]
%{
\includegraphics[width=\linewidth]{figures/Rearrang_att_matrix_H2_vdc_condensed.pdf}%\label{fig:subfig36}}
% \hspace{-2.5cm}
% \hfill
% \subfloat[Subfigure 2]{\includegraphics[width=0.5\linewidth]{figures/Rearrang_att_matrix_DCgain_vdc.pdf}\label{fig:subfig37}}
% \subfloat[Subfigure 3]{\includegraphics[width=0.5\linewidth]{figures/Rearrang_att_matrix_H2_freq.pdf}\label{fig:subfig36}}
% \hspace{-3cm}
% \hfill
% \subfloat[Subfigure 4]{\includegraphics[width=0.5\linewidth]{figures/Rearrang_att_matrix_DCgain_freq.pdf}
%\label{fig: stab_maps}}
\caption{Stability maps considering all CCRCs (left) and a reduced set with better performance (top right) based on $\mathcal{H}_{2,V_{DC}}$. The bottom right panel presents a box plot comparing the distribution of the indicator.}
\label{fig: stab_maps}
\vspace{-0.3cm}
\end{figure}
Following the selection process outlined in \cite{10202947}, a reduced set of CCRCs is identified. Fig.~\ref{fig: stab_maps} presents the stability map for this reduced set, derived by extracting the stability maps of the selected CCRCs from Fig.~\ref{fig: stab_maps}. The figure shows how the chosen CCRCs ensure stable operation across the entire operating space, with the best possible performance. Additionally, the box plot in Fig.~\ref{fig: stab_maps} compares the distribution of the $\mathcal{H}_{2,V_{DC}}$ indicator when using all CCRCs versus the selected subset, highlighting a significant reduction in its values.
% \begin{figure}[htbp]
% \centering
% \subfloat[]{\includegraphics[width=0.5\linewidth]{figures/Rearrang_att_matrix_sel_comb_H2_vdc_mod.pdf}\label{fig: stab_maps_red}}
% % \hspace{-2.5cm}
% \hfill
% \subfloat[]{\includegraphics[width=0.5\linewidth]{figures/boxplot_h2vdc_mod.pdf}\label{fig: boxplot_hvdc}}
% \caption{(a) Stability map of the set of selected CCRCs. (b) Comparison of the $\mathcal{H}_{2,V_{DC}}$ distributions.}
% \end{figure}
% \begin{figure}
% \centering
% \includegraphics[width=\linewidth]{figures/boxplot_indicators.pdf}
% \caption{Caption}
% \label{fig:enter-label}
% \end{figure}
The CCRC clustering and selection process is also applied to the $\mathcal{H}_{2,f}$, $\mathcal{K}_{V_{DC}}$, and $\mathcal{K}_{f}$ indicators. As outlined in Section~\ref{sec: CCRCs_selection}, the clustering results are visualized using a set representation, as shown in Fig.~\ref{fig: venn}. This representation facilitates the final selection of the minimum number of CCRCs required to ensure good stability performance across the entire operating space, considering multiple indicators. The number of CCRCs is minimized by retaining only one from each group of CCRCs that exhibit similar stability performance within the same regions of the operating space.
%Table~\ref{tab: set_tables} details the CCRCs corresponding to the sets depicted in Fig.~\ref{fig: venn}. Each CCRC is described by specifying the control roles assigned to each IPC and is grouped based on the clustering intersection results. The CCRCs selected within each group are highlighted in bold and are the elements of $\mathcal{R}$.
% \begin{table}[b!]
% \vspace{-0.5cm}
% \caption{}
% \label{tab:my-table}
% \begin{tabular}{ccccllll}
% \toprule
% Indicator & mean & std & min & \multicolumn{1}{c}{25\%} & 50\% & 75\% & max \\ \midrule
% $\mathcal{H}_{2,f}$ & 0.20 & 0.05 & 0.16 & 0.18 & 0.20 & 0.22 & 1.29 \\
% $\mathcal{H}_{2,V_{DC}}$ & 2.41 & 3.51 & 0.12 & 0.41 & 1.28 & 2.64 & 60.36 \\
% $\mathcal{K}_f$ & \multicolumn{1}{l}{0.021} & \multicolumn{1}{l}{0.004} & \multicolumn{1}{l}{0.015} & 0.017 & 0.022 & 0.024 & 0.029 \\
% $\mathcal{K}_{V_{DC}}$ & \multicolumn{1}{l}{0.113} & \multicolumn{1}{l}{0.091} & \multicolumn{1}{l}{0.013} & 0.027 & 0.087 & 0.177 & 0.384 \\ \bottomrule
% \end{tabular}
% \end{table}
\subsubsection*{Insight on Indicators Behavior}
% \begin{table}[b!]
% \vspace{-0.5cm}
% \caption{Clusters of CCRCs with similar dynamic behavior.}
% \label{tab: set_tables}
% \resizebox{0.98\linewidth}{!}{$
% \begin{tabular}{lcccccc}
% \toprule
% CCRC & {\color[HTML]{000000} IPC-A} & {\color[HTML]{000000} IPC-B} & {\color[HTML]{000000} IPC-C} & {\color[HTML]{000000} IPC-D} & {\color[HTML]{000000} IPC-E} & {\color[HTML]{000000} IPC-F} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{3182BD} 1} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{6BAED6} 2} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{C6DBEF} 3} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{E6550D} 4} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{FDAE6B} 5} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{FDD0A2} 6} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{74C476} 7} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} \\
% \cellcolor[HTML]{74C476} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\
% \cellcolor[HTML]{74C476} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\
% \cellcolor[HTML]{74C476} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\
% \cellcolor[HTML]{74C476} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\
% \cellcolor[HTML]{74C476} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\
% \cellcolor[HTML]{74C476} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{A1D99B} 8} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\
% \cellcolor[HTML]{A1D99B} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} DC-GFM} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{756BB1} 9} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} \\
% \cellcolor[HTML]{756BB1} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} DC-GFM} \\
% \cellcolor[HTML]{756BB1} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} DC-GFM} \\
% \cellcolor[HTML]{756BB1} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} GFL} \\
% \cellcolor[HTML]{756BB1} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} DC-GFM} \\
% \cellcolor[HTML]{756BB1} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} DC-GFM} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{BCBDDC} 10} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} \\
% \cellcolor[HTML]{BCBDDC} & {\color[HTML]{000000} GFL} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} DC-GFM} & {\color[HTML]{000000} AC-GFM} & {\color[HTML]{000000} AC-GFM} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{DADAEB} 11} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{969696} 12}& {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{BDBDBD} 13} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \midrule
% \multicolumn{1}{c}{\cellcolor[HTML]{D9D9D9} 14} & {\color[HTML]{000000} \textbf{GFL}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} & {\color[HTML]{000000} \textbf{AC-GFM}} & {\color[HTML]{000000} \textbf{DC-GFM}} \\ \bottomrule
% \end{tabular}
% $}
% \end{table}
Fig.~\ref{fig: ind_behav} illustrates the trends of the stability indicators by comparing their values across 100 random OPs, considering only the stable cases. Specifically, \( \mathcal{H}_{2,f} \) is plotted against the \( \mathcal{H}_{2,V_{DC}} \), while \( \mathcal{K}_{f} \) is compared with the \( \mathcal{K}_{V_{DC}} \). The results indicate that, for all CCRCs, the indicators related to DC voltage and frequency exhibit a positive correlation. This suggests that, when a single CCRC is applied, operating points with improved frequency dynamics also tend to display better DC voltage dynamics, and vice versa. However, when comparing the same operating points under different CCRCs, a trade-off emerges: CCRCs that enhance frequency dynamics tend to deteriorate DC voltage dynamics. This effect is particularly evident when selecting a single operating point and analyzing its indicator values across all CCRCs for which it remains stable. For the selected operating point (marked with a red cross in Fig.~\ref{fig: ind_behav}), applying CCRC number 46 %7
yields the best DC voltage dynamics but the worst frequency dynamics. Conversely, to achieve optimal frequency dynamics, the assigned CCRC should be CCRC number 41. %6.
Fig.~\ref{fig: npf50} compares the DC bus voltage dynamics and AC grid frequency dynamics for the selected operating point, considering the application of CCRCs number 17, 41, 46. %3, 6, 7.
A 1\% load increase at Load 1 is applied as a disturbance. The plots confirm the insights provided by the stability indicators, demonstrating that CCRC number 46 %7
achieves the smallest maximum DC voltage deviation. However, this CCRC does not yield good frequency dynamics, which are instead better with CCRC number 41. %6.
Lastly, the dynamics resulting from CCRC number 17 %3
are included to illustrate the worst-case performance scenario in terms of DC voltage deviation.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{figures/index_vs_index_labels.pdf}
\caption{Relations and trends of indicators.}
\label{fig: ind_behav}
\vspace{-0.7cm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\linewidth]{figures/N_PF50_insihgt_notime.pdf}
\caption{DC voltages and AC frequencies comparison for the same OP, considering the application of CCRCs providing best and worst performances.}
\label{fig: npf50}
\vspace{-0.5cm}
\end{figure}
\vspace{-0.5cm}
\subsection{Generated Data Sets}
\vspace{-0.1cm}
The data generation process is carried out considering the application of the 14 CCRCs in $\mathcal{R}$. The data set for training $\hat{\Psi}$, $\mathcal{D}_\Upsilon$, collects 11.421 instances, with 46\% of stable cases. %\textcolor{red}{add something on other datasets}%The arrays of input variables have dimensionality $X_{PF} \in \mathbb{R}^{1\times??}$, $X_{IPC} \in \mathbb{R}^{1\times??}$, $X_{C} \in \mathbb{R}^{1\times??}$.
%\subsection{Features Engineering}
\vspace{-0.6cm}
\subsection{Models Training}
\subsubsection{Training of $\hat{\Psi}$}
Given the composition of the data set $\mathcal{D}_\Upsilon$, the metrics used to compare and assess the quality of the trained model is the score $F_{\beta = 2}$. An initial model comparison is performed, including a Dummy classifier that predicts a constant value (\( \hat{\Upsilon_i} = 1 \)). The other models considered in the comparison are DT Classifier (DTC), Neural Network (NN) based on Multi-Layer Perceptron (MLP), Extreme Gradient Boosting DTC (XGB-DTC), and Logistic Regression. All algorithms, except XGB-DTC, are implemented using the \textit{Scikit-Learn} library%\cite{scikit-learn}
. The XGB-DTCs model is implemented using the library \textit{XGBoost}. %provided in \cite{Chen:2016:XST:2939672.2939785}.
For this comparison, all algorithms are tested with their parameters set to default values. It emerges that the MLP and XGB achieve the highest $F_\beta$ score %. Given their
with comparable performances. Therefore, both are selected as candidates for training the surrogate model \( \hat{\Psi} \) and subjected to a refined training process. This process includes Features Selection, using PFI, and hyperparameter tuning by GSkFCV. For the MLP model, the tuning process involves testing different activation functions (\textit{relu}, \textit{logistic}, and \textit{tanh}) as well as varying the number of hidden layers and the number of neurons within each layer. For the XGB-DTC model, different values are tested for the learning rate (to control the degree of shrinkage applied to feature weights), the maximum tree depth, and the subsample ratio. Table~\ref{tab: mlp_vs_xgb} summarizes the key features of the models that achieved the highest \( F_\beta \) scores in the GSkFCV and their corresponding performance metrics. Both models achieve a high F-beta score, with only a minimal difference between them. The data-driven MCDM implementation will be demonstrated using the XGB-DTC, as it performs slightly better.
% \begin{figure}
% \centering
% \includegraphics[width=\linewidth]{figures/models_comparison_class.pdf}
% \caption{Caption}
% \label{fig: model_comp_class}
% \end{figure}
% \begin{table*}[]
% \caption{}
% \label{tab: mlp_vs_xgb}
% \begin{tabular}{cccccc}
% \toprule
% Model & \begin{tabular}[c]{@{}c@{}}GSkFCV\\ Hyperparameters\end{tabular} & $F_\beta$ & Precision & Recall & Accurcay \\ \midrule
% MLP & \begin{tabular}[c]{@{}c@{}}activation function: \textbackslash{}textit\{relu\}\\ 1 hidden layer with 87 neurons\\ ( = number of inputs)\end{tabular} & 0.9758 $\pm$ 0.0027 & 0.9725 $\pm$ 0.0075 & 0.9715 $\pm$ 0.0209 & 0.9741 $\pm$ 0.0102 \\
% XGB & & 0.9769 $\pm$ 0.0037 & 0.9683 $\pm$ 0.0108 & 0.9848 $\pm$ 0.0083 & 0.9781 $\pm$ 0.0062 \\ \bottomrule
% \end{tabular}
% \end{table*}
\begin{table}[]
\caption{Best models hyperparameters and scores.}
\label{tab: mlp_vs_xgb}
\begin{tabular}{ccc}
\toprule
Model & \begin{tabular}[c]{@{}c@{}}GSkFCV\\ Hyperparameters\end{tabular} & $F_\beta$ \\ \midrule
MLP & \begin{tabular}[c]{@{}c@{}}Activation function: \textit{relu}\\ 1 hidden layer with 87 neurons\\ ( = number of inputs)\end{tabular} & 0.9758 $\pm$ 0.0027 \\ \midrule
XGB-DTC & \begin{tabular}[c]{@{}c@{}}Learning rate = 0.3\\ Maximum depth of trees = 6\\ Subsamples ratio= 100\%\end{tabular} & 0.9769 $\pm$ 0.0037 \\ \bottomrule
\end{tabular}
\vspace{-0.2cm}
\end{table}
\subsubsection{Training of $\hat{\Gamma}$}
To compute the stability performance indicators, a separate regression model is trained for each combination of indicator type (\(\mathcal{H}_{2}\)-norm or \(\mathcal{K}\)), signal (\(y \in \{f, V_{DC}\}\)), and CCRC (\(\in \mathcal{R}\)). This approach ensures that each regression model is specifically tailored to its target variable, maximizing accuracy. Consequently, each model is trained independently using the most suitable regression technique among Dummy regressor (that predicts the average value of the target in the training data), Linear Regression, Ridge Regression, DT Regressor (DTR), Neural Network (NN) based on MLP, %\cite{scikit-learn},
Extreme Gradient Boosting DTR (XGB-DTR). % \cite{Chen:2016:XST:2939672.2939785}.
PFI is applied only when it enhances the model accuracy. During the training process, %Winsorization is applied to handle extreme values beyond the 95th percentile.
values are winsorized to the 95th percentile. Model accuracy is evaluated both individually (omitted here for brevity) and in an aggregated manner based on their respective tasks, i.e., models predicting the same indicator for all CCRCs. In the latter case, results are summarized in Table~\ref{tab: rgr_r2}, showing that models predicting the \(\mathcal{K}\) indicators achieve very high accuracy, while those predicting the \(\mathcal{H}_2\)-norm indicators show improved accuracy when points out from Winsorization range are neglected. Fig.~\ref{fig: ind_acc_zoom} compares the predicted and actual indicator values for a test data set, highlighting in the zoomed regions that the regression models for \(\mathcal{H}_2\)-norm indicators achieve better accuracy for lower values. This is particularly relevant, as these lower ranges are of highest interest for obtaining reliable predictions during the execution of the data-driven MCDM algorithm.
\begin{table}[]
\caption{Regression models scores.}
\label{tab: rgr_r2}
\begin{tabular}{ccccc}
\toprule
& $\mathcal{H}_{2,f}$ & $\mathcal{H}_{2,V_{DC}}$ & $\mathcal{K}_{f}$ & $\mathcal{K}_{V_{DC}}$ \\ \midrule
$R^2$ & 0.3357 & 0.6666 & 0.9993 & 0.9998 \\
\begin{tabular}[c]{@{}c@{}}$R^2$ on\\ winsorized range\end{tabular} & 0.7022 & 0.9906 & 0.9998 & 0.9999 \\ \bottomrule
\end{tabular}
\vspace{-0.3cm}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{figures/indicators_accuracy_with_zoom.pdf}
\caption{Predicted vs. exact indicators values.}
\label{fig: ind_acc_zoom}
\vspace{-0.6cm}
\end{figure}
\vspace{-0.5cm}
\subsection{Data-driven MCDM Algorithm Implementation}
%The data-driven MCDM algorithm is designed for online power system operation, enabling the optimal selection of CCRCs to maintain system performance despite deviations from day-ahead generation and demand forecasts.
This section demonstrates the implementation of the data-driven MCDM algorithm in a simulated real-world application. To achieve this, two operating scenarios are considered: one based on the day-ahead operation forecast and the other reflecting intra-day forecast updates that approximate actual system conditions. %The description of these operating scenarios is provided in Appendix~\ref{appendix: appendix}.
Using the day-ahead forecast, the system operator can determine a CCRC assignment schedule by running the MCDM algorithm with exact models, as sufficient computational time is available. Once the intra-day forecast is available, the operator can then re-schedule the CCRC assignment using the proposed data-driven MCDM algorithm.
The effectiveness of the data-driven re-scheduling approach is assessed by comparing its accuracy to the results that would be obtained if the \textit{exact} models were used in the MCDM algorithm for the same operating points. Additionally, the improvement in stability performance achieved through dynamic CCRC re-scheduling is evaluated against the alternative of maintaining the day-ahead CCRC assignment. Finally, an additional scenario is considered in which no MCDM algorithm is used, and CCRC changes are minimized to the extent necessary to ensure system stability. This comparison highlights the stability benefits of a dynamic CCRC assignment. The sequence of CCRC assignments obtained using each scheduling method is compared in Fig.~\ref{fig: CCRC_schedules}, while Fig.~\ref{fig: methods_ind_boxplot} compares their performance based on stability indicator values.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/CCRC_schedules_noweigths.pdf}
\caption{CCRCs assignment schedules comparison.}
\label{fig: CCRC_schedules}
\vspace{-0.5cm}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/CCRC_shedules_boxplot_indicators.pdf}
\caption{Distributions of the indicators values obtained with the different CCRCs assignment schedules.}
\label{fig: methods_ind_boxplot}
\vspace{-0.3cm}
\end{figure}
%\vspace{-0.1cm}
The \textit{Exact re-schedule} serves as the target sequence of CCRC assignments to be achieved. The \textit{No MCDM} approach ensures stable operating points in 100\% of the cases by using only two CCRCs, among the ones that show the highest percentage of stable cases in the training data set. However, the \textit{No MCDM} approach matches the target CCRC assignments only in 5.2\% and leads to very poor stability performance, as shown by the stability indicator distribution in Fig.~\ref{fig: methods_ind_boxplot}. The \textit{Day-ahead re-schedule} matches the target sequence in 43.7\% of cases. However, for the remaining operating points, it does not always guarantee stability, resulting in instability in 14.5\% of cases. In the box plot distribution of Fig.~\ref{fig: methods_ind_boxplot}, these unstable cases have been excluded, yet the performance still remains worse than that of the \textit{Data-driven re-schedule}. The \textit{Data-driven re-schedule} aligns with the target sequence in 89.6\% of cases while also achieving similar stability performance, making it the most effective approach among the alternatives.
According to the \textit{Data-driven re-scheduling}, the sequence of control roles assigned to each IPC is shown in Fig.~\ref{fig: sched_IPC}. In addition to IPC-E, which is constrained to AC-GFM by its operating principle, the other IPCs of the DC-grid 2 also maintain fixed control roles — one in AC-GFM and the other in DC-GFM. The IPCs of the DC-grid 1, however, are the ones subjected to control role switches. The data-driven MCDM was solved by setting the maximum number of allowed simultaneous control switches equal to 1 ($\gamma^* = 1$). This constraint is satisfied in all cases except one, where it is necessary to relax it and allow up to 2 simultaneous switches.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/CCRC_schedules_IPC_noweigths.pdf}
\caption{IPCs control roles assignment sequence obtained with the data-driven MCDM CCRCs schedule.}
\label{fig: sched_IPC}
\vspace{-0.7cm}
\end{figure}
%In the scenario where the MCDM is not applied, CCRCs changes are allowed only to maintain system stability, without aiming to improve stability performance. Since no single CCRC can guarantee stability across all operating points derived from the intraday forecast, a minimum of two CCRCs is required. CCRCs $X_C$ 3 and $X_C$ 2 are employed. $X_C$ 3 is selected because it stabilizes the highest percentage of operating points (87.2\%) in the training dataset $\mathcal{D}_\Upsilon$. $X_C$ 3 fails to ensure stability for two operating points, for which $X_C$ 2 is used. Concerning the stability performances, this approach results in very poor performances as shown by the indicators distributions comparison of Figure\ref{fig: methods_ind_boxplot}.
% \subsubsection{Data-driven CCRC Re-scheduling}
% This scenario emulates the real-world application of the data-driven MCDM algorithm. It assigns the CCRC that provides the best stability performance for each operating point, as determined by the intra-day forecast, which adjusts the day-ahead forecasted operating points.
% \subsubsection{Exact CCRC Re-scheduling}
% This scenario is used to evaluate the accuracy of the data-driven CCRC assignment. It involves running the MCDM algorithm for the intra-day forecasted operating points using exact models in the algorithm.
% \subsubsection{Exact Day-ahead Schedule}
% Here, the CCRC assignment determined by executing the MCDM algorithm with day-ahead forecasted operating points is directly applied to the intra-day forecasted points. Since there is more time available in the day-ahead stage, exact models can be used to construct the performance matrix.
% When equal weighting is applied, the resulting CCRCs assignment sequence is shown in Figure~\ref{fig: sol_now_dd}. Over the 96 instances, 88.5\% has a correct assignment, matching the one that would be obtained using the exact models.
% \begin{figure}
% \centering
% \includegraphics[width=\linewidth]{figures/solutions_noweights_datadriven.pdf}
% \caption{Caption}
% \label{fig: sol_now_dd}
% \end{figure}
% \begin{figure}
% \centering
% \includegraphics[width=\linewidth]{figures/indicators_results.pdf}
% \caption{Caption}
% \label{fig: ind_sol}
% \end{figure}
% \begin{table}
% \caption{Comparison of Scheduling Methods}
% \label{tab: schedulmethod}
% \centering
% \begin{tabularx}{\columnwidth}{l c *{8}{X}}
% \hline
% & \multirow{2}{*}{Stable} & \multicolumn{2}{c}{$\mathcal{H}_{2,f}$} & \multicolumn{2}{c}{$\mathcal{H}_{2,V_{DC}}$} & \multicolumn{2}{c}{$\mathcal{K}_{f}$} & \multicolumn{2}{c}{$\mathcal{K}_{V_{DC}}$} \\ \cline{3-10}
% & Cases & min & max & min & max & min & max & min & max \\ \hline
% No MCDM & 100\% & 0.187 & 0.347 & 2.18 & 30.15 & 0.017 & 0.027 & 0.121 & 0.368 \\
% Day-ahead & 85.4\% & 0.171 & 0.336 & 0.15 & 24.71 & 0.017 & 0.030 & 0.013 & 0.363 \\
% Data-driven & 100\% & 0.177 & 0.341 & 0.21 & 12.64 & 0.017 & 0.026 & 0.017 & 0.346 \\
% Exact & 100\% & 0.177 & 0.336 & 0.18 & 10.81 & 0.017 & 0.026 & 0.016 & 0.347 \\ \hline
% \end{tabularx}
% \end{table}
\vspace{-0.3cm}
The computational performance of the proposed data-driven MCDM is evaluated in comparison to the \textit{exact} MCDM. During the execution of the MCDM for the 15-minute operating points over a 24-hour period, the computing time required for solving each instance was recorded for both methods. The PC used is equipped with an Intel Core i7-11390H (4 cores / 8 threads, 3.6 GHz), 16 GB RAM. Fig.~\ref{fig: comptime} illustrates the computing time as a function of the number of CCRC alternatives (i.e., the CCRCs in $\mathcal{R}'_\cap$) considered per instance. For the data-driven MCDM, the total computing time is further decomposed into the time required solely for solving the MCDM using the \textit{surrogate} models, and the additional time needed to verify the solution with the \textit{exact} model. On average, the data-driven MCDM achieves a computing time reduction of 59.8\% compared to the \textit{exact} MCDM. Moreover, unlike the \textit{exact} MCDM, its computing time is considerably reduced for any number of CCRC alternatives at each execution. This characteristic makes it a promising approach for application to larger systems with a higher number of IPCs and CCRCs.
\vspace{-0.5cm}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\linewidth]{figures/computing_time_V2.pdf}
\caption{Computing time}
\label{fig: comptime}
\vspace{-0.7cm}
\end{figure}
% Generated by IEEEtran.bst, version: 1.14 (2015/08/26)
\begin{thebibliography}{10}
\providecommand{\url}[1]{#1}
\csname url@samestyle\endcsname
\providecommand{\newblock}{\relax}
\providecommand{\bibinfo}[2]{#2}
\providecommand{\BIBentrySTDinterwordspacing}{\spaceskip=0pt\relax}
\providecommand{\BIBentryALTinterwordstretchfactor}{4}
\providecommand{\BIBentryALTinterwordspacing}{\spaceskip=\fontdimen2\font plus
\BIBentryALTinterwordstretchfactor\fontdimen3\font minus \fontdimen4\font\relax}
\providecommand{\BIBforeignlanguage}[2]{{%
\expandafter\ifx\csname l@#1\endcsname\relax
\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}%
\typeout{** loaded for the language `#1'. Using the pattern for}%
\typeout{** the default language instead.}%
\else
\language=\csname l@#1\endcsname
\fi
#2}}
\providecommand{\BIBdecl}{\relax}
\BIBdecl
\bibitem{ramosready4dc}
J.~Ramos, I.~Jahn, M.~Moritz, J.~Moore, and C.~Kjaer, ``Ready4dc whitepaper wg4: Long-term view for hvdc technology,'' Available online: https://www. ready4dc. eu, Tech. Rep., 2023.
\bibitem{9729620}
D.~Groß, E.~Sánchez-Sánchez, E.~Prieto-Araujo, and O.~Gomis-Bellmunt, ``Dual-port grid-forming control of mmcs and its applications to grids of grids,'' \emph{IEEE Trans. Power Del.}, vol.~37, no.~6, 2022.
\bibitem{9714816}
Y.~Li, Y.~Gu, and T.~C. Green, ``Revisiting grid-forming and grid-following inverters: A duality theory,'' \emph{IEEE Trans. Power Syst.}, vol.~37, no.~6, pp. 4541--4554, 2022.
\bibitem{10508461}
H.~Xin, C.~Liu, X.~Chen, Y.~Wang, E.~Prieto-Araujo, and L.~Huang, ``How many grid-forming converters do we need? a perspective from small signal stability and power grid strength,'' \emph{IEEE Trans. Power Syst.}, vol.~40, no.~1, pp. 623--635, 2025.
\bibitem{cui2024control}
G.~Cui, Z.~Chu, and F.~Teng, ``Control-mode as a grid service in software-defined power grids: Gfl vs gfm,'' \emph{IEEE Trans. Power Syst.}, 2024.
\bibitem{10017135}
J.~A. Soler, D.~Groß, E.~P. Araujo, and O.~G. Bellmunt, ``Interconnecting power converter control role assignment in grids with multiple ac and dc subgrids,'' \emph{IEEE Trans. Power Del.}, vol.~38, no.~3, 2023.
\bibitem{10443561}
J.~Arévalo-Soler, D.~Groß, E.~Prieto-Araujo, and O.~Gomis-Bellmunt, ``Scheduling role assignment and system reconfiguration in ac/dc segmented systems interconnected by power converters,'' \emph{IEEE Trans. Power Del.}, vol.~39, no.~3, pp. 1524--1537, 2024.
\bibitem{floudas2008encyclopedia}
C.~A. Floudas and P.~M. Pardalos, \emph{Encyclopedia of optimization}.\hskip 1em plus 0.5em minus 0.4em\relax Springer Science \& Business Media, 2008.
\bibitem{BHOSEKAR2018250}
A.~Bhosekar and M.~Ierapetritou, ``Advances in surrogate based modeling, feasibility analysis, and optimization: A review,'' \emph{Computers \& Chemical Engineering}, vol. 108, pp. 250--267, 2018.
\bibitem{kundur1994power}
P.~Kundur, \emph{Power System Stability and Control}, ser. The EPRI Power System Engineering Series.\hskip 1em plus 0.5em minus 0.4em\relax New York, NY, USA: McGraw-Hill, 1994.
\bibitem{skogestad2005multivariable}
S.~Skogestad and I.~Postlethwaite, \emph{Multivariable feedback control: analysis and design}.\hskip 1em plus 0.5em minus 0.4em\relax john Wiley \& sons, 2005.
\bibitem{PRIETOARAUJO2017}
E.~Prieto-Araujo, A.~Junyent-Ferré, C.~Collados-Rodríguez, G.~Clariana-Colet, and O.~Gomis-Bellmunt, ``Control design of modular multilevel converters in normal and ac fault conditions for hvdc grids,'' \emph{Electric Power Systems Research}, vol. 152, pp. 424--437, 2017.
\bibitem{triantaphyllou_2000}
E.~Triantaphyllou, \emph{Applied optimization}.\hskip 1em plus 0.5em minus 0.4em\relax Kluwer, 2000.
\bibitem{gomis2020principles}
O.~Gomis-Bellmunt, E.~S{\'a}nchez-S{\'a}nchez, J.~Ar{\'e}valo-Soler, and E.~Prieto-Araujo, ``Principles of operation of grids of dc and ac subgrids interconnected by power converters,'' \emph{IEEE Trans. Power Del.}, 2020.
\bibitem{10202947}
F.~Rossi, S.~C. Dilmé, J.~A. Soler, O.~G. Bellmunt, and E.~P. Araujo, ``Converters control roles selection for operating hybrid ac/dc grids with enhanced small-signal stability performance,'' in \emph{2023 IEEE PowerTech}.
\bibitem{6876017}
A.~Lex, N.~Gehlenborg, H.~Strobelt, R.~Vuillemot, and H.~Pfister, ``Upset: Visualization of intersecting sets,'' \emph{IEEE Trans. Vis. Comput. Graph.}, vol.~20, no.~12, pp. 1983--1992, 2014.
\bibitem{rossi2022data}
F.~Rossi, E.~P. Araujo, M.~C. Ma{\~n}e, and O.~G. Bellmunt, ``Data generation methodology for machine learning-based power system stability studies,'' in \emph{2022 IEEE ISGT-Europe}.
\bibitem{sammut2011encyclopedia}
C.~Sammut and G.~I. Webb, \emph{Encyclopedia of machine learning}.\hskip 1em plus 0.5em minus 0.4em\relax Springer Science \& Business Media, 2011.
\end{thebibliography}