Machine learning based binding contingency pre-selection for AC-PSCOPF calculations

Aug 5, 2026

Upload by Anda

Proposes a supervised machine learning oracle to predict the set of binding contingencies for an alternating-current preventive security-constrained optimal power flow (AC-PSCOPF) solver. The approach is validated on the Nordic32 benchmark system and employs a steady-state security assessment (SSSA) module to evaluate prediction effectiveness. An offline training dataset is built from solved PSCOPF instances, then an online loop runs PSCOPF only with predicted binding contingencies and iteratively enlarges the set when security violations are detected.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages3

Reading Time8 min

AccessFree to download

I. INTRODUCTION

II. ML-BASED BINDING CONTINGENCY DETECTION

A. Proposed multi-label classification formulation

Index Terms—machine learning,multi-label classification,pre-ventive security-constrained optimal power flow,steady-statesecurity assessment # I.INTRODUCTION The AC-PSCOPF is a nonlinear optimization problem aim-ing at finding an optimal operating point of a power systemthat is also secure with respect to a set of contingencies四.Forlarge power system models with tens of thousands of buses,lines,and contingencies,this leads to nonlinear optimizationproblems with billions of variables and constraints,which arecurrently not solvable accurately in a tractable way. A way to make PSCOPF calculations more tractable wouldbe to use an oracle to guess the subset of binding contingen-cies;the latter are contingencies that at the PSCOPF solutionhave at least one non-zero Lagrange multiplier associatedto their branch loading or bus voltage limits.Solving thePSCOPF with only these contingencies explicitly modelledmay be much more efficient and would yield a solution thatwould also be secure for the non-binding contingencies. In order to identify an as small as possible subset of con-tingencies containing all the binding ones,heuristic iterativeapproaches have been proposed in the literature 2.Thesemethods progressively grow the subset of covered contingen-cies,until the resulting PSCOPF solution is found to be secureagainst all other (not explicitly modelled)contingencies.Theytypically need several iterations and converge to a superset ofthe subset of binding contingencies.In practice however,thenumber of potentially binding contingencies they finally selectremains less than a few tens,even for very large systems 3. While machine learning has been proposed recently toenhance AC-PSCOPF in various ways 4,15,we proposein this letter,for the first time,to use supervised MachineLearning(ML)to predict binding contingencies for the AC-PSCOPF.Our method works in the following way: 1)in off-line mode,we generate a dataset of solvedPSCOPF instances and train an ML-based oracle thatcan predict binding contingencies, The authors acknowledge the funding from the FNRS-Belgium forthe project ML4SCOPF under grant T.0258.20,and the funding fromFNR-Luxembourg in the framework of the project ML4SCOPF(INTER/FNRS/19/14015062).Computational resources have been provided by theCÉCI-FNRS under Grant No.2.5020.11 and by the Walloon Region. *Institut Montefiore,Université de Liège,Liege,Belgium(email:{n.popli, 1.wehenkel}@uliege.be) Luxembourg Institute of Science and Technology,Esch-sur-Alzette,Lux-embourg(email:{elnaz.davoodi,florin.capitanescu}@list.lu) Machine learning based binding contingencypre-selection for AC-PSCOPF calculations Nipun Popli*Elnaz Davoodi'Florin Capitanescu Louis Wehenkel* Abstract—We propose to use machine learning to constructan oracle able to predict the set of binding contingencies foran alternating current preventive security-constrained optimalpower flow(AC-PSCOPF)solver.The method is show-casedon the Nordic32 benchmark system.A steady-state securityassessment (SSSA)module is also used to assess the effectivenessof the learnt binding contingency prediction oracle. 2)in on-line mode, a)we run the PSCOPF with only the contingenciespredicted as binding by the learnt ML-oracle, b)we run SSSA,to check whether some of theremaining contingencies lead to security violations;if this is the case,we include them in the set ofpotentially binding ones and return to step a. 3)we continuously enrich the training set with the set ofbinding contingencies found in on-line mode and werefresh the ML-oracle by off-line retraining. Fig.1:Proposed on-line workflow # II.ML-BASED BINDING CONTINGENCY DETECTION In Fig.1.we present the on-line workflow for evaluating ourclassifier-assisted PSCOPF solver.It begins with computingan OPF solution (i.e.without contingencies)by applying aphysics-based alternating-current solver on the given loadprofile.The OPF solution is then fed into the contingencyclassifier,which works as an orac

FAQs

What problem does the proposed method address in AC-PSCOPF calculations?

It targets the computational burden of large AC-PSCOPF instances by identifying an as small as possible subset of binding contingencies to model explicitly.

How does the ML oracle work in the proposed workflow?

Offline training builds a dataset from solved OPF/PSCOPF instances; online, the oracle predicts binding contingencies, then PSCOPF is run using only those contingencies.

How is security assessed and the contingency set updated?

A steady-state security assessment (SSSA) checks whether the classifier-assisted PSCOPF solution is insecure; any contingencies causing violations are added and the loop continues with an updated potentially binding set.

Aug 5, 2026

Upload by Anda