A Learning-based Hybrid System Approach for Detecting Contingencies in Distribution Grids with Inverter-Based Resources

TL;DR

This paper introduces a machine learning-based stochastic hybrid system framework for real-time detection of unobservable contingencies in distribution grids with inverter-based resources, improving grid reliability.

Contribution

It develops a novel SHS modeling approach combined with time series learning for detecting contingencies that traditional sensors cannot identify.

Findings

96% detection accuracy in simulations

Effective detection of both physical and measurement contingencies

Applicable to active distribution networks with inverter-based resources

Abstract

This paper presents a machine-learning based Stochastic Hybrid System (SHS) modeling framework to detect contingencies in active distribution networks populated with inverter-based resources (IBRs). In particular, this framework allows detecting unobservable contingencies, which cannot be identified by normal sensing systems. First, a state-space SHS model combining conventional and IRB-based resources is introduced to formulate the dynamic interaction between continuous states of distribution networks and discrete contingency events. This model forms a randomly switching system, where parameters or network topology can change due to contingencies. We consider two contingency classes: (i) physical events, such as line outages, and (ii) measurement anomalies caused by sensor faults. Leveraging multivariate time series data derived from high-frequency sampling of system states and network outputs, a time series-based learning model is trained for real-time contingency detection and classification. Simulation studies, carried out on the IEEE 33-bus distribution system, demonstrate a 96% overall detection accuracy.