Controlled Comparison of Machine Learning Models for Fault Classification and Localization in Power System Protection

Researchers conducted a controlled comparison of machine learning models for fault classification and localization in power systems, finding that advanced nonlinear models achieve 98%+ accuracy at 10ms decision windows while topology-dependent factors significantly influence localization performance across different grid segments.

Modern power grids face unprecedented complexity as distributed renewable energy sources and inverter-based resources reshape traditional protection architectures. This research addresses a critical gap in machine learning applications for power system protection by establishing standardized benchmarking methodology. Rather than comparing fragmented studies with inconsistent datasets and assumptions, the authors created controlled conditions using identical sensing parameters, decision horizons, and validation approaches on a common electromagnetic transient dataset. The findings reveal important distinctions between fault classification and fault localization tasks, each with different information requirements and performance characteristics. High-capacity machine learning models demonstrate remarkable speed in fault classification, reaching above 98% F1 scores within 10 milliseconds—a timeframe critical for protection relay operation. Conversely, fault localization presents a more nuanced challenge, with top performers stabilizing at approximately 10% normalized error while performance variance across different grid segments suggests that topology and network structure matter as much as algorithm selection. This topology-dependent difficulty undermines assumptions that longer decision windows automatically improve accuracy, indicating that localization fundamentally depends on understanding specific grid characteristics. The research provides essential quantitative reference points for utilities and protection equipment manufacturers evaluating machine learning integration. Grid operators considering ML-based protection schemes can now benchmark candidate models against established performance tiers, reducing deployment uncertainty. These findings particularly matter as inverter-based resources proliferate, creating transient phenomena that conventional protection schemes struggle to classify rapidly and accurately.

• →Nonlinear ML models achieve 98%+ fault classification accuracy within 10ms, meeting protection-critical timing requirements.
• →Fault localization accuracy plateaus at approximately 10% error regardless of decision window length, indicating topology-dependent limitations.
• →Performance varies significantly across different grid segments, suggesting network structure influences ML model effectiveness more than temporal context alone.
• →Controlled benchmarking on common datasets enables standardized comparison of ML protection schemes across different studies and implementations.
• →Findings support ML adoption for power system protection while identifying specific performance boundaries and grid-dependent constraints.