Adaptive Model Selection for Electrical Fault Detection Using Reinforcement Learning

Electrical faults exhibit significant variations based on their type, primarily influenced by fluctuations in current and voltage levels, as well as harmonic distortions. Low impedance faults (LIFs) generate large fault currents, making them relatively easier to detect, whereas high impedance faults (HIFs) produce subtle signals that often evade detection. Consequently, each fault type necessitates a tailored detection approach aligned with its distinct electrical profile. This paper presents a dynamic fault detection system that integrates historical fault data with a Reinforcement Learning (RL) framework. The proposed system operates in two stages: first, it detects the occurrence of a fault using a preliminary fault identification module; second, it identifies the specific fault type and the circuit on which the fault occurs by continuously learning from historical data and adapting its model selection based on fault characteristics. The RL agent evaluates various machine learning (ML) models tailored for LIFs and HIFs, optimizing model selection through feedback derived from a reward function. This adaptability enables the system to refine its detection strategy in response to evolving fault patterns effectively. The adaptive model selection approach ensures enhanced detection accuracy, minimizes false positives and false negatives, and maintains robustness amid changing grid conditions. Ultimately, the proposed fault detection system evolves with the operational environment, offering improved classification capabilities, and increased system resilience and reliability, especially in scenarios of wildfires caused by ground faults and faulty circuits.