Deep reinforcement learning differential protection system for electric power networks

Abstract

A method for diagnosing a fault in an electric power distribution network includes measuring three-phase current signals from local and remote terminals of a protected zone on the network, and computing differential current signals based on the measured three-phase current signals. The differential current signals are preprocessed using a filter to smooth the differential current signals. A maximal overlap discrete wavelet transform is applied on the smoothed differential current signals to obtain a plurality of detail coefficients. Observation signals comprising the plurality of detail coefficients are provided to a deep reinforcement learning (DRL) agent comprising a temporal convolution attention-based neural network (TCAN). The TCAN DRL agent is trained using a proximal policy optimization (PPO) method to propose a trip action corresponding to a trip or no trip command for a fault condition. Responsive to receiving a trip command, a trip signal is transmitted to a circuit breaker.