Optimal Load Shedding During Service Restoration in Electrical Secondary Distribution Network Based on Reinforcement Learning

Abstract

Increased stress in traditional power systems results in blackouts due to voltage instability attributed to a mismatch between available capacity and load demand, especially in distribution networks. Service restoration schemes are designed to return power supply to the affected parts of the networks. The availability of insufficient supply is a complex problem that requires operational experience or an automatic system. The stochastic nature of load demand significantly impacts service restoration as it results in increased restored demand in case a fault occurs during off-peak hours and helps reduce overload if the fault occurs during peak hours. The study adopts an experimental design methodology to develop the Reinforcement Learning-based service restoration algorithm considering the stochastic nature of load demand. Three reinforcement learning models were used to develop the optimal load shedding model, including Actor-Critic (A2C), a Deep Q Network (DQN) and Proximal Policy Optimization (PPO2), and compared to maximize restored customers, satisfaction of operational constraints, and balancing of power supply and demand. The Particle Swarm Optimization (PSO) algorithm, a metaheuristic algorithm, was also implemented to compare with the proposed approach. The proposed solution has been tested using data from a real electrical secondary distribution network. The proposed solution considered the stochastic nature of load demand, resulting in more restored customers. The computation time during restoration has been improved by 69.8% compared to the metaheuristic approach.