Reinforcement learning based operational control of active distribution networks

Abstract

The dissertation explores voltage control in active distribution networks (ADNs) and proposes a reinforcement learning-based Voltage/Var control (VVC) strategy utilizing PV inverters to mitigate voltage fluctuation and reduce network energy loss in ADNs. The methodology involves framing the VVC problem as a Markov Decision Process (MDP) and implementing deep reinforcement learning (DRL) algorithms, specifically Deep Deterministic Policy Gradient (DDPG) and Actor-Attention-Critic (AAC).To better assess the voltage regulation performance, a load-weighted voltage deviation index is adopted. In the framework, two objectives, i.e., network energy loss and voltage deviations, are considered. The research demonstrates the effectiveness of these methods through simulations conducted on an IEEE 33-bus distribution system, comparing centralized and decentralized control strategies. The results indicate that the centralized DDPG approach outperforms others, achieving faster and more effective voltage regulation while minimizing active power losses.