Online Voltage Control Strategy: Multi-Mode Based Data-Driven Approach for Active Distribution Networks

Active distribution network (ADN) is faced with significant challenges, including frequent and fast voltage violations, due to the increased integration of intermittent renewable energy resources. This paper proposes a two-stage multi-mode voltage control strategy based on a deep reinforcement learning (DRL) algorithm, designed to alleviate voltage violations in ADN and minimize network power loss. In the first stage, a DRL algorithm, the soft actor-critic (SAC), is introduced to determine the hourly dispatch of on-load tap changers and capacitor banks, ensuring voltage security during the day-ahead stage. A multi-mode voltage regulation strategy is then proposed to obtain real-time dispatch of PV inverters, aiming to save energy and enforce voltage constraints under various conditions. The real-time voltage regulation problem is formulated as a Markov decision process and solved using a multi-agent SAC integrated with an attention mechanism. All agents undergo centralized offline training to learn the optimal coordinated voltage control strategy, then make decentralized online decisions based on locally available information only. The effectiveness of the proposed approach is confirmed through extensive testing on the IEEE 33-bus distribution system, with simulation results conclusively demonstrating its ability to address voltage violation challenges.