Volt/VAR optimization for photovoltaic-storage-charging station high-permeability power distribution networks: A data-knowledge hybrid driven reinforcement learning method

Abstract

The integration of a high proportion of electric vehicles and photovoltaic systems has increased the uncertainty of power flow state transitions in power distribution networks (PDN). To address the intermittency of power and the complexity of operational constraints caused by the integration of photovoltaic-storage-charging station systems into the PDN, this paper proposes a Volt/VAR optimization (VVO) framework driven by a data-knowledge hybrid approach. In this framework, a novel DistFlow-based soft actor-critic (DFSAC) algorithm is first introduced, which constructs an expert knowledge safety layer based on the DistFlow equations to describe the coupling relationship between the reinforcement learning actions and PDN voltage, ensuring the safety of the VVO strategy. A novel data augmentation technique and a linearized power flow calculation method are then proposed, enhancing the diversity and completeness of the state-action pairs for the reinforcement learning agent and increasing the speed of interaction with the environment. Finally, numerical experiments using real PDN operational data and the IEEE 34-bus system are conducted. The results show that the proposed method outperforms other state-of-the-art VVO methods, demonstrating better performance and good robustness under extreme source-load power conditions. The DistFlow linearization safety layer also shows good scalability in large-scale real power systems. Additionally, the proposed data augmentation method improves the DFSAC performance by approximately 75 % to 85 %, and the linearized power flow calculation method increases the overall training speed of the DFSAC agent by about 3 times.