Data-Driven Robust Voltage/VAR Control Using PV Inverters in Active Distribution Networks

Modern power systems are heading towards a sustainable future with rapidly increasing penetration of distributed renewable energy sources such as solar photovoltaics (PV), especially in active distribution networks (ADNs). However, temporal and spatial variations and intermittency of renewable power generation bring challenges on voltage control. As controllable VAR sources, PV associated inverters can supply flexible and fast-responsive reactive power to achieve voltage/VAR control (VVC) in the ADNs. Conventional VVC mainly relies on rule-based, mathematical or heuristic methods, which can be inefficient or even infeasible as the system becomes large and when complex uncertainty conditions are considered. To this end, this paper proposes a deep reinforcement learning (DRL) based optimization model for the inverter-based VVC under uncertainties. The DRL agent learns to optimally determine inverter reactive power output setpoints through exploration in the virtual environment and neural network training, with multiple objectives of minimizing bus voltage deviations, PV active power curtailments and network power losses. Temporally and spatially uncertain PV power generation and loads are fully considered in the proposed DRL model. A numerical case study on the IEEE 123-bus test system has indicated the VVC decisions are efficiently obtained and are robust against uncertainty realization.