Reinforcement Learning Models for Adaptive Low Voltage Power System Operation

The fast-paced installation of electric vehicles (EV) and photovoltaic units (PV) in low voltage distribution grids calls for sophisticated control strategies that ensure a safe grid operation. In view of this necessity, the goal of this paper is to eliminate voltage and current bound violations while maximizing EV and PV penetration in the low voltage grid by optimally controlling the EV charging processes and curtailing PV generation. The paper presents a reinforcement-learning (RL) solution using the Soft-Actor Critic (SAC) algorithm that can be generalized in any grid since it does not require a grid model and is based on limited measurements. The developed RL agent is evaluated using datasets in both low and high time-resolutions showing that it is capable of ensuring a safe grid operation by reducing violations by more than 98% with a low penalty in EV and PV curtailment compared to the optimal solution. Also, it is shown that the algorithm performs well in previously unseen grids. Finally, it is compared with two heuristic baselines demonstrating its superior performance.