Deep Reinforcement Learning Approach to Optimize Distribution Network with Prediction Embedding

As the integration of renewable energy sources into the power grid continually accelerates, traditional methods of devising operational strategies through predictive optimization fall short in responding rapidly to random fluctuations. Consequently, deep reinforcement learning (DRL) approaches, enabling adaptive system operation, have emerged at the forefront of research. Notwithstanding, traditional reward metrics, heavily reliant on cost, struggle to objectively assess the efficacy of these strategies and fail to leverage short-term precise predictive data to boost training and application efficiency. This paper addresses the optimization of economic operation within a distribution network (DN) using DRL, specifically examining the influence of state prediction information on strategy training. Initially, we refine the Markov Decision Process (MDP) model used for DN optimization, treating uncontrollable state transitions independently and introducing a corresponding policy compensation structure. Subsequently, we create an interaction model for the DN predicated on Optimal Power Flow (OPF), and based on this, we establish the specific form of incorporating predictive information into the reward function. Finally, we assess the effectiveness of embedding predictive information in enhancing the training process, under varying algorithms and reward configurations.