Optimal Energy System Scheduling Based on Safe Deep Reinforcement Learning Algorithms

The utilisation of distributed energy resources (DERs) can improve energy efficiency. However, due to the uncertainty and system complexity of renewable energy sources, there are difficulties in energy scheduling for DERs systems. Current scheduling algorithms based on Deep Reinforcement Learning (DRL) have the advantages of real-time learning and dynamic environment adaptation, but still have some problems, such as the difficulty of strictly satisfying the balance between supply and demand. In addition, it is difficult to establish an accurate environment model for DERs, which leads to insufficient training of DRL algorithms, which may not be able to make safe decisions in the face of unknown situations. To address these issues, this paper proposes an approach that uses a pre-trained neural network to learn the sensitivity of energy system supply and demand constraints, builds a security layer in combination with a security adjustment policy, and integrates it into a deep deterministic policy gradient (DDPG) algorithm to enhance the security of the policy network. Experimental results show that this approach can monitor and dynamically adjust the power in real time to effectively reduce the security risk.