Knowledge Distillation Based Dueling DQN for Real-Time Power Grid Dispatching Operation and Control

Deep reinforcement learning (DRL) has proven to be effective for real-time Power grid Dispatching and Control (PDC) but requires intensive computing resources and has low efficiency. Most research employs expertise demonstrations to accelerate the training process, but neglects the rules derived from the experience. To boost the DRL efficiency with PDC rules, a Knowledge Distillation based Dueling Deep Q-Iearning (KD3QN) algorithm is proposed which uses behavioral cloning to lead the agent to learn from PDC rules. Behavioral cloning may introduce bias in the policy, thus knowledge distillation is employed to extract the knowledge instead of directly following the rules. Besides, a balance strategy is proposed to control the interference of rules. To verify the proposed algorithm, extensive experiments have been performed on Grid20p platform with the 14- bus and 36- bus cases. The experimental results demonstrate that the proposed algorithm increases the training efficiency and yields agents with better performance.