Multi-Timescale Distribution Network Optimization Based on Deep Transfer Reinforcement Learning

With the development of the power grid to the direction of intelligence and digitization, distributed generations have gradually become an important part of the distribution network. However, the randomness of distributed generations' output influenced by surrounding environment, together with the high-dimensional distribution network scheduling model, brings great challenges to the formulation and implementation of distribution network scheduling optimization. Therefore, this paper proposes multi-timescale distribution network optimization based on deep transfer reinforcement learning. Firstly, considering difference of devices' action speed, the coordination of the discrete device and the continuous device is realized through cooperation between different timescale. Therefore, action plan for discrete device is determined by stochastic optimization in the day-ahead phase. Secondly, a deep reinforcement learning optimization model of continuous devices is established in intraday phase and trained according to short term prediction data to lay the model foundation for knowledge transfer. Then, the trained model is used to do parameter transfer. In this way, the scheduling strategy can be adjusted according to the real-time output data, so as to ensure a more stable and optimized operation of the distribution network. The experimental results show that the proposed method can adjust the scheduling strategy quickly, the algorithm is more convergent, and the exploration is faster.