Pruning-Based Knowledge Transfer Method for Power System Scheduling

In order to enhance the learning optimization efficiency of deep reinforcement learning algorithms in the scenario of unit shutdown and maintenance, this study delves into knowledge transfer issues arising from the reduction of target task state and action vector dimensions compared to the source task. A dimensionality reduction transfer method for deep reinforcement learning based on neural network pruning is proposed. Firstly, a method for calculating the information richness of neural network nodes is introduced, utilizing L1 norm and information entropy. Subsequently, guided by the richness of information, the significance of each neural network node in the source task to the target task is analyzed, enabling effective network pruning. Furthermore, a mapping relationship between the neural networks of the target task and the source task is established, facilitating the transfer of deep reinforcement learning network parameters. Finally, through numerical simulations, the effectiveness of the proposed method is validated by comparing the transfer effects of different transfer methods, pruning techniques, and pruning rates.