Boosting Reinforcement Learning for Tertiary Voltage Control by Transfer Learning from a Supervised Power Flow Simulation Task

This work explores how transfer learning can improve reinforcement learning for tertiary voltage control, which is a simulation-intensive process. The model is pre-trained over the supervised learning of a power flow simulator, as a way of incorporating physics inside the model. Two transfer strategies are proposed and compared against a transfer-free baseline. The case60nordic test case, which provides diversified operating conditions, including topological variations, is used to assess performance. Results indicate that fine-tuning the pre-trained model can effectively improve performance or reduce training time on the target task. This study emphasizes the potential of transfer learning for accelerating the training of power grid control downstream tasks.