End-to-end learning for stochastic preventive dispatch of renewables-rich power systems in abnormal weather conditions

Power systems are currently increasingly threatened by abnormal weather conditions, which are exacerbated by global warming. In particular, the proliferation of renewable energy sources (RES), which are quite variable, might cause power systems to have insufficient energy supply to cope with the risk of severe power imbalances in such conditions. In this paper, a preventive power system dispatch model is proposed that allows power systems with large RES (specifically wind energy) installations to mitigate excessive load curtailments under abnormal weather conditions. Compared to the conventional predict-then-optimize (PTO) framework of power system dispatch that relies on accurate RES forecasting, the proposed end-to-end learning framework mitigates the impact of forecasting errors on decision-making, which reduces the dependence on the accuracy of RES forecasting and effectively enhances the resilience of the power system. The proposed framework consists of a probabilistic RES forecasting module, a preventive dispatch module, and a resilience-oriented assessment module. A backpropagation mechanism based on Karush-Kuhn-Tucker conditions is proposed to unify the dispatch optimization problem as deep neural networks, which ensures the reliability of optimal solutions as well as provides necessary gradients for backpropagation, successfully bridging the information gap among the three modules. The proposed model is updated iteratively in an online implementation scheme composed of the decision-making part and the reflection & training part. The results of numerical experiments on the IEEE 6- and 118-bus systems verify the effectiveness and superiority of the proposed model in maintaining the resilience of RES-based power systems.