A Wasserstein Distributionally Robust Learning Approach for Power Grid Dispatch

Within the context of power grid digitalization and intelligent transformation, Artificial Intelligence (AI)-driven power dispatching holds immense promise, yet its trustworthiness hinders real-world applications. Mainstream methods such as end-to-end learning offer computational efficiency, but lack model robustness and explainability. Therefore, we investigate the potential of Distributionally Robust Optimization (DRO) in power grid AI applications, which offers a principled methodology for ensuring robustness in the face of uncertainties. However, DRO and its learning-based extension, Distributionally Robust Learning (DRL), can both suffer from conservatism, especially under data pollution. To alleviate this conservatism, we propose a novel DRL method. Our approach innovatively incorporates a tighter, dimension-independent radius selection to reduce ambiguity set size during its construction, and leverages Partial Optimal Transport theory to handle contaminated training data. Simulation results validate the effectiveness and potential value of our method.