Trustworthiness Evaluation of AI Training Data for Power Grid Dispatch Based on Wasserstein Distance

Trustworthy artificial intelligence (AI) depends on the trustworthiness of training data, but existing research on data trustworthiness is limited, particularly in the emerging field of trustworthy AI for power grids. The distribution shift captures the difference between data samples and the true distribution, which directly affects the trustworthiness of AI models in applications. However, effective methods to evaluate distribution shifts without access to the true distribution are still lacking. To address this, we propose a novel approach to evaluate the distribution shifts using the ambiguity set radius in distributionally robust optimization (DRO), which is proportional to the distribution shift. To overcome limitations in existing radius selection methods, we introduce a new data-driven radius selection method. Simulation results demonstrate the effectiveness of the proposed radius selection method in quantifying distribution shifts under the Wasserstein metric. Furthermore, AI models trained for power grid dispatch with more trustworthy datasets exhibit superior performance, highlighting the practical value of our method.