面向拓扑鲁棒变量的可解释电力系统拥塞事件预测

Network congestion is a frequent challenge that power systems need to handle in real-time operation. Though learning-based congestion event prognosis (CEP) is a promising way for early warning, its inadequate adaptability to topology alterations and lack of interpretability may hinder its practical applications. This article first proposes a feature combinatorial optimization (FCO) method to explore a variable set that can provide robust contributions for CEP on different topologies, where an information-assisted scheme is designed to facilitate the FCO efficiency and result. Then, using topologically robust variables, an explainable CEP model is built based on the tensorized learning network and mixture attention mechanism, where the contribution from individual variables can be explicitly tracked via variable-wise hidden states. Next, the data of static covariates are encoded into models to improve CEP performance. These components finally drive a CEP model with decent topology robustness and explainability. Numerical results validate the efficacy of the proposed method, indicating that it can effectively mitigate the degradation in CEP performance caused by topology alterations by at least ∼32% and achieves ∼95% average CEP accuracy even encountering unseen topological changes.