Online probabilistic early warning for power system dynamic congestion events with enhanced trustworthiness

The development of renewables and electricity demand induces the growth of inter-area power transfer of power systems, potentially causing dynamic congestion events (DCEs) of transmission interfaces that threaten system stability and adequacy. The lack of procedures for perceiving real-time DCEs often leads to inadequate time for decision-making and resource mobilization. In this paper, a trustworthiness-enhanced learning approach is proposed for online probabilistic early warning (PEW) of DCEs, promoting the system proactivity in resisting risks. An explainable PEW network is first built utilizing tensorized neural blocks and tandem attention, enabling more explicit variable-wise information encoding. Next, an interface for static covariates is designed to fuse heterogeneous data and improve prediction performance. Then, an ensemble confidence calibrator is developed to modify the mapping between logits and congestion posterior probabilities via negative log-likelihood minimization, outputting more credible multi-class probability estimates. Test results verify the efficacy of the proposed trustworthiness-enhanced PEW method. By combining class predictions with calibrated probability estimations, more comprehensive decisions can be made.