A multivariate prediction framework for flood-induced substation damage based on generative adversarial network and MPformer-based two-stage model

Abstract

Frequent extreme rainstorms have significantly increased the flooding risk, threatening the security and stability of electrical substations. The process of flood-induced substation damage is complex and nonlinear, challenging traditional predictive methods. Therefore, a novel predictive framework is proposed for flood-induced substation damage. This framework uses a generative adversarial network (GAN)-based model to capture complex data relationships and generate realistic samples, which mitigates training data imbalance. A multivariate predictive Transformer network (MPformer), integrating three improved modules: time embedding, multi-factor fusion encoding, and attention-based encoder, is proposed to capture temporal dependencies and complex interactions between influencing factors and flood-induced damage. Based on MPformer and sensitive cost learning, a two-stage integrated model is designed to reduce the problem of sample imbalance further and realize the simultaneous prediction of the substation damage probability, severity, and time. The experimental results show that the GAN-based method is superior to the traditional method in terms of sample balancing, and the MPformer-based two-stage model outperforms the mainstream model, with a 12.30 % average increase in F1 score for probability prediction and reductions of 38.56 % and 45.31 % in RMSE for severity and time predictions, respectively. A case study shows that the proposed method can offer reliable pre-disaster prediction.