Early disaster analysis by prediction of deformation trend on pumped storage dam based on intelligent transformer model

Pumped-storage dams are essential infrastructure for energy management and storage. However, they are vulnerable to deformation due to sudden environmental changes and cyclic stress. Existing models struggle to capture complex long-term deformation patterns and provide timely early warnings. In this study, we propose DEFT-Net (Deformation Forecasting Transformer Network). This novel hybrid model combines the advantages of the dynamic Equipoise CNN (E-CNN) for spatial anomaly detection with a lightweight Transformer architecture, utilizing Squeezed Head Attention (SHA) and a Bidirectional Processing Layer (BPL) for enhanced temporal dependency modeling. Unlike previous CNN-transformer hybrid models that focus mainly on accuracy at the expense of efficiency, DEFT-Net introduces SHA to reduce memory and computation costs while maintaining precision significantly, and integrates BPL to capture bidirectional temporal dependencies. DEFT-Net is designed explicitly for multimodal input fusion and optimized for low-latency applications. The study simulation is performed using the PSPS dam dataset from China. The results demonstrate that the suggested DEFT-Net notably reduces MAE by 22%, improves anomaly detection accuracy to 90.3%, and achieves R2 of 0.95, showing the model’s effectiveness in early disaster detection and real-time deformation prediction.