An Adaptive Hybrid Framework for Runoff Prediction Using Temporal-Spatial Transformers and Dynamic Decomposition-Reconstruction Optimization

Hydropower and watershed management depend on projected runoff dependability and accuracy. because to climate changes and human involvement, runoff production is difficult to anticipate because to its non-linearity, stochasticity, and temporal unpredictability. In response, this paper offers an Enhanced Multi-Stage Hybrid Model for Runoff Prediction using spatiotemporal feature extraction and dynamic decomposition. This paper recommends using EEMD and VMD for thorough runoff signal decomposition and an integrated reconstruction phase for feature clarity. The Temporal-Spatial Transformer Network (TSTN) idea is used to extract temporal and spatial relations of runoff data reliably and effectively. The adaptive Attention-Augmented Mechanism (AAM) improves temporal feature recognition. The model uses Bayesian hyperparameter tuning for optimal prediction performance and resistance. The proposed method and existing models like LSTM, CEEMDAN-TCN-Attention, CEEMDAN-VMD-LSTM-Attention, and others were compared to actual monthly runoff data from 1979–2018 at Pingshan and Yichang Hydrological Stations for validation experiments. These results showed significant improvements with the new test set MAE: 923.45, RMSE: 912.67, MAPE: 14.56, and NSE: 0.942 for Pingshan station and 976.33, RMSE: 1078.29, MAPE: 15.28, and 0.938 for Yichang station. Such results validate the model's benefits, such as higher predicted accuracy, stability, and statistical reliability. As hydrological situations got more complicated, the Enhanced Multi-Stage Hybrid Model for runoff forecasting improved in interpretability, efficiency, applicability, and nonlinear flexibility.