Short-Term Water Level Prediction for Long-Distance Water Diversion Projects Using Data-Driven Methods with Multi-Scale Attention Mechanism

Short-term water level prediction plays a crucial role in the control of gates for long-distance water diversion projects, enabling decision makers to make informed gate control decisions and ensure the safe and stable operation of water resources allocation projects. In this study, a short-term water level prediction coupling model for long-distance water diversion projects, It comprises three components: the Wavelet threshold denoising (WTD) method, Reversible Instance Normalization (RevIN), and Crossformer. The model acquires data that is more conducive to the model’s mining of the water level change law through low-frequency information extraction and data smoothing. Through multi-scale learning, time dimension and feature dimension correlation synchronization analysis, the full excavation of water level change law is realized. The generalization ability and robustness of the proposed method are verified by ablation test, comparison test with benchmark model, application of different forecast periods and different control gates. The results indicate that the coupling of WTD, RevIN, and Crossformer can improve the accuracy of short-term water level prediction, and the data smoothing method utilizing RevIN is more effective in improving the model. Compared with LSTM and Transformer, the accuracy of the proposed method is significantly improved, with a maximum increase of 82.25%, which is more in line with the short-term (6 ~ 24h) water level prediction of long-distance water diversion projects. This study can provide theoretical and technical support for the safe dispatching of long-distance water diversion projects, and provide useful reference for time series prediction of other similar water conservancy projects.