An Ensemble Graph and Series Decomposition Model for Water Level Prediction at Both Real and Virtual Stations in the Yangtze River

The middle reaches of the Yangtze River are important water sources and economic zones in China. Accurate water level forecasts are critical for water resource management and flood mitigation. However, most existing water level prediction models do not consider the unique river-lake interaction in the middle reaches of the Yangtze River and virtual stations. This study proposes an ensemble graph and series decomposition model for water level prediction at both real and virtual stations in the Yangtze River. We present a module for water level prediction in regions without monitoring stations, filling gaps in traditional water level monitoring systems. Our model utilizes graph convolutional network (GCN) to model spatial dependencies between monitoring stations, and graph attention network (GAT) to capture river-lake interaction. Time series are decomposed into trend and seasonal series, followed by Transformer to extract long-range dependencies. In addition, a hierarchical ensemble method combining hierarchical clustering, stacking, and AdaBoost is introduced to enhance the robustness and generalization of our method. Numerical experiments conducted on real historical water level data in the middle reaches of the Yangtze River in China and the Mackenzie River in Canada validate the effectiveness of our model. Compared to existing studies, our model achieves an average reduction of 35.88% in MAE and 26.28% in RMSE in 12-timestep water level prediction.