耦合CNN-SE Attention-GRU机制的中长期径流智能预测研究OA

In order to improve the effect of the mid-long term runoff forecasting,a deep learning model combining convolutional neural networks(CNN)model,gated recurrent unit(GRU)and squeeze-and-excitation attention(SE Attention)was proposed.The model extracts spatial features of hydrometeorological data through CNN,learns long-term temporal patterns via GRU,and enhances the influence of key information using SE Attention,thereby constructing a CNN-SE Attention-GRU hybrid model for medium-and long-term runoff forecasting.The study was conducted in the Futou Lake basin,employing hydrometeorological data from 1971 to 2017 to perform multivariate runoff predictions at the Jinkou and Shihaoqiao stations.The proposed CNN-SE Attention-GRU model was compared with single models(LSTM,GRU),hybrid models(CNN-LSTM,CNN-GRU),and a composite model with the same architecture(CNN-SE Attention-LSTM).The results demonstrate that the CNN-SE Attention-GRU model exhibiting optimal performance,the model attained root mean square error(RMSE)values of 17.86 and 5.18,mean absolute error(MAE)values of 10.50 and 3.10,mean absolute percentage error(MAPE)values of 15.03％and 15.15％,coefficient of determination R2 values of 0.765 and 0.874,and Nash efficiency coefficient(NSE)values of 0.79 and 0.88,respectively.