The frequency and recurrence of extreme rainfall events has resulted in many cities facing serious flooding prob-lems.The ability to accurately and efficiently predict changes in water levels at urban waterlogging sites is an important component in the prevention and control of urban flooding.In order to effectively improve the accuracy and efficiency of ur-ban storm water prediction,this paper establishes an urban storm water prediction model based on convolutional neural net-work(CNN)-gated recurrent unit(GRU)-attention mechanism(ATT).First,CNN and GRU are used to extract the lo-cal spatial features and deep temporal features of the water level data,and then ATT is introduced to enhance the memory of the key information in the rainfall sequence,and finally the water level prediction of urban water accumulation points is completed.The model was validated by using the measured water level at a waterlogged site in Kaifeng and compared with previous CNN-GRU,ATT-CNN-LSTM and CNN-LSTM models.The results show that the loss function of the model is converged at epoch=20,and the value of the loss function is finally stabilized at 0.000 2,which is a good convergence effect.In addition,compared with the other three models,the CNN-GRU-ATT model has the best performance in terms of prediction accuracy,with the root mean square error of 1.39％,the mean absolute percentage error of 4.32％and the coef-ficient of determination of 0.995 4.The model also has the shortest training and prediction time and the highest operational efficiency,which indicates that the model can accurately and efficiently predict the water level changes at the water accu-mulation points.The model will provide an effective scientific basis for early warning of storm water flooding and the formu-lation of flood control and drainage plans.
维普
万方数据