Abstract: In order to effectively extract the information features of runoff time series and improve the stability of the high-dimensional nonlinear fitting ability and prediction performance of the runoff prediction model, the convolutional neural network(CNN),BiLSTM and attention mechanism were combined to construct a runoff combination model of CNN-BiLSTM-attention.The runoff of Hankou station in the middle reaches of the Changjiang River Basin was simulated and verified by this model.The error characteristics of monthly runoff simulated by five runoff prediction models, namely BiLSTM,CNN,BiLSTM+attention, CNN-BiLSTM and CNN-BiLSTM-attention were analyzed.FA-SSA,GWO,BAO were used to optimize the hyperparameters of the novel model, including the number of convolutional nuclei, number of BiLSTM hidden layer neurons, number of fully connected hidden layer neurons, dropout layer, batch size, and learning rate, respectively, to explore the effects of the three optimization algorithms on the monthly runoff prediction performance of the novel model.The results show that the prediction error of BiLSTM-attention is the largest, followed by BiLSTM,and the overall prediction accuracy of CNN-BiLSTM-attention is the highest.The novel model can more effectively and accurately capture key information and further master the rule of runoff timing changes, and the predicted runoff value is in good agreement with the actual value.FA-SSA optimization algorithm is superior to GWO and BAO algorithms, which is conducive in optimizing the hyperparameter values of the novel model and further improving the prediction accuracy of the model.