Abstract: The strong nonlinearity presented by the runoff process constrains the prediction performance of the existing hydrological models.The artificial intelligence methods such as deep learning with strong nonlinear fitting ability can break through the current bottleneck to a certain extent. To effectively extract the nonlinear time-varying feature information of runoff sequences and improve the accuracy of runoff simulation and the multi-step-head forecasting performance,the ForecastNet based runoff prediction model with time-varying structure has been established. The Yajiang River Basin in the upper reaches of the Yalong River is taken as a case study,and comparative analyses among the ForecastNet,traditional hydrological model of SWAT（Soil and Water Assessnent Teol）,and neural network models of RNN（Recurrent Neural Network）,LSTM（Long Short-Term Memory）and RNN-LSTM are carried out. The results show that the ForcastNet model has strong applicability in long term prediction,and can effectively improve the accuracy of runoff simulation and multi-step-ahead forecasting,thus providing technical support for high-precision real-time runoff prediction.