Accurate runoff forecasting is of great significance for flood control, drought prevention, reservoir scheduling, and ecological protection. To explore the applicability of deep learning networks combined with signal processing techniques in runoff forecasting, an ICEEMDAN-VMD-CNN-LSTM daily runoff forecasting model for the flood season was developed. First, the original runoff series was decomposed using the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN). Then, the complex series was further decomposed using Variational Mode Decomposition (VMD) to reduce data complexity. Next, each mode component was input into a Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) combined model to extract local features of the data and capture long-term dependencies of the time series. Finally, the predicted values were reconstructed to obtain the final prediction results. Using the measured daily runoff data from the Hekou station in the Diaojiang basin as an example, the results showed that the ICEEMDAN-VMD-CNN-LSTM achieved testing MAE and NSE of 5.232 m3/s and 0.977, respectively, demonstrating excellent forecasting accuracy.