Runoff forecasting plays a vital role in flood control, drought mitigation, and water resources management. To address the issues of mode mixing and complex residuals associated with traditional decomposition methods, this study proposes a runoff forecasting model—IVCL (ICEEMDAN-VMD-CNN-LSTM)—that integrates dual decomposition and deep learning. First, the improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is applied to preliminarily decompose the original runoff series to reduce complexity. Then, the most complex sub-series undergoes secondary decomposition using Variational Mode Decomposition (VMD) to extract deeper features. Subsequently, a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) model is used to model and predict each sub-series. The predicted sub-series are summed and reconstructed to obtain the final runoff forecast series. The model is validated using observed daily runoff data during the flood season from three hydrological stations with varying climatic, surface, and hydrological characteristics: Hekou Station in Guangxi, Zhengyangguan Station in Anhui, and Lanzhou Station in Gansu. The results show that the IVCL model outperforms other benchmark models across most evaluation metrics, exhibiting lower errors, more accurate extreme value predictions, and higher stability. Moreover, the order of decomposition significantly affects model performance, and a properly designed decomposition scheme can effectively enhance runoff forecasting accuracy.