Hybrid Secondary Decomposition and CNN BiLSTM Model for Nonlinear Monthly Runoff Prediction

Addressing the strong nonlinear and non-stationary characteristics of monthly runoff series, which pose challenges to single decomposition or single-model prediction approaches, this paper proposes a hybrid prediction model (ISCBL) combining ICEEMDAN-SSA secondary decomposition with CNN-BiLSTM deep learning. The secondary decomposition is guided by sample entropy to identify and further decompose the most complex intrinsic mode function (IMF), thereby enhancing feature stationarity. The model is validated using monthly runoff data (1971–2023) from the Lanzhou hydrological station in the upper Yellow River, with data split into training (1972–2013) and testing (2013–2023) sets. The results show that the ISCBL model achieves a Nash-Sutcliffe efficiency coefficient (NSE) of 0.9733 and a mean absolute percentage error (MAPE) of 8.35% on the test set, significantly outperforming baseline and primary decomposition models. This study verifies the effectiveness of the sample entropy-guided secondary decomposition strategy in extracting multi-scale features and the advantage of CNN-BiLSTM in capturing spatiotemporal dependencies. The proposed framework provides a robust and accurate method for complex hydrological sequence prediction, offering valuable technical support for water resource management in the Yellow River Basin.