A novel optimized coupled runoff model is developed based on the concept of “decomposition-prediction-reconstruction”

Runoff refers to the quantity of water that flows over the surface of the ground from precipitation, snowmelt, or other sources, playing a crucial role in water resource management. Accurate runoff prediction in water resource modeling aids in managing water resources, forecasting floods and droughts, optimizing reservoir operations, and formulating reasonable water use policies. Advanced modeling techniques, enable more precise capture of the temporal characteristics of runoff, thereby improving the accuracy and reliability of predictions and playing a significant role in ensuring the sustainable use of water resources. To enhance the precision of runoff forecasts, a novel approach has been introduced. This methodology integrates the Adaptive Noise Complete Ensemble Empirical Modal Decomposition (CEEMDAN) with a Bidirectional Long Short-Term Memory (BiLSTM) model, further optimized through the application of the Sparrow Search Algorithm (SSA). The coupling of the SSA-BiLSTM model has led to substantial optimization of several parameters, including the number of iterations, the quantity of hidden layer nodes, and the learning rate. The resulting model, termed the CEEMDAN-SSA-BiLSTM, offers an advanced and integrated solution for predicting both short-term and long-term runoff scenarios, thereby facilitating more effective water resource management and environmental preservation within the basin. Daily runoff data from 2016 to 2022 were analyzed at four hydrological stations—Huayuankou, Jiahetan, Gaocun, and Lijin. The approach involved using 80% of the daily runoff data for training and 20% for prediction. The performance of the CEEMDAN-SSA-BiLSTM model was compared against several other models, including LSTM, BiLSTM, and CEEMDAN-BiLSTM, using various evaluation indices. The error results for the CEEMDAN-SSA-BiLSTM model compared to the aforementioned models are as follows: For the HuaYuankou station, the RMSE is 97.42, the MAPE is 5.46%, the MAE is 56.9, and the NSE is 0.96. At the JiaHetan station, the RMSE is 950.36, the MAPE is 6.76%, the MAE is 59.33, and the NSE is 0.96. For the GaoCun station, the RMSE is 92.38, the MAPE is 5.53%, the MAE is 54.85, and the NSE is 0.97. Finally, for the LiJin station, the RMSE is 88.31, the MAPE is 6.49%, the MAE is 52.68, and the NSE is 0.95. The ultimate results indicate that the CEEMDAN-SSA-BiLSTM model demonstrates superior accuracy in forecasting daily runoff, with fewer errors relative to the other models.