[Objective] Improving the prediction accuracy of medium- and long-term hydrological forecast is of great significance for water resources scheduling, flood control and drought relief, and agricultural production. This study aims to select reliable, efficient, and practical hybrid machine learning models to improve forecasting performance for highly irregular, complex nonlinear, and multi-scale variable medium- and long-term hydrological forecasts, providing new approaches for enhancing hydrological forecast accuracy in changing environments. [Methods] To improve the accuracy of hydrological forecasts, based on the measured monthly runoff series at Wanxian Station in the Three Gorges Reservoir area, the mutual information method was used to screen forecasting factors. Then, Long Short-Term Memory (LSTM) models optimized by the Whale Optimization Algorithm (WOA), Grasshopper Optimization Algorithm (GOA), and Sparrow Search Algorithm (SSA) were established. Combined with Time-Varying Filtered Empirical Mode Decomposition (TVF-EMD), Complementary Ensemble Empirical Mode Decomposition (CEEMDAN), and Variational Mode Decomposition (VMD), multiple hybrid prediction models were established. Their prediction performance was evaluated using five indicators: mean absolute error (MAE), root mean square error (RMSE), Nash-Sutcliffe efficiency coefficient (NSE), mean absolute percentage error (MAPE), and correlation coefficient (R). [Results] The forecast factor scheme selected by the mutual information method provided optimal model input, with a lag of 15 months achieving the maximum mutual information value and minimum MASE, representing the best input configuration. Among the three single machine learning models, LSTM and SVM outperformed BP, with LSTM and SVM showing similar performance. LSTM was preferred due to its sensitivity to temporal sequences, enabling better handling of nonlinear runoff prediction, and was thus used in coupling with different methods for runoff forecasting. The hybrid models following the “decompose-reconstruct” strategy outperformed single LSTM models: the VMD-LSTM model improved the NSE of the test set by 0.12 compared with the single LSTM model, exceeding CEEMDAN-LSTM and TVF-EMD-LSTM. Further integration with robust optimization algorithms enhanced accuracy: the VMD-SSA-LSTM model outperformed VMD-LSTM, VMD-GOA-LSTM, and VMD-WOA-LSTM, showing superior adaptability, generalization, and overall predictive performance. [Conclusions] Machine learning models provide effective runoff forecasting methods for regions with limited hydrological and meteorological data. The approach of combining forecasting factor screening, data preprocessing, and integrating robust optimization algorithms with the model can further improve the accuracy of a single hydrological forecasting model. The established VMD-SSA-LSTM model achieved test period performance of MAE=32.65, RMSE=43.44, NSE=0.95, MAPE=12.9%, and R=0.98, representing the highest accuracy among compared models. This model meets practical production and daily life requirements and can provide a reference for water resource management and industrial and agricultural production in the studied basin.