To extend the lead time and improve the accuracy of medium-to-long-term runoff forecast, teleconnection factors were incorporated as physical determinants of runoff formation alongside antecedent precipitation-runoff relationships. An RF-Informer model for monthly runoff forecast was developed by integrating the random forest (RF) model for predictor selection with the Informer model, which has demonstrated superior performance in long-sequence time-series forecasting. The proposed model was applied to forecast monthly inflow runoff of the Lianghekou, Jinxi, and Ertan reservoirs in the Yalong River Basin. Results indicate that the integration of teleconnection factors effectively extends the lead time while enhancing the accuracy for forecasting basin-scale monthly runoff. By resolving nonlinear couplings between predictors that linear methods overlook, the RF-driven feature selection advances the physical interpretability and predictive skill of runoff forecast; compared with the RF-LSTM, RF-SVM, and RF-BP neural network models, the RF-Informer model achieves the minimum forecast error and highest precision in monthly runoff forecast.

Based on Vine Copula, which can accurately describe the correlation between high-dimensional variables, and considering the spatial correlation of short-term runoff forecasting errors, a risk estimation model for short-term power generation scheduling of cascade reservoirs was constructed. The model was applied to the Xiluodu, Xiangjiaba, and Three Gorges reservoirs in the upper reaches of the Yangtze River, and the short-term power generation scheduling risks of single reservoirs and cascade reservoirs caused by runoff forecasting errors were analyzed. The results show that the joint distribution based on C-vine Copula can better describe the error characteristics of daily runoff forecasting at Pingshan Station, Zhutuo Station, Cuntan Station, and Wulong Station. As the adjustable safety range of the reservoir increases, the risk rates of insufficient power generation and water abandonment in a single reservoir decrease, and the risk rate of insufficient power generation, joint risk rate of water abandonment, and co-occurrence risk rate of cascade reservoirs decrease. That is to say, the larger the regulating capacity of the reservoir, the smaller the risk it bears.