Abstract: The probabilistic forecast performance of a single runoff model is highly dependent on its deterministic forecast accuracy, while traditional ensemble forecasting methods generally fail to account for the constraints imposed by initial hydrological states. The probabilistic distributions generated by multiple models using the Copula-based Hydrologic Uncertainty Processor (CHUP) were integrated into the Bayesian Model Averaging (BMA) framework. A coupled CHUP–BMA ensemble probabilistic forecasting method was developed that simultaneously incorporates multi-model forecast distributions and initial hydrological state constraints. Three inflow forecasting models, including a dual-attention deep learning, an extreme gradient boosting, and an operational forecasting model, were developed for the Three Gorges reservoir interval basin to construct an ensemble probabilistic runoff forecasting. The results indicate that when the deterministic forecasting accuracy of individual models is relatively low, the correction performance of the CHUP method is limited. The proposed CHUP–BMA method effectively corrects forecast errors, reducing the mean absolute error by an average of 7% relative to the best-performing member. Moreover, it achieves superior probabilistic forecasting performance by narrowing the forecasting intervals while maintaining or improving the coverage rate of confidence intervals. These findings demonstrate that the CHUP–BMA method can provide reliable risk information to support flood control and reservoir operation decisions in the Three Gorges reservoir.