Time-varying Bayesian Model Averaging for ensemble probabilistic forecasting of reservoir inflow floods

Individual flood forecasting models struggle to achieve consistently optimal performance across diverse forecasting scenarios. Although traditional Bayesian Model Averaging (BMA) integrates information from multiple models, its fixed global weights restrict the exploitation of the local strengths of individual member models. To address this, a sliding window mechanism was introduced to dynamically update the weights of member models. Consequently, a Sliding Window BMA (SW-BMA) ensemble probabilistic forecasting framework for predicting reservoir inflow floods, integrating five machine learning algorithms, was developed and applied to the Wuxikou Reservoir. The results demonstrated that the SW-BMA achieved the highest prediction accuracy across all seven lead times during the testing period. Furthermore, at longer lead times of 4 to 7 hours, the forecasting error of the SW-BMA grew more gradually, effectively mitigating the rapid error accumulation typically associated with single models as lead times extend. In terms of probabilistic forecasting, while maintaining a confidence interval coverage rate comparable to that of BMA, the SW-BMA reduced the continuous ranked probability score by an average of 27.7% and narrowed the prediction interval width by an average of 24.2% across the seven lead times, thereby achieving a reliable quantification of forecast uncertainty. These findings can offer robust decision-making support for reservoir flood control operations.