A dual-phase ensemble framework for enhancing reservoir inflow forecasting

Abstract. Growing reliance on reservoir storage makes reservoir inflow forecasting essential to water supply across many regions. We present a dual-phase ensemble modeling framework to enhance the accuracy and timeliness of flow forecasting. The first phase of the ensemble consists of primary members developed based on calibrations of a semi-distributed hydrologic model against selected events across diverse hydroclimatic conditions. Through parameter perturbation and optimization, the primary members are further tuned to address the uncertainty in initial soil moisture conditions, resulting in a set of derived models. The primary and derived models together form a complete ensemble prediction band. The second phase involves statistical ensemble band using the quantile regression forests approach. The outputs of ensemble members are converted into a probabilistic range around the ensemble mean to better accommodate modeling uncertainties. The inflow forecast further extended by three days using linear regression, enhancing operational value in specific catchments. This ensemble approach was applied to two major reservoirs in the Nueces River Basin in Texas, USA. Ensemble members were established using the HEC-HMS model, and the ensemble was validated and tested with seven-day precipitation forecast for events in summer 2025. As measured by weighted performance metrics, the ensemble forecast outperformed the National Water Model, providing a more accurate estimation of reservoir inflows while preserving the probabilistic characteristics of the weather inputs. These results highlight the benefits of ensemble forecasts for effectively dealing with the uncertainties and complexities in watershed hydrologic responses and associated data inputs, providing an improved basis to support forecast-informed reservoir operations.