A hybrid framework for forecasting monthly reservoir inflow based on machine learning techniques with dynamic climate forecasts, satellite-based data, and climate phenomenon information

In this study, we developed and evaluated a hybrid framework for reservoir inflow forecast. This framework is unprecedented, which integrates new quasi-globally available observation-, satellite-, or model-based datasets using machine learing models to forecast inflow at the local scale. Under this framework, we compared random forests, gradient boosting machine, extreme learning machine, M5-cubist, elastic net, as well as their multi-model ensemble using Bayesian model averaging (BMA), and tested contributions from different input datasets, including retrospective forecast (reforecast) from florecast oriented low ocean resolution (FLOR) dynamic climate model, satellite-based hydrologic products, and climate phenomenon information. The performance was evaluated using Kling-Gupta efficiency (KGE) and correlation coefficient (R) in two headwater reservoirs, Harris reservoir in the humid Alabama-Coosa-Tallapoosa river basin and the Navajo reservoir in the arid Upper Colorado River Basin (UCRB). The results showed that for the Harris reservoir, the BMA combining five models with antecedent inflow and satellite-based hydrologic information as model inputs provided the best performance (KGE = 0.66, R = 0.76). For the Navajo reservoir, the gradient boosting machine model with all variables combined as input showed the best performance (KGE = 0.76, R = 0.83). Satellite-based soil moisture and evaporation consistently showed significant contributions to the inflow forecast. Benefits from climate indices and FLOR reforecast varied by locations, with more benefits coming from climate indices than FLOR potential evaporation reforecast at the Navajo reservoir in UCRB. Given the global coverage of the model inputs, our approach can be potentially applicable to improve reservoir inflow forecasts in different regions of the world.