An Uncertainty-Aware, Machine Learning-Enabled Reservoir Inflow Forecast Model

Forecasts of reservoir inflows are needed for hydropower operators to make climate-resilient decisions. However, streamflow forecasts are challenging due to multiple influential factors, probabilistic nature, and consequent uncertainties. Recently, researchers have started using machine learning (ML) techniques by leveraging a deluge of diverse data and ML’s predictive power. But current ML applications do not consider the non-stationarity of hydroclimatic data, cannot understand climate influence, and lack uncertainty quantification (UQ) on the forecasts. To address these challenges, we develop an uncertainty-aware, ML-enabled forecast method. It includes an interpretable ML model and a UQ capability. The ML model predicts reservoir inflows several days ahead using diverse input data, including long-term, non-stationary, historical observations of meteorological forcing and streamflow. Furthermore, the ML model calculates the importance of each input to reservoir inflow prediction, which helps decision-makers understand hydroclimatic impacts as well as identify influential variables for data collection. Additionally, the UQ capability produces confidence intervals on inflow forecasts to support decision-making. We demonstrate the proposed methods to several multipurpose reservoirs in the southeastern and western United States. Results indicate that the method can reasonably predict streamflow variation and quantify predictive uncertainty. Additionally, we observe that as the lead time increases, the prediction accuracy decreases with increasing uncertainty. The method may help operators better understand long-term climate change impacts on water management and hydropower production and support climate-resilient decision-making under uncertainty.