Advancing subseasonal reservoir inflow forecasts using an explainable machine learning method

Abstract

Study RegionUpper Colorado River Basin and Great Basin in the United StatesStudy FocusAccurate subseasonal reservoir inflow forecasts and understanding the influence of hydrometeorological forcings on these forecasts are crucial for improving water resources management. Machine learning (ML) techniques, such as long short-term memory (LSTM) networks, perform well for short-term inflow forecasts but have deficiencies in subseasonal forecasts and lack interpretability. To address these limitations, we propose an explainable ML method that integrates an encoder–decoder LSTM (ED-LSTM) network to improve long-term reservoir inflow forecasts and a gradient-based explanation method to quantify the importance of individual hydrometeorological forcings and their interactions on inflow forecasts.New Hydrological Insights for the RegionThe ED-LSTM model outperforms the standard LSTM in the 30-day inflow forecasts at all 30 reservoirs. At the 1-day lead time, ED-LSTM produces NSEs exceeding 0.75 at 29 reservoirs; at the 15-day lead time, about half of reservoirs maintain this high-accurate performance, and when forecasting 30 days ahead, ED-LSTM achieves NSEs exceeding 0.5 at most reservoirs. The variable importance identifies past inflow and temperature as crucial drivers for predicting inflow dynamics. When considering interactions between hydrometeorological forcings, precipitation contributes significantly to inflow forecasting through its interaction with temperature and historical inflow. The proposed method enhances subseasonal reservoir inflow forecasts and the understanding of the impact of hydrometeorological factors, which supports decision-making in reservoir operations.

Keywords

• Subseasonal forecasting
• Explainable machine learning
• Variable importance
• Encoder-Decoder LSTM networks
• Reservoir inflow