Identifying the Hydrometeorological Decision Factors Influencing Reservoir Releases over the Upper Colorado Region using a Long Short-Term Memory Network
Abstract
Due to the capability of managing water resources, reservoirs play imperative roles in water supply, power generation, flood control, drought relief, and other function goals. Effective reservoir operation can bring significant benefits in alleviating the water scarcity and potential impacts caused by climate change. Because of complex hydrological processes and varying conditions from human demands, the daily discharge predicted from the reservoir models deviates from the practical release value and can't be directly used to make the reservoir operation decision. Since the reservoir historical data contain both natural factors (storage, inflow, temperature, and precipitation) and human operating management experience, using date-driven machine learning models can help provide a fast and accurate reservoir release prediction. Furthermore, the changing climate has increased the intensity and frequency of extreme events, where dry regions/seasons get drier and wet regions/seasons get wetter. It is of necessity to check the influence of meteorological information, such as temperature and precipitation, on the forecast of reservoir outflow. In this study, we predict 30 reservoir daily discharges over the Upper Colorado Region using long short-term memory (LSTM) neural networks with 7-day delayed information of the four input variables, including reservoir inflow and storage, precipitation, and temperature. We find that with 7-day delayed information, the daily releases for all reservoirs reach a good statistical performance with the following range: 0.832<= CORR <= 0.995, 0.691 <= NSE <= 0.99, 0.52 <= KGE <= 0.984, 0.213 <= RMSE <= 9.349, 0.101 <= RSR <= 0.555, -34.15% <= PBIAS <= 22.98%. Furthermore, the observational meteorological data can help solve the over- or under-estimation of reservoir outflow at low-flow regimes. Through comparing prediction performance using 7-day with 2-day delayed information, we find that providing longer delayed information can gain substantial improvement in the daily discharge prediction accuracy. In this study, we demonstrate that using machine learning model LSTM can accurately mimic human's decisions on reservoir daily release, which can assist the management of water resources and the early-warning for extreme weather events.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H21C..04F