Evaluation of Machine Learning Assisted Reservoir Operation Models for Long-Term Water Management Simulation
Abstract
The representation of reservoir operations is crucial for streamflow modeling in regulated river basins. The operation of multipurpose reservoirs with competing constraints is complex and dependent upon the decisions of experienced operation engineers, which cannot be fully captured by the generic rule-based reservoir management model (RMM). While machine learning (ML) models, Long Short-Term Memory (LSTM) in particular, have shown promising results in short-term reservoir release predictions, given their data driven nature, they could yield a physically infeasible solution in long-term water management simulation, especially under changing climate and water usage conditions. In this study, we explore several different strategies to simulate long-term water management using both RMM and ML models. Two major multipurpose reservoirs in the southeastern United States, Allatoona Lake and Lake Sidney Lanier that serve the greater Atlanta metropolitan area, are selected in this study. First, a standalone RMM was developed to simulate daily release and storage during Water Year (WY) 19802015 using long-term reservoir inflow observations and reservoir pertinent information. Next, a standalone LSTM model was trained based on reservoir inflow and meteorological observations to simulate reservoir release. A mass balance calculation was then performed to calculate the corresponding reservoir storage. We consider three hybrid models by combining the RMM and LSTM, one using RMM output as an additional LSTM input, another using LSTM as the initial release estimate in RMM, and the third combining the first two strategies. We evaluate the performance of the hybrid models and compare to the standalone RMM and LSTM to understand their strengths and limitations. In general, our results indicate that hybrid models can capture the best features of both RMM and LSTM and provide a more robust simulation capacity for use in high-resolution, large-scale streamflow modeling applications.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H31B..07G