Simulating Operation Behaviors of Cascade Reservoirs Using Physics-Based Machine Learning Models: A Case Study for Gunnison River Basin
Abstract
Rule-based reservoir management models (RMM) are commonly used to simulate reservoir operation. Predefined operating rules are embedded as hard constraints in RMM, making their applications data-intensive and complex. Further, many ad-hoc decisions made by operators to achieve reservoir-specific objectives are difficult to capture in RMM, especially considering the cascade nature of multi-reservoir systems. Over the last decade, machine learning (ML) based techniques for reservoir operation have become available, and among them, the short-term memory (LSTM) model has demonstrated great potential in assisting various aspects of reservoir operation due to its ability to learn long-term dependencies between the input and output of the network. In this study, we investigate the applicability of physics-based ML models developed by combining RMM and LSTM to predict the behaviors of a nine-reservoir system in the Gunnison River Basin in the Upper Colorado Region of the United States. RMM are developed by gathering reservoir operation information from publicly available long-term time-series data (> 30 years), physical characteristics of the reservoir, and through information provided by the operators. Hybrid models are then implemented to evaluate the improvement in predictability compared to standalone RMM and LSTM for long-term water management simulation using various statistical measures (root-mean-square error, Nash-Sutcliffe efficiency, correlation coefficient, and mean absolute error). Lastly, we explore strategies to incorporate these physics-based ML models in distributed hydrologic models, such as the Advanced Terrestrial Simulator (ATS), to expand our modeling capabilities to address a variety of issues in human-regulated watershed systems.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H45L1536I