Data-driven Generalized Release Policies for Reservoirs in the Tennessee River Basin
Abstract
In highly regulated river basins, reservoir operations, rather than precipitation or snowmelt, tend to dictate streamflow. These operations are driven by anomalies in inflow, such as droughts or floods, seasonal changes in inflow or demand, varying storage rule curves, and the multi-purpose nature of the allocation. Though it is a challenge to include all the information that influences release decisions into a generalized modeling framework, such a framework could help incorporate the release policies into distributed hydrologic models. This work leverages general reservoir characteristics, such as residence time, inflow sources, and the nature of the system - run-of-the-river or storage reservoir - to classify reservoirs and then develop several release parameterizations based on physical variables, such as storage and inflow, using hierarchical regression and regression trees. Within the Tennessee River Basin, 9 reservoirs are used to fit these equations while 18 are used for validation. Across the training reservoirs, day-ahead release is predicted with a Nash-Sutcliffe efficiency (NSE) of 0.984 with an average root mean squared error (RMSE) of 22.9% of the daily mean release. For the testing reservoirs, the NSE is 0.970 with an average RMSE of 32.9%. The worst performing reservoir in the training set has an NSE of 0.672 and an RMSE of 42.6% while the worst performing reservoir in the validation set has an NSE of 0.623 and an RMSE of 57.6%. Additionally, this work illuminates important aspects and relationships that may be useful in future efforts. For example, for periods of low release, the average release over the past week is an important predictor for day-ahead release but has a much smaller effect during periods of high release. Similarly, while reservoir storage has little effect on predicted release values in upstream reservoirs, if it is greater than the average storage over the past week, reservoir releases tend to be higher. Though the parameterizations developed here produce accurate day-ahead release predictions for untrained reservoirs in the Tennessee River Basin, they may not translate perfectly to other river basins. Thus, there is a need to generalize reservoir operations in the form of simple equations for incorporating them into distributed hydrologic models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H25A1055F