Derivation of monthly reservoir release policies

TLDR

Hoover Reservoir on Big Walnut Creek in central Ohio serves as the case study. A backward‑looking dynamic program was used to optimize releases for a single multiple‑purpose reservoir, solving one‑ and two‑sided quadratic loss functions, and monthly policies were derived by regressing optimal releases on input and state variables, with linear and nonlinear policies developed and compared through simulation. For a two‑sided quadratic loss, linear policies perform as well or better than nonlinear ones, whereas for a one‑sided loss nonlinear policies outperform linear ones, and the maximum R² criterion for selecting policies may not always be suitable.

Abstract

A single multiple‐purpose reservoir is analyzed using a backward looking dynamic program algorithm to obtain optimal releases. The dynamic program is solved for both one‐sided and two‐sided quadratic loss functions. Monthly policies are derived by regressing the optimal set of releases on the input and state variables. Linear and nonlinear release policies are developed, then verified and compared through simulation. For a two‐sided quadratic loss function, linear policies are as good or better than nonlinear policies. However, for a one‐sided quadratic loss function, nonlinear policies give improved performance over linear policies. It is also illustrated that the maximum R 2 criterion for selecting release policies may not always be appropriate. Hoover Reservoir, located on Big Walnut Creek in central Ohio, is used as a case example.