Generalizing Reservoir Operations Using a Piecewise Classification and Regression Approach

Abstract

AbstractInflow anomalies at varying temporal scales, seasonally varying storage mandates, and multipurpose allocation requirements contribute to reservoir operational decisions. The difficulty of capturing these constraints across many basins in a generalized framework has limited the accuracy of streamflow estimates in land‐surface models for locations downstream of reservoirs. We develop a Piecewise Linear Regression Tree to learn generalized daily operating policies from 76 reservoirs from four major basins across the coterminous US. Reservoir characteristics, such as residence time and maximum storage, and daily state variables, such as storage and inflow, are used to group similar observations across all reservoirs. Linear regression equations are then fit between daily state variables and release for each group. We recommend two models—Model 1 (M1) that performs the best when simulating untrained records but is complex and Model 2 (M2) that is nearly as performant as M1 but more parsimonious. The simulated release median root mean squared error is 49.7% (53.2%) of mean daily release with a median Nash‐Sutcliffe efficiency of 0.62 (0.52) for M1 (M2). Long‐term residence time is shown to be useful in grouping similar operating reservoirs. Release from low residence time reservoirs can be mostly described using inflow‐based variables. Operations at higher residence time reservoirs are more related to previous release variables or storage variables, depending on the current inflow. The ability of the models presented to capture operational dynamics of many types of reservoirs indicates their potential to be used for untrained and limited data reservoirs.