Multiobjective Direct Policy Search Using Physically Based Operating Rules in Multireservoir Systems

Abstract

This study explores the ways to introduce physical interpretability into the process of optimizing operating rules for multireservoir systems with multiple objectives. Prior studies applied the concept of direct policy search (DPS), in which the release policy is expressed as a set of parameterized functions (e.g., neural networks) that are optimized by simulating the performance of different parameter value combi-nations over a testing period. The problem with this approach is that the operators generally avoid adopting such artificial black-box functions for the direct real-time control of their systems, preferring simpler tools with a clear connection to the system's physics. This study addresses this mismatch by replacing the black-box functions in DPS with physically based parameterized operating rules, for example by directly using target levels in dams as decision variables. This leads to results that are physically interpretable and may be more acceptable to operators. The methodology proposed in this work is applied to a network of five reservoirs and four power plants in the Nechí catchment in Colombia, with four interests involved: average energy generation, firm energy generation, flood hazard, and flow regime alteration. The release policy is expressed depending on only 12 parameters, which significantly reduces the computational complexity compared to existing approaches of multiobjective DPS. The resulting four-dimensional Pareto-approximate set offers a variety of operational strategies from which operators may choose one that corresponds best to their preferences. For demonstration purposes, one particular optimized policy is selected and its parameter values are analyzed to illustrate how the physically based operating rules can be directly interpreted by the operators.

Compound flood impact forecasting: integrating fluvial and flash flood impact assessments into a unified system

“…The rules on which dam operators base their release decisions are typically unknown (e.g. Nazemi and Wheater, 2015;Ritter et al, 2020b), usually hindering an adequate representation of the effects of dams in hydrological models. Secondly, large parts of south-east Spain, and in particular the Segura Basin, are situated on a highly karstic topography (Goldscheider et al, 2020), in which hydrologi-cal simulations generally show high uncertainties (Hartmann et al, 2014).…”

Multi-Objective Optimization of Reservoir Operation using Machine Learning Models. Case study: Hatillo Reservoir in the Dominican Republic

“…Giuliani et al (2014) and Giuliani et al (2016) used radial basis networks (RBNs), demonstrating their benefits as universal approximation functions. Likewise, Ritter et al (2020) applied DS using physically based operating models.…”

Deep Reinforcement Learning for Real-Time Optimization of Pumps in Water Distribution Systems

“…Savic et al [7] found that many state-of-the-art (SOTA) optimization techniques in WDS operation optimization involves some kind of metaheuristics. Ritter et al [8] presents a multiobjective direct policy search technique where different heuristic approaches share the individuals of a population among different metaheuristics. This way the different beneficial properties are combined from different metaheuristics but incorrectly chosen hyperprameters of the distinct methods can alter the whole optimization process.…”