Distributionally Robust Decision-Making under Ambiguity: Case Study of Water Environmental Management

Abstract

Decision-making under uncertainty is ubiquitous in environmental project planning. Environmental processes such as a streamflow discharge often present a subexponential memory, where the autocorrelation persists for a long time. In addition, optimization problems driven by environmental processes encounter the issue of model ambiguity because of a lack of sufficient data for model identification. To facilitate decision-making for the management of aquatic environments (e.g., flood mitigation, water abstraction for hydropower generation), we formulate a unified distributionally robust stochastic optimization problem based on a mixed moving average (MMA) process. The MMA process is a superposition of infinite-dimensional affine stochastic processes that is seemingly complex, but the affine property helps with the formulation and computation of the optimization. Our problem is based on a convex objective with a nonsmooth conditional value-at-risk measure. We present a convergent regularization to obtain its smooth and strictly convex counterpart. The model ambiguity is represented as a distortion of the probability density of the target dynamics, and it is penalized by a divergence with which the optimization problem remains convex and becomes computable. As a case study, we apply the optimization problem to two cases with identified parameter values. The performance of the optimized dynamics is evaluated through a statistical simulation. This paper serves as a multidisciplinary work covering both the theory and application of distributionally robust optimization.