Sample average approximation and stability tests applied to energy system design

Abstract

This paper uses confidence intervals from sample average approximation (SAA) and stability tests to evaluate the quality of the solution of a long-term energy system model with stochastic wind power production. Using poorly designed scenarios can give stochastic model results that depend on the scenario representation rather than the actual underlying uncertainty. Nevertheless, there is little focus on the quality of the solutions of stochastic energy models in the applied literature. Our results demonstrate how too small a sample size can give a poor energy system design and misrepresent the value of the stochastic solution (VSS). We demonstrate how to evaluate the number of scenarios needed to ensure in-sample and out-of-sample stability. We also show how replication and testing of many candidate solutions using SAA iterations can provide a solution with a satisfactory confidence interval, including when the samples contain fewer scenarios than required for stability. An important observation, though, is that if SAA repeatedly solves the model with a sample size that satisfies in-sample and out-of-sample stability, the confidence interval is narrow, and the solutions are of high quality in terms of providing a tight bound for the optimal solution.