A local search method for costly black-box problems and its application to CSP plant start-up optimization refinement

Abstract

A variety of engineering applications are tackled as black-box optimization problems where a computationally expensive and possibly noisy function is optimized over a continuous domain. In this paper we present a derivative-free local method which is well-suited for such problems, and we describe its application to the optimization of the start-up phase of an innovative Concentrated Solar Power (CSP) plant. The method, referred to as rqlif, exploits a regularized quadratic model and a linear implicit filtering strategy so as to be parsimonious in terms of function evaluations. After assessing the performance of rqlif on a set of analytical test problems in comparison with three well-known local algorithms, we apply it in conjunction with a global algorithm based on RBFs interpolation to the start-up optimization of the CSP plant developed in the PreFlexMS H2020 project. For the test problems, rqlif provides good quality solutions in a limited number of function evaluations. For the application, the global–local strategy yields a substantial improvement with respect to the reference solution and significantly reduces the thermo-mechanical stress suffered by the plant components.