Pouya Aghaei-Pour, Tobias Rodemann, Markus Olhofer, Jussi Hakanen, Kaisa Miettinen, "On Surrogate Management in Interactive Multiobjective Building Energy System Design", ECCOMAS Thematic Conference Computational Sciences and AI in Industry (CSAI), 2019.

When thinking of possible extensions of energy systems, decision making for larger buildings consists of a series of complex investment decisions. The consideration involves multiple objectives like investment and annual operation costs, CO 2 emissions and module lifetime to be considered simultaneous. Thus, in building energy system management, methods of multiobjective optimization are needed to support decision making. We have a system upgrade problem with five objective functions: initial investment cost, running cost, CO 2 emissions, resilience to power outages and battery lifetime. We consider hardware additions and modifications of system controllers and use a building simulation software to simulate energy flows to analyze different investment options. We use evolutionary algorithms for optimization. Unfortunately, they are very time consuming since they deal with populations of solution candidates and simulations may take days or even weeks. In our case study of a building energy system design, we can vary the simulation period leading to a trade-off between the simulation time (seconds to minutes) and the accuracy. In order to solve optimization problems with a desired accuracy, longer simulation periods have to be used but this results with long evaluation times for the objective functions. Therefore, using surrogate models help to expand the simulation period, so that we can increase the accuracy. More specifically, we apply Kriging (or Gaussian processes) as surrogate models because they can provide us uncertainty information about the surrogates. As an evolutionary algorithm, we apply the reference vector guided evolutionary algorithm (RVEA) and, as said, incorporate surrogate models in it. Instead of settling with surrogate-assisted RVEA that tries to widely repre- sent all solutions with different trade-offs among the objectives, we introduce an interactive method. Such a method has not been used before to solve this particular problem. Among the advantages of interactive methods are that the decision maker can learn about the trade-offs involved conveniently and can concentrate on those solutions that are interesting. The challenge here is that the decision maker usually has a limited time to spend on the solution process but updating the surrogate model may take a lot of time. In this research, we focus on an adaptive method to update the Kriging model based on the decision maker’s preferences. For example, we can spend more time on training the Kriging model before the decision maker is involvedwith the solution process, and after the first interaction, we only update the Kriging model in local areas that the decision maker is interested in. Moreover, we test Kriging models with different kernels for this problem and compare their performances with each other.