Evolutionary Air-Conditioning optimization Using an LSTM-Based Surrogate Evaluator

摘要

We propose a fast air-conditioning temperature optimization system using a surrogate solution evaluator. Simulation-based evolutionary optimization does not require the formulation of a mathematical model of the optimization problem and allows the treatment of the problem by considering a black box. However, the simulation-based solution evaluation is often time-consuming. Because evolutionary algorithms need to generate many candidate solutions for optimization, the time-consuming solution evaluation will be a bottleneck, which is encountered in air-conditioning temperature optimization. The building simulator provides the thermal comfort level and the power usage by inputting a candidate air-conditioning temperature setting through a complicated simulation. Although the results are useful compared with those obtained by a simple mathematical model, it is time-consuming. To accelerate the optimization process, we propose a surrogate evaluator based on the time-series predictive long short-term memory, which is a recurrent neural network architecture. Instead of the time-consuming building simulator, the surrogate evaluator outputs the two time-series data of the thermal comfort and power consumption, and schedules are optimized with the multi-objective particle swam-based optimizer OMOPSO. Experimental results show that the proposed system could obtain practical schedule sets, and the optimization was accelerated by using the surrogate evaluator.