Intelligent co-simulation: neural network vs. proper orthogonal decomposition applied to a 2D diffusive problem

Abstract

One possibility to improve the accuracy of building performance simulation (BPS) tools is via co-simulation techniques, where more accurate mathematical models representing particular and complex physical phenomena are employed through data exchanging between the BPS and a specialized software where those models are available. This article performs a deeper investigation of a recently proposed co-simulation technique that presents as novelty the employment of artificial intelligence as a strategy to reduce the computational burden generally required by co-simulations. Basically, the strategy, known as intelligentco-simulation, constructs new mathematical models through a learning procedure (training period) that is performed using the input–output data generated by a standard co-simulation, where the models of specialized software are employed. Once the learning phase is complete, the specialized software is disconnected from the BPS and the simulation goes on by using the synthesized models, requiring a much lower computational cost and with a low impact on the accuracy of the results. The synthesis of accurate-and-fast models is performed through machine learning techniques and the purpose of this paper is precisely a deep investigation of two techniques – recurrent neural networks and proper orthogonal decomposition reduction method, whose main goal is to reduce the training time period and to improve the accuracy. The case study focuses on a co-simulation between Domus and CFX programs, performing a two-dimensional diffusive heat transfer problem through a building envelope. The results show that for a standard co-simulation of 14 h, the intelligent co-simulation provided a reduction of 90% in the computer run time with accuracy error at the order of