Fast prediction of annual hourly indoor non-uniform unsteady environment through coupled model of daylighting, CFD, energy simulation and machine learning

Abstract

Predicting the non-uniform indoor temperature field is time-consuming by using computational fluid dynamic (CFD) methods. This process usually becomes even more complex when sunlit shine on ground which creates a dynamic hot spot boundary, which is largely overlooked by previous studies. Currently, it is common to conduct annual building energy simulation, but fast generation of annual indoor CFD results is still unapproachable. This study proposes a new model for forecasting indoor temperature distribution, developed by combining POD algorithms with machine learning techniques. Through a series of experimental and numerical validations, the results indicate that the proposed POD-ML model can accurately and rapidly predict indoor temperature fields, and it performs well under various model settings, with errors ranging from 1.26 % to 9.11 %. The model allows for continuous simulation of indoor temperature fields across the year using real meteorological data, providing architects and HVAC system designers with deeper insights into indoor temperature distribution, which aids in achieving greater energy savings and improving indoor environmental conditions.