Control-oriented inverse modeling of the thermal characteristics in an office

Abstract

This article puts forward 12 graybox models at varying sensory model inputs and parameters. The parameters of each model were recursively estimated using the extended Kalman and the particle filtering methods. A systematic way to evaluate the predictive performance and the appropriateness of the models was introduced by using the data gathered in three perimeter offices. The simplest feasible models that can capture the timing and magnitude of local extrema were the models with five parameters and six inputs collected with low-cost sensors. They could robustly predict the indoor temperature at less than ±0.6°C mean absolute error over a 2-day horizon.