Estimating Multi-point Indoor Temperature from Different Season Data based on Correlation-based Two-Step Learning

Abstract

This paper presents a method to estimate multi-point temperature in a large-scale indoor space on the basis of indoor temperature and Heating, Ventilation, and Air-Conditioning (HVAC) temperature measured for one day in the same season and data measured for a longer period in a different season. Existing studies have tried to learn an estimation or prediction model from such learning data on the basis of fine-tuning or transfer learning in which the loss function is calculated from differences such as mean squared error or accuracy between measured and estimated data. However, it is difficult for existing methods to estimate on the basis of data from a different season because the difference between indoor temperature and HVAC temperature depends on the season. In this paper, we focus on not the difference but the correlation between indoor temperature and HVAC temperature, which does not depend on seasons. We propose correlation-based two-step learning in which the loss function is calculated from the correlation between indoor temperature and HVAC temperature at the first learning. We evaluate the effectiveness of our proposal using measured indoor temperature and HVAC temperature data in a real building.