Prediction of Unsteady Indoor-Temperature with Few Pattern Data Learning and Prediction Model Selection Based on Feature Contribution

The aim of this paper is to predict indoor-temperature by machine learning under three main constraints: 1) indoor-temperature is unsteady due to people flow, 2) only data with few control patterns of air-conditioning can be used for training, and 3) indoor-temperature is to be accurately and plausibly predicted under unknown air-conditioning control patterns not included in training data. Previous studies tried to predict indoor-temperature in buildings without people but with air-conditioning data for various control patterns. However, these constraints make predictions of indoor-temperature irregular because of unsteady indoor-temperature and inadequate data patterns. To solve the problems, we propose a model selection method based on prediction accuracy and feature contribution. The method can select the prediction model appropriate for the observed instability and can augment data patterns. We demonstrate the effectiveness of our proposal using measured sensor data.