Action-based personalized dynamic thermal demand prediction with video cameras

Abstract

Real-time and accurate prediction of indoor thermal demands is of great significance for improving building energy control and human satisfaction. This study proposes a system for dynamically predicting personal thermal demand with ordinary cameras, which integrates action acquisition, action feature extraction, action classification and thermal demand prediction. Three deep learning models were developed and integrated with multiobject tracking algorithms to build a personal thermal demand model. The proposed system and models were validated through a field measurement. The results show that the defined thermal adaptive behaviors can characterize human thermal sensation well. The action feature extraction model constructed based on the apparent feature extraction network combined with the proposed action classification model can determine whether the occupants have taken a thermal adaptive behavior with 92% accuracy and classify the behavior into specific categories with 86% accuracy. The thermal demand prediction model can achieve 95% accuracy, and the overall accuracy of the personal thermal demand model is 91%. This study provides technical and theoretical support for designing intelligent building decision-making schemes based on noninvasive thermal demand prediction.