Room Occupancy Prediction from Temperature Data with Deep Convolutional Neural Networks

This study aims to estimate the number of people in a room using data from temperature sensors placed in a room. The study utilizes an open-source dataset comprising time-dependent temperature sensor data. The days when the number of people in the dataset was always zero were removed to avoid misleadingly increasing the success of the model, and the number of data points was reduced by averaging every 10 measurements. The temperature data were converted into RGB images of 28 x 28 pixels, and the measurements from each sensor were assigned to a different region in the image. A convolutional neural network model was trained by dividing these images into training, validation, and test sets. The model was able to predict the no-person and low-person classes with high accuracy. However, at higher headcounts, the model’s performance degraded. In particular, prediction errors increased in transition situations where the number of people changes rapidly. The accuracy of the model on the test dataset is obtained as 93.33%. The results show that temperature data can be effectively used to predict occupancy levels. This study lays a foundation for headcount prediction based on temperature data and offers significant potential in applications such as smart building systems.