Predictive AC Control Using Deep Learning: Improving Comfort and Energy Saving

Abstract

The growing global population and the availability of energy-hungry smart devices are critical factors in today's alarmingly high electricity usage. The majority of energy used in urban areas is consumed by buildings, with heating, ventilation, and air conditioning (AC) systems accounting for a significant amount of energy use. This project proposes an AC controlling algorithm that uses the Internet of Things sensors and a deep learning framework in temperature prediction to control a single AC unit. The algorithm consists of a Long Short-Term Memory (LSTM) model to predict the indoor temperature for the next J minutes. The highlight of this model is its capacity to predict the future temperature based on the predetermined AC status, whether it is switched on or off. The AC unit will be turned off if the J-minute predicted temperature is within the desired thermal comfort range, and it will be turned back on if the sensor readings exceed the upper pre-set threshold. The experiment is performed on the dataset collected by Chulalongkorn University Building Energy Management System (CU-BEMS). The LSTM prediction model developed using CU-BEMS data yields an average Root Mean Squared Error and Mean Absolute Error of 0.08 and 0.03, respectively. A half-day simulation is also performed in controlling the AC unit from 7:39 a.m. to 11:35 a.m. The proposed algorithm shows that 49.00% of the time, the AC unit can be turned off while the thermal range is maintained between 27ºC to 27.9ºC, providing a strategy for managing the AC unit and achieving energy savings.