Forecast of air-conditioning duration in office buildings in summer using machine learning and Bayesian theories

Abstract

Air-conditioning accounts for the bulk of a building’s energy use. Researchers have focused on various affecting parameters (e.g., ventilation, thermostat setpoint) to understand occupants’ interactions with air-conditioning to reduce energy consumption. However, there has been little research on the duration of air-conditioning usage, which is the leading source of energy consumption. Conventional thermal comfort models do not fully reflect the individual preferences of the occupants. This paper presents a method to predict occupant behaviors of air-conditioning, under the influences of different usage patterns and multiple environmental factors. A three-star green office building is supposed to conduct field measurements to obtain data concerning indoor use of air-conditioning and outdoor environmental parameters. The fitting occupant behavior model is embedded in building energy consumption simulation, which considers the occupants’ perceived control and the resulting comfort and energy performance. Predictions of the air-conditioning usage are made by establishing the measured data with the simulation data through Bayesian inference and by iterative optimizations with the Monte Carlo method. The research results show that machine learning can predict the occupant's air conditioning usage, and different methods have different prediction accuracy, among which the GP model using Bayesian theory has better applicability with guaranteed accuracy.