AI-Enabled Predictive Control of HVAC Systems for Energy-Efficient Buildings: An Analysis of Thermal Comfort, Load Forecasting, Power Consumption, and Smart Control Performance

Abstract

Artificial intelligence can aid predictive HVAC control only as long as forecasts are accurate enough for decisions to be made without degrading indoor conditions. The approachability of control assesses the practicality of using a real, high-resolution subset of Lawrence Berkeley National Laboratory Building 59. It uses 2,976 fifteen-minute intervals in January 2020 that were either measured or modified and includes north and south HVAC electrical demand, outdoor temperature and relative humidity, solar radiation, zone temperature and heating setpoint, supply airflow and return supply/return fan speeds. Descriptive statistics, Pearson correlations, and HAC-robust regression were applied, as well as chronological one-hour-ahead forecasts. The average combined HVAC demand was 30.97 kW, with a standard deviation of 9.43 kW and a range of 7.70 to 60.22 kW. The bivariate correlation of the supply airflow and demand was r = .514. The multivariable model accounted for 55.4% of the contemporaneous demand. For one-hour-ahead forecasting, the RMSE of the linear regression was 4.29 kW with R² of .730, and the persistence model had an MAE of 2.75 kW, thus showing that nonlinear models did not outperform the strong short-horizon baseline. Mean absolute zone-to-heating-setpoint deviation was 0.40°F. These results support the case for using a predictive control architecture where simultaneously estimating forecast quality, thermal state tracking, and actuator signals is analyzed, as provocative estimates of energy savings potential are avoided for the January no-control scenario.