Personalized Thermal Comfort Forecasting for Smart Buildings via Locally Weighted Regression with Adaptive Bandwidth

A personalized thermal comfort prediction method is proposed for use in combination with smart controls for building automation. Occupant thermal comfort is traditionally measured and predicted by the Predicted Mean Vote (PMV) metric, which is based on extensive field trials linking reported comfort levels with the various factors. However, PMV is a statistical measure applying to large populations, and the actual thermal comfort could be significantly different from the predicted value for small groups of people. Moreover it may be hard to use for a real-time controller due to the number of sensor readings needed. In the present paper, we propose Robust Locally Weighted Regression with Adaptive Bandwidth (LRAB), a kernel based method, to learn individual occupant thermal comfort based on historical reports. Using publicly available datasets, we demonstrate that this technique is significantly more accurate in predicting individual comfort than PMV and other kernel methods. Therefore, is a promising technique to be used as input to adpative HVAC control