Adaptive Sensing and Personalized Thermal Comfort Prediction for Building Energy Efficiency
Advisor: Professor Nora El-Gohary
Abstract:
Building energy consumption has increased over the past decades due to the population growth and the increasing demands for building services and comfort. Sensor-based data-driven approaches could play an important role in identifying energy-saving strategies that balance the tradeoff between energy saving and thermal comfort. However, most of the existing approaches (1) rely on data from fixed sensors, which are limited in coverage and adaptiveness to the building dynamics; and (2) use pre-defined thermal comfort standards without sufficient consideration of the personal comfort preferences of the occupants. To address this gap, this research proposes an adaptive and personalized approach for building energy efficiency using flexible sensing and machine learning. The approach aims to support building energy efficiency through (1) flexible and adaptive sensing of the indoor environmental conditions towards accurate building energy consumption prediction; and (2) personalized and automated occupant thermal comfort prediction based on occupant behavior data from video recordings.
To achieve the aforementioned goals, the research methodology included seven primary tasks. First, conducting a comprehensive literature review on related topics such as methods and approaches for occupant thermal comfort prediction, indoor condition monitoring using sensors, building energy consumption prediction, and machine learning in building energy efficiency. Second, conducting a survey to discover occupant cues that building occupants might engage with to adjust their thermal comfort and developing a deep learning-based model to recognize these cues from RGB videos. Third, studying the distribution of indoor physical parameters over time and space using field data from multimodal sensors, and proposing a consensus clustering-based method to determine optimal sensing locations for indoor physical condition monitoring that are more robust to the changes in indoor conditions over time and space. Fourth, developing several machine learning models to study the impact of sensor deployment configurations (e.g., number and location) for physical parameter data collection on building energy consumption prediction accuracy. Fifth, developing a deep reinforcement learning-based method for learning optimal sensor locations for accurate building energy consumption prediction – over time and in a way that can be adaptive. Sixth, developing a deep learning-based model for predicting personalized occupant thermal comfort preferences using occupant features (e.g., poses, clothing) extracted from RGB videos and indoor environmental features extracted from sensor measurements. Seventh, integrating the aforementioned models and methods into a proof-of-concept framework for optimizing energy consumption and thermal comfort, and testing the framework using a case study.
The developed computational models and methods were evaluated individually, showing their potential for identifying optimal sensor locations for indoor environmental data collection and predicting building energy consumption and (individual) occupant thermal comfort. The integrated framework showed good performance, achieving mean MAE and CV of 1.2 and 16.3% in predicting energy consumption, respectively, and mean accuracy and weighted F1-measure of 79.4 and 84.4% in predicting occupant thermal comfort preferences. The optimization results showed potential energy savings of up to 28.8%, while taking into consideration individual thermal comfort preferences.
