Infusing Domain Knowledge into Gaussian Process Regression: Portable Prediction Model for Indoor Temperature

Abstract: Accurate indoor temperature prediction models are essential for smart building control systems. Gaussian Process Regression (GPR) is widely used due to its prediction accuracy and error guarantees. However, real-life applications face challenges, including GPR’s poor scalability with large datasets and the manual selection of input features. This study employs LoG-GP, a distributed GPR method, for continuous model updating and fast predictions.The wrapper method is used for sensitivity analysis to identify optimal input features for various building types. A metadata-driven workflow is proposed to automate feature selection and dataset retrieval, achieving computationally efficient, accurate, and adaptable models tested on new buildings.