An online learning framework for self-adaptive dynamic thermal modeling of building envelopes

Abstract

Modeling dynamic thermal behavior plays a crucial role for developing effective energy management strategies of buildings. In this work, we introduce an improved framework for online training of the GP-integrated 3R-2C model for building envelopes using a modified Ensemble Kalman filter (EnKF) algorithm. The framework combines the R-C model for defining physical constraints with Gaussian process (GP) predictor to capture time-varying thermal dynamics. This modified EnKF algorithm is designed based on Bayesian estimation theory and adapted for online learning, which eliminates the need for offline data collection and training. Numerical experiments using varied synthetic datasets were conducted to evaluate the effectiveness of the proposed framework. The results show that this framework significantly outperforms conventional schemes using the augmented EnKF algorithm: its relative deviations from benchmark results are only half of those achieved by the conventional ones. Importantly, this framework can gradually improve itself and be adapted to the incidental variations in heat transfer characteristics of building surfaces. Furthermore, it is highly flexible, which can be extended to deal with various thermal modeling problems through the utilization of different R-C models and configurations of GP predictors or machine learning techniques.