An adaptive grey-box model for nearly zero energy building operations in challenging climate

Grey-box model predictive control (MPC) can utilize building thermal mass to act as “thermal batteries” in nearly zero-energy buildings (NZEBs). However, conventional grey-box models demonstrate limited reliability under rapid weather changes. Specifically, conventional adaptive (CA) models trained on sequential historical data lack representativeness during extreme weather conditions, reducing predictive accuracy. This study proposes a Weather Similarity-based Adaptive (WSA) method for grey-box model calibration through similarity-based training dataset selection. The WSA method identifies representative historical conditions using key meteorological variables. Both predictive and control performance of the proposed method in an NZEB were evaluated using a high-fidelity co-simulation testbed across multiple climate scenarios, including analysis of hyperparameter influences. Results: demonstrated that training data representativeness proves more critical than recency for model reliability. WSA models achieve up to 30% reduction in prediction error compared to Linear Time-Invariant (LTI) models during extreme weather events. The method maintains comparable performance during stable conditions while requiring less training data. The WSA method with partial parameter updating achieved optimal control performance, minimizing both energy costs and thermal discomfort. This approach proved particularly effective during the cold snap, where CA methods fail. Overall, the proposed WSA method improves MPC reliability in NZEBs under challenging climates by ensuring the representativeness of the training data.