Evaluating single-zone grey-box thermal models: Impact of model structure on prediction accuracy and computational cost

Low-order grey-box resistance-capacitance (RC) models are widely used in building energy simulations due to their physical interpretability and computational efficiency. The RC structure determines the trade-off between prediction accuracy and computational cost for tasks such as model predictive control; however, prior comparisons have been limited to a narrow set of configurations, lacking systematic and quantitative assessments. This paper systematically evaluates the impact of model structure in single-zone grey-box thermal models by introducing a parametric RC model set with variants in model order, solar-gain node placement, infiltration treatment, and solar-radiation data fidelity. Model parameters are estimated within a factorial experimental design using high-fidelity data from multiple BESTEST envelope cases under three distinct climates, with weekly validation periods distributed across an entire year. The results indicate that oversimplified structures and low-fidelity solar inputs are the dominant drivers of prediction error, yielding 0.5-1.0 K RMSE in zone-air temperature predictions. Conversely, three recommended 3C to 4C configurations provide robust accuracy with RMSE values within 1 K. Further structural fine-tuning within 3C-4C models reduces RMSE by only 0.25 K, at a thirty-fold increase in computational cost for parameter estimation. Consequently, selecting an appropriate complexity level and ensuring input accuracy outweigh the benefits of further structural fine-tuning. For single-zone, envelope-dominant cases, 3C-4C configurations are optimal, minimizing computational effort while maintaining accuracy. These findings offer recommendations for designing RC model structures to mitigate structural uncertainties and enhance thermal predictions in buildings.