Assessing the effect of network order on epistemic uncertainty quantification for reduced-order grey-box energy models

Grey-box building energy models are becoming extremely popular for modeling building thermal energy performance and subsequently evaluating base case energy consumption, establishing efficiency scenarios, implementing model predictive control and forecasting building thermal behavior. Energy simulation inputs and model parameters in such models introduce uncertainty and hence, highly affect the accuracy and reliability of energy simulation results. Furthermore, increasing the reduced-order model complexity eventually increases the epistemic uncertainty (lack of knowledge) in energy simulation results due to an associated increase in number of model parameters. Existing studies often provide disintegrated analysis of model complexity, accuracy and uncertainty when implementing reduced-order grey-box models. This study proposes a framework to create reduced-order grey-box energy models and henceforth, quantify and analyze the effect of epistemic uncertainties through variation of network order. The devised framework further enables the identification of a balance between network complexity, accuracy and model uncertainty. A strong relationship exists between network order and model parameter uncertainty. Increasing the model complexity has no significant effect on model accuracy (CVRMSE reduces from 3.65% to 2.55%). The epistemic spread of uncertainties increases by a significant amount (~ 10%).