In the present paper, the energy performance of buildings forming a school centre in the northwest of Spain was analyzed using a transient simulation of the energy model of the school, which was developed with TRNSYS, a software of proven reliability in the field of thermal simulations. A deterministic calibration approach was applied to the initial building model to adjust the predictions to the actual performance of the school, data acquired during the temperature measurement campaign. The buildings under study were in deteriorated conditions due to poor maintenance over the years, presenting a big challenge for modelling and simulating it in a reliable way. The results showed that the proposed methodology is successful for obtaining calibrated thermal models of these types of damaged buildings, as the metrics employed to verify the final error showed a reduced normalized mean bias error (NMBE) of 2.73%. It was verified that a decrease of approximately 60% in NMBE and 17% in the coefficient of variation of the root mean square error (CV(RMSE)) was achieved due to the calibration process. Subsequent steps were performed with the aid of new software, which was developed under a European project that enabled the automated calibration of the simulations.

In this paper, two different methodologies are applied to the parameter estimation problem of a computational model of a Round-Robin test box. The numerical model is developed using TRNSYS. A global sensitivity analysis is carried out to determine the most important parameters that should be considered in the subsequent calibration procedure. Using the Bayesian (probabilistic) approach, the posterior distribution of the unknown input parameters is estimated via simulation techniques. Using the deterministic approach, executed in GenOpt, the calibration is performed by the minimization of an objective function that measures the differences between model predictions and real measured data. Parameter estimation results obtained with both methodologies are then compared and discussed. A reduction of the Coefficient of Variation of the Root Mean Square Error (CV (RMSE)) after calibration over 40% with both methods has been obtained, being the CV (RMSE) for calibration and validation periods on average 3.21% and 2.40%, respectively.