Experimental calibration of dynamic thermal models is required for model predictive control and characterization of building energy performance. In these applications, the uncertainty assessment of the parameter estimates is decisive; this is why a Bayesian calibration procedure (selection, calibration and validation) is presented. The calibration is based on an improved Metropolis-Hastings algorithm suitable for linear and Gaussian state-space models. The procedure, illustrated on a real house experiment, shows that the algorithm is more robust to initial conditions than a maximum likelihood optimization with a quasi-Newton algorithm. Furthermore, when the data are not informative enough, the use of prior distributions helps to regularize the problem.

Dynamic behaviour of a system may be described by models with different forms: thermal (RC) networks, state-space representations, transfer functions, and ARX models. These models, which describe the same process, are used in the design, simulation, optimal predictive control, parameter identification, fault detection and diagnosis, and so on. Since more forms are available, it is interesting to know which one is the most suitable by estimating the sensitivity of the model to transform into a physical model, which is represented by a thermal network. A procedure for the study of error by Monte Carlo simulation and of factor prioritization is exemplified on a simple, but representative, thermal model of a building. The analysis of the propagation of errors and of the influence of the errors on the parameter estimation shows that the transformation from state-space representation to transfer function is more robust than the other way around. Therefore, if only one model is chosen, the state-space representation is preferable.