Master's Thesis Julian Schaefer

 

Calibration of dynamic building models under uncertainty using machine learning

Metamodel Copyright: EBC Metamodel of a building archetype using the statistical figure R2 (black dots: training data, colored surface: metamodel)

The dynamic simulation of energy systems at an urban level is becoming increasingly important as an additional decision criterion. An essential aspect is the dynamic simulation of the heating and cooling requirements of buildings. Typically, very little information about the topographical and physical structure of buildings is provided. To circumvent this problem, building archetypes are a widely used approach. With the help of these archetypes, missing information is substituted with statistically determined values. Studies have shown that with the use of archetypes, the simulated dynamic demand on large groups of stock buildings can be accurately reproduced, however, this is not the case for individual buildings. This is particularly noticeable if the actual characteristics of the individual building deviate from the statistical average.

In order to improve the demand prediction of individual buildings based on archetypes, a Bayesian calibration is implemented using the software tool TEASER. This methodology provides an automated calibration of archetypes by using high-resolution measurement data, factoring in uncertainties in the building model, in the calibration parameters and in the measurement data. In addition, a method is presented that enables Bayesian calibration on the basis of statistical figures. The method is evaluated using real measurement data from selected buildings of Jülich Research Centre. The results show that both in conventional implementation as well as based on selected statistical, Bayesian calibration figures achieve a significant improvement in individual demand forecasting. As a result of the calibration, the root of the mean square deviation between the measured and the predicted heat demand can be reduced by up to 57%. Calibration from statistical figures provides a significant improvement with regards to runtime with a moderate reduction of the calibration accuracy.