Master's Thesis Asad Esmailzadeh

 

Automated calibration of simulation models for energy systems supported by machine learning

parameter optimization Copyright: EBC Design of a calibration process consisting of a classification, a sensitivity analysis and a parameter optimization

The integration of model based control has been gaining attention in recent years due to its ability to increase the efficiency of energy systems for instance in model predictive or adaptive control applications. During the process of calibration, the simulation model is adjusted to the physical reference system by changing the model parameters. With reference to the performed methods for calibration in practice, most of them belong to the manual calibration category, which defines a heuristic trial and error process that apart from being time consuming, is also highly dependent on the engineers expertise and experience.

In the scope of this work, an automatic calibration method is developed, which can be roughly separated into three modular parts. The first part divides the time series data into different classes based on its properties by using a supervised classification algorithm. Second, for each of those classes a sensitivity analysis is executed. The sensitivity based parameter ranking decides about in which class the parameters are set as variables for the optimization in the third part. The developed calibration method is validated with monitoring time series data by using a dynamic simulation

model of a boiler and a heat pump. The Cross-Validation results show that the calibrated model’s prediction quality is decent in case of the boiler and very high for the heat pump model. Furthermore, the time needed for calibrating the latter one can be reduced by more than 99 percent in comparison to a manual approach executed in a previous work. In a following examination the application of an unsupervised machine learning algorithm is tested with the objective to automatically create the classes from the time series data to train the classifier. Therefore an unsupervised cluster analysis is used for the purpose of creating physical classes. The results show clearly that the automatically generated classes lead to a further increase in prediction accuracy (approximately 100 percent) for the supervised classification process without effecting the total calibration results.