Bachelor's thesis Vincent Evenschor
Classification of time series data of building automation systems based on machine learningCopyright: EBC
Optimized building automation systems are highly promising to contribute for significant energy savings related to the building sector. In practice, most operate inefficiently though, since their analysis and fault detection is difficult, as there exists no standardized scheme for the structure and denomination of data points in building automation systems yet. However, growing numbers of deployed sensors and actuators are encouraging to integrate machine learning for knowledge discovery by using only the produced data itself.
So far most research was focused on unsupervised clustering algorithms or supervised classification of univariate time series. Motivated by the task of separating concrete core activation from other energetic subsystems, a multivariate time series classification approach was chosen to assign sets of data points to classes of energy systems that appear in heating, ventilation and air conditioning by applying state-of-the-art deep learning models.
In use of flow and return temperatures, as well as volume flow that were retrieved from the monitoring database of the E.ON Energy Center, located in Aachen, Germany, multivariate time series datasets were created and preprocessed. The best trained models achieve accuracies beyond 95% on the test data in classifying concrete core activations and façade ventilation unit’s heating and cooling water systems, being able to even distinguish between systems of the same class on energy distribution and utilization level. Further, from an application on datasets with less strict to almost no applied reprocessing, advice for the filtering of raw data is derived in order to allow for good classification performance. Altogether the results indicate great potential to gain contextual knowledge about data points using the applied multivariate classification techniques, while for generalization purposes further research is necessary.