START Conference Manager    

Modelling energy labels using open data and machine learning

Sanne Hettinga

(Submission #289)


National governments are opening up (3D) geospatial data, both of their own accord and in light of the INSPIRE project. This data contains a great deal of information by itself. By using it as input data for geospatial models, this information can be enriched and used to generate smart solutions for current day challenges. The energy transition is one of the challenges that requires such smart data solutions. One of them is described in the European Energy Performance of Buildings Directive (EPBD), which describes an Energy Index and an energy label that is to be assigned to all buildings in the European Union. Most countries are still in the process of assigning these labels. In the Netherlands, only 33% of dwellings have a registered energy label. To enable national governments to assign an estimated energy label to dwellings, such that they can develop policy to stimulate the refurbishment of these dwellings, the TABULA project has come up with national building typologies that give an indication of the energy performance of dwellings and an indication of its Energy Index and energy label. However, after assigning the according energy labels to all Dutch dwellings that already have a certified energy label, actually only 21% of the energy labels appeared to be predicted correctly. This study investigates whether we can use open data and machine learning techniques to increase the percentage of correct predictions. We started to enrich the input data with geospatial data such as actual exposed perimeter and actual surface area. However, the accuracy of predicting the energy label decreased to the even worse 16%. Therefore, we abandoned the use of the TABULA building typology and decided to use machine learning in combination with a large variety of (3D) geospatial data, that cover building information, socio-economic information and data on energy consumption. The advantage of using machine learning is that it is able to find patterns in data that are difficult to find manually. It is therefore more suited to classify each dwelling in the Netherlands into the proper energy label class. Initial research has indicated that with a simple machine learning algorithm, the energy labels can be predicted with an accuracy of 76%. Further research into the optimization of the use of machine learning algorithms is ongoing.


Topic Area:  [1.9] Energy (renewable – non-renewable)
Abstract Type:  Oral Presentation

Additional Fields

Academic:   Yes
Data Provider:   Yes
Data User:   Yes
INSPIRE Implementer (IT):   No
INSPIRE newbies:   No
Policy Officers:   No
Public Administration (MS/Regional/Local):   No
Thematic specialists:   Yes
Comments:   Energy and Location, Energy Labels, Machine learning

START Conference Manager (V2.61.0 - Rev. 5269)