| Abstract [eng] |
Buildings account for about 36 % of all greenhouse gas emissions. To lower the CO2 emissions of buildings it is important to be able to evaluate the thermal losses of buildings as efficiently as possible. Although thermal images are used in this context for detecting cold or hot bridges, they can also be utilized for creating three-dimensional representations of buildings. In the methodology of this work, an approach for using such thermophotogrammetry models for building heat loss evaluation is studied. A dataset for this task was created, consisting of two different captures using a DJI Mavic 2 Enterprise drone: one in March, and one in August. The dataset consists of in total 654 thermal images, and the thermophotogrammetry models that are generated from these thermal images consist of about 12 000 000 points. Surface temperature values were extracted from thermal images by assigning a RGB value to each surface temperature value. After converting the thermophotogrammetry model to a thermal point-cloud, machine learning algorithms for point-cloud segmentation can be used. After analyzing three different building segmentation algorithms it was determined that KPConv provided the best results. The segmentation results were further refined by statistical outlier detection, which slightly improved their accuracy. Segmented point-clouds were converted to a mesh and each mesh triangle was assigned a surface temperature value to its‘ surface area based on RGB values. For evaluating the heat loss calculations, results from a white-box simulation were used, in which the heat losses of the building were evaluated using geometry and material properties. Heat loss evaluation results were more accurate for the non-heating season. The predicted heat loss for the heating season is higher than the actual heat losses. The average error for heat loss calculations was -11 kW for the heating season and 0,3 kW for the non-heating season. When applying the methodology described in this work it is possible to perform heat loss evaluations by using deep learning-based methods and thermophotogrammetry. The results are more accurate during the non-heating season, but during the heating season it is still possible to identify that there are heat losses occurring. The methodology can be useful for interested parties regarding the CO2 footprint of buildings, and it can be applied for more larger scale heat loss evaluations, e.g., neighborhood or city-scale. It is especially relevant nowadays when the fight against climate change is important and often reiterated via the United Nations Sustainable Development Goals and the European Green Deal. Further works could explore creating a more accurate building segmentation model that would be trained on a thermal dataset. Also, different facade elements could be segmented with the goal of applying different heat loss coefficients and emissivity parameters. This could lead to a more accurate evaluation of surface temperatures. |
