Development of a model for analysis of census data using statistical methods and machine learning

Abstract

In the current energy crisis faced by Germany, building an energy map holds great importance for effective energy management and decision-making. This thesis contributes relevant information for constructing an energy map by classifying the number of residents per building. Building population estimation holds importance to identify the energy trends and patterns of building occupants which plays a significant role in managing energy needs and formulating appropriate energy management strategies. The study utilizes two primary datasets, namely 3D building models for the particular study area and the 2011 German census data. Due to data privacy policies, information about individual building occupants is unavailable and this particular information is important to understand not just the final energy demand per building or household but also understand energy consumption behaviour patterns. This thesis addresses this challenge by developing a specialized and comprehensive databank that integrates the building models with census data on population, housing, and buildings. Additionally, to classify the building’s number of residents, a supervised machine learning approach using the XGBoost algorithm was employed. In the absence of training datasets due to data laws, a synthetic training dataset was prepared by incorporating various building types such as single-family houses, multi-family houses, and other types. Three separate models were constructed, and their results were subsequently combined to enhance accuracy. The evaluation of accuracy per grid cell provides insights into the effectiveness of the classification process. The developed classification model showed an average accuracy of 62%, where the models are closely related to the building form classification. The findings of this research have implications for better energy management and decision-making in the face of the country’s energy crisis.