Automatic Selection of Predictive Models for Building Energy Demand

Buildings account for more than one-third of total energy demand in many countries worldwide, positioning them as one of the major contributors to greenhouse gas emissions. Effective prediction of building energy demand is essential for optimizing electricity consumption and mitigating power shortages or waste. The importance of this task is increasing due to two significant transitions: the electrification of heating and transportation and the growing dependence on renewable energy sources. The former refers to the widespread adoption of electric vehicles and electric heating systems, such as heat pumps. The latter involves the expansion of renewable energy generation, which introduces challenges in balancing supply and demand, as electricity from sources like wind and solar is inherently intermittent and not always available when consumption peaks. The principal objective of this project, building further on previous work at KU Leuven University, Belgium, was to develop machine learning models to create building energy demand forecasts for the future (day-ahead time scale). It therefore focuses on establishing a complete procedure for energy demand prediction including preprocessing, forecasting, data analysis, time series feature extraction and model recommendation. A crucial task in time-series forecasting is the identification of the most suitable forecasting model. The challenges associated with energy demand forecasting arise from a wide range of predictive algorithms and models can be adapted through minor or significant modifications which is time intensive. Additionally, the diversity of buildings, each with distinct energy consumption patterns, further complicates the forecasting process. Consequently, the project first constructs a comprehensive dataset comprising buildings with diverse energy demand patterns to evaluate the performance of various forecasting models across these buildings. Then, based on this analysis, it designs a recommender system to identify and suggest the most suitable forecasting models for a newly encountered building. To implement this recommender system effectively, time series feature extraction is a crucial step. Feature extraction facilitates the summarization of long time-series data into a lower-dimensional set of key values or statistics, thereby reducing data complexity while preserving essential patterns and information. Finally, based on the developed dataset comprising five distinct building energy demand profiles along with their extracted features and the performance evaluation of over thirty forecasting models on these buildings, the recommender system is designed to identify the most suitable forecasting models. When the features of a previously unseen building are input into the system, it recommends the forecasting models that best align with the specific characteristics of that building. The evaluation results indicate that the top-performing models outperformed the baseline methods by approximately 12%, demonstrating the advantage of advanced forecasting techniques and the recommender system achieved an accuracy of approximately 0.05 in terms of MAE, (the average discrepancy between the predicted and actual values), highlighting its capability to provide reliable model recommendations based on the extracted building features.