ML-Powered Energy Label Prediction System for Residential Buildings in the Netherlands

Many residential buildings in the Netherlands still lack an official energy label, making it challenging for organizations to plan renovations and contribute to national climate targets. The AMI Research Group previously explored the use of open data for predicting energy labels but faced fragmented data handling, manual workflows, and no mechanism to deliver results to operational partners. As a result, predictions could not be easily accessed or applied in practice. This project addressed these challenges by designing and implementing a complete, automated, and explainable machine learning system that transforms the AMI research group’s exploratory work into a functional platform for energy label prediction and data exploration. The system integrates Dutch open datasets—EP-Online, the Common Data Factory, and 3DBAG—within a modular pipeline built with Dagster, automating data ingestion, preprocessing, model training, and deployment. Experiments were conducted on approximately 500,000 dwellings with official energy labels. The Random Forest Classifier achieved 82% exact-label accuracy, an ordinal MAE of 0.29 label steps, and ±1-label accuracy of 95%. Compared to the related Dutch study by Hettinga et al., which achieved 71% exact and 93% ±1 accuracy on a larger national dataset, the results in this project indicate that the chosen model and feature set perform competitively on a smaller subset of Dutch dwellings. While further validation on the full national dataset is needed, the outcomes demonstrate the system's technical feasibility and potential usefulness for large-scale energy-label prediction. Predictions and SHAP-based explanations are made accessible through a Next.js web application, allowing users to search for addresses, visualize predicted labels on an interactive map, and understand why the model assigned each label. Feedback from stakeholders confirmed that the web application was practical for exploring predictions and that the explanations were understandable for non-technical users.