This study investigates the application of machine learning (ML) for predicting building energy performance at city scale, with a focus on reducing heating and cooling demands under current and future climate scenarios. A two-part methodology was adopted, involving: (i) large-scale building energy simulation and (ii) ML model development. Using Rotterdam, Netherlands as a case study, a computational workflow was created to automate data collection, processing, and energy simulation for 20,000 residential buildings under both the present and two projected climate conditions. Results highlight the influence of building layout, envelope thermal properties, and air tightness on reducing energy demand across a diverse range of building archetypes. An artificial neural network (ANN) was subsequently developed to enable rapid prediction of energy demands for both existing building conditions and retrofit scenarios. The analysis demonstrates that a shallow ANN is an effective ML model in terms of time efficiency, usability, and accuracy, particularly for predicting heating demands. The study highlights both the strengths and limitations of ML-based approaches relative to traditional energy modelling, offering valuable insights for energy planning and targeted retrofit decision-making at city-scale.

The scripts created for the main computational workflow of this project are shared to the following github repository: https://github.com/elenarduzzi/buildingenergymetamodels