The need for energy efficiency in neighborhood-scale architectural design is driven by environmental imperatives and escalating energy costs. This study identifies three key phases in a design process framework where machine learning can be applied to optimize energy consumption in early design stages. The overall framework integrates machine learning tools into the design workflow, enhancing design exploration from concept level and enabling targeted energy assessments. This paper focuses on the first phase (Phase 1) of the framework, which employs machine learning for building energy forecasting using only the few inputs available in a business-as-usual early-stage design workflow. The CatBoost model was selected for its high accuracy in predicting energy consumption using minimal input data. A preliminary application to a case study in New York City showed high predictive accuracy while reducing the input needed, with R2 scores of 0.88 for both cross-validation and test datasets. Shapely additive explanation analysis validated the selection of key influencing parameters such as building area, principal building activity, and climate zones. The test demonstrated discrepancies between the test data-driven model and a physics-based energy model values ranging from −8.69% to 11.04%, which can be considered an acceptable result in early-stage design. The remaining two phases, though outside the scope of this study, are introduced at a conceptual level to provide an overview of the full framework. Phase 2 will analyze building shape and elevation, assessing the total energy use intensity, while Phase 3 will apply district-level energy optimization across interconnected buildings. The findings from Phase 1 underscore the potential of machine learning to integrate energy efficiency considerations into neighborhood-scale design from the earliest stages, providing reliable predictions that can inform sustainable design.
