This is like building a smart "thermostat brain" for a building: it studies how past energy use changes with weather and other factors, then uses that learning to predict how much energy the building will need in the future.
Architects, engineers, and building owners need to understand and predict how much energy a building will consume under different conditions (weather, occupancy, retrofit options) without running expensive simulations or waiting for years of utility bills. This work compares multiple regression algorithms and applies the best ones to create accurate multivariable energy signatures that can be used for design decisions, retrofits, and operational optimization.
Domain-specific feature engineering and access to high-quality, longitudinal building energy and weather datasets can create a moat; otherwise, the underlying regression methods are widely available and commoditized.
Classical-ML (Scikit/XGBoost)
Structured SQL
Medium (Integration logic)
Data quality and representativeness of historical building energy and weather data; model performance will be limited if meters, BMS points, or weather feeds are sparse, noisy, or inconsistent.
Early Majority
Compared to typical single-variable energy signatures (e.g., only using outdoor temperature), this approach evaluates multiple regression algorithms on multivariable inputs (weather, schedules, building properties, etc.), enabling more accurate and robust models that are directly applicable to real buildings in operation.