AI‑Powered Measurement & Verification: Building Interpretable Counterfactual Models to Verify Energy Savings in Buildings (Tutorials Track) Spotlight
Benedetto Grillone (Ento)
Abstract
The building sector is a critical frontier for climate action, but a key barrier to scaling energy efficiency retrofits is the challenge of reliably verifying their impact. This tutorial provides a hands-on, end-to-end guide to applying machine learning for Measurement and Verification (M&V), transforming energy savings from an ambiguous claim into a verifiable, investment-grade asset. Using a real-world dataset of hourly meter and weather data, participants will learn to build a robust counterfactual energy baseline with a LightGBM (Gradient Boosting Machine) model. The workflow covers essential practical steps, including feature engineering for occupancy proxies, automated detection of non-routine events, and time-series cross-validation. The tutorial moves beyond simple prediction to address two essential aspects of real-world M&V: uncertainty quantification, where we calculate a statistically rigorous confidence interval for the savings estimate in alignment with ASHRAE Guideline 14; and model interpretability, where we use global and regional Partial Dependence Plots (PDPs) via the effector library to validate that our model has learned physically plausible relationships, ensuring our results are both accurate and defensible. By the end of this notebook, users will have implemented a complete workflow to transform raw time-series data into an investment-grade savings report. This tutorial aims to equip practitioners with the practical skills needed to bridge the gap between advanced machine learning and its application in building decarbonization, empowering them to develop the transparent and scalable solutions required for the climate transition.