Active learning cuts building energy model errors by 54% in new study

A team of researchers from an undisclosed institution (Nam T. Nguyen and Truong X. Nghiem) has released a comprehensive study on arXiv (2606.25301) comparing active learning (AL) strategies for optimal experimental design in building energy system identification. The paper addresses a critical bottleneck in machine learning-based modeling: the quality of training data. Instead of relying on uniformly random inputs (passive learning), AL algorithms intelligently select which experiments to run to maximize information gain. The study tested 14 different AL acquisition functions across two model classes—deterministic feedforward neural networks and stochastic Gaussian processes—and categorized them into four groups: data space, uncertainty, information gain, and model change algorithms.

Using the high-fidelity building simulator BOPTEST, the team evaluated models on HVAC thermal dynamics under various initial dataset sizes and control input constraints. Results showed that AL-based models consistently outperformed passive learning baselines, with error reductions reaching up to 54% in root mean square error. However, the magnitude of improvement varied significantly across acquisition functions and operating regimes. The findings provide practical guidance for deploying data-efficient ML models in smart building control, potentially reducing the cost and time needed to train accurate energy system models for real-world applications.