Minute-Level Dataset from a Naturally Ventilated Building for Benchmarking and Learning-Based Modeling

This paper presents a one-year, one-minute interval dataset from HouseZero^®, a naturally ventilated and ultra-low-energy building in Cambridge, MA, USA. Filtering techniques were developed to process millions of raw data points recorded from 190 sensors and meters installed throughout the building. The filtering process detects and removes erroneous data through three stages: (1) a system error filter, that identifies faults in the building management system (BMS); (2) a subsystem error filter, that detects errors from devices between the BMS and the sensors; and (3) a sensor-level filter, that flags errors at the individual sensor level. To ensure access to up-to-date data, an automated algorithm processes the incoming data on a weekly basis and stores the results in a database. Various visualizations are employed to support technical validation of the data, emphasize its fidelity, and examine relationships among key features. This dataset from HouseZero^® is intended to serve as a benchmark for low-energy buildings, enhance data analysis methodologies, and support data-driven, learning-based approaches to building thermal modeling.