Time-series Forecast for Indoor Zone Air Temperature with Long Horizons: A Case Study with Sensor-based Data from a Smart Building

Liping Sun, Yucheng Guo, Siliang Lu, Zhenzhen Li

Abstract

Heating, Ventilation, and Air Conditioning (HVAC) systems are significant energy consumers and major contributors to carbon emissions. Efficient management is often hampered by the limitations of traditional physics-based simulations and purely data-driven models, especially for long-term forecasting. This study presents a hybrid "gray-box" model utilizing XGBoost and thermodynamic principles to predict indoor temperatures up to two weeks in advance. Using real-world sensor data from a smart building, the model achieves high accuracy (MAE of 4.20°F), enabling proactive energy load shifting and cost reduction while maintaining occupant comfort.

Introduction

HVAC systems are critical to building operations but contribute substantially to global carbon emissions. As extreme weather events become more frequent, the need for efficient HVAC control grows. Current modeling approaches often fall into two categories: computationally expensive "white-box" physics simulations that lack adaptability, and "black-box" data-driven models that require excessive input variables and struggle with interpretability. There is a notable gap in providing accurate, long-horizon (e.g., 14-day) temperature forecasts necessary for advanced energy optimization.

Methodology

The researchers developed a hybrid gray-box approach that bridges the gap between physical reality and statistical efficiency. By integrating domain knowledge of building thermodynamics with machine learning, the model reduces the complexity of required inputs.

Overview of the algorithm-optimized modeling workflow for zone air temperature prediction.

Key technical strategies included:

  • Algorithm Selection: After testing various regression models, XGBoost combined with self-regression was identified as the most effective performer.
  • Feature Engineering: The model focuses on high-impact variables: outdoor air temperature, actual zone temperature, air flow rates, and heating/cooling setpoints.
  • Temporal and Spatial Scaling: Data was downsampled to 15-minute and 1-hour intervals to reflect the building's thermal mass. The model also distinguishes between "core" and "perimeter" zones to account for varying exposure to external conditions.

Data and Implementation

The study utilized a comprehensive dataset from a smart building equipped with a Variable Air Volume (VAV) system, spanning January 2022 to June 2024. The dataset included over 1,000 variables and 2GB of sensor readings. Training and validation were conducted on 6-month blocks, with the first half of 2024 serving as the final test set for the 14-day forecasting window.

Results and Discussion

The XGBoost-based self-regression model demonstrated remarkable performance over the 14-day horizon, achieving a Mean Absolute Error (MAE) of 4.20°F and a Root Mean Square Error (RMSE) of 4.8°F.

A 2-week forecast comparing true versus predicted zone air temperature, demonstrating high fidelity over a long horizon.

The model's ability to track temperature fluctuations between 65°F and 75°F across different building zones highlights its potential for real-world application. By leveraging the building's thermal mass, these long-term forecasts allow HVAC systems to "pre-cool" or "pre-heat" during off-peak hours, optimizing for both cost and carbon footprint.

Conclusion

While successful, the model's accuracy could be further enhanced by incorporating solar radiation data, particularly for south-facing perimeter zones. Future work will also explore the interdependencies between VAV boxes and the scalability of these gray-box models for cloud-based deployment across commercial districts. This research provides a robust foundation for the next generation of smart building energy management systems.

Summary updated

BibTeX
@inproceedings{sun2025time,
  author = {Liping Sun and Yucheng Guo and Siliang Lu and Zhenzhen Li},
  title = {{Time-series Forecast for Indoor Zone Air Temperature with Long Horizons: A Case Study with Sensor-based Data from a Smart Building}},
  booktitle = {CO-BUILD @ ICML},
  year = {2025},
  url = {https://openreview.net/forum?id=ozxp9aQeOK},
}
Figure