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

AI summary

Overview

Research area: machine learning for smart buildings, specifically time-series forecasting of indoor zone air temperature in a building with a variable air volume (VAV) HVAC system. Technical level: Intermediate. Scope: a case study using sensor data from a smart building to develop a 2-week-horizon zone air temperature forecast model and explore HVAC control applications.

What This Paper Is About

HVAC systems need fast and accurate indoor temperature predictions to support energy-efficient and flexible control, but simulation tools such as EnergyPlus and DOE2 require high computing power and lack flexibility when input parameters change. Data-driven models can be faster, but they often need many input features, which makes reliable datasets harder to obtain and model predictions harder to understand. This paper develops a hybrid data-driven and thermodynamic time-series forecast model for zone air temperature over a 2-week horizon, using selected sensor features from a smart building and examining spatial differences across zones.

Key Contributions

1. Long-horizon auto-regressive forecasting: develops a 2-week zone air temperature forecast model using multivariate auto-regressive regression, with XGBoost combined with self-regression selected as the final model.
2. Hybrid data-driven and thermodynamic feature set: uses outside air temperature, actual zone air temperature, actual zone air temperature setpoint, and actual supply pressure setpoint of the last n weeks as inputs, reducing the feature burden after preprocessing.
3. Spatial-temporal zone investigation: selects three VAVs in different locations to illustrate performance differences and discusses location and orientation corrections for different thermal zones, such as perimeter versus core zones.
4. Potential application framework: outlines energy flexibility optimization and data-driven model predictive control for HVAC setpoints using flexible tariffs and convex optimization.

Main Findings

• Best-performing model: XGBoost combined with self-regression was selected to predict zone air temperature with a horizon of 2 weeks.
• Error metrics: the average MAE is 4.20 and the average RMSE is 4.8, indicating an air temperature error of approximately 4 °F.
• Predicted temperature range: for three selected VAVs over 2 weeks, zone air temperature fluctuated between 65°F and 75°F.
• Selected VAVs: VAV RH 2-2-50 with No. 2620112368775269 on the 2nd floor; VAV RH 1-1-19 with No. 2622037806906769 in the north room on the first floor; and VAV CO 1-1-51 with No. 2758068039436455 in the south room on the first floor.
• VAV performance variation: prediction performance differed among the three VAVs. Two potential reasons are that the model is independent of spatial location, which may affect the impact of solar radiation that is not an input parameter, and that devices have different functions.
• Data scale: the dataset spans January 2022 to June 2024 with five datasets, 2022a, 2022b, 2023a, 2023b, and 2024a, totaling more than 2GB and more than 1,000 variables.
• Temporal resolution: the original dataset timestep is 5 minutes, and the authors down-sampled the data to develop models with a timestep of 15 minutes and 1 hour.
• Data splits: for 2022 and 2023, the first 6 months of each year were used as the train dataset and the last 6 months as the validation dataset; the first half of 2024 was used as the test set.
• Per-VAV metrics: individual MAE and RMSE values for the three selected VAVs are not reported; the paper provides aggregate metrics and illustrative figures.

Methodology in Plain English

The authors collected data from a smart building equipped with a typical all-air VAV system consisting of a chilled and cooling water system, an air handling unit, and multiple VAV units in each thermal zone. The dataset includes observation values and action values of the HVAC system, outside weather data, HVAC device information, and the floor plan with device distribution.

They preprocessed the data with feature selection, missing data imputation, and outlier removal. For each zone air temperature forecast, the input features were outside air temperature, actual zone air temperature, actual zone air temperature setpoint, and actual supply pressure setpoint of the last n weeks, with the output being the predicted temperature for the next n weeks. Because the original 5-minute timestep leads to frequent predictions that may not be useful given thermal mass effects in building envelopes, they also down-sampled the data and developed models with a 15-minute and 1-hour timestep.

The data were split into train, validation, and test sets as described above, and sliced into 2-week durations to support long-horizon predictions. The authors compared several regression algorithms, including random forecast regression, Gaussian process regression, Adaboost regression, gradient boosting regression, and XGBoost regression, all developed with a multivariate auto-regressive approach. They used a hierarchical prediction module and selected the best model with a model selection algorithm called time-series cross-validation. Mean absolute error (MAE) and root mean square error (RMSE) were used to evaluate forecast performance. In the first phase, they assumed that temperature changes in each VAV zone are independent and trained the model with data from each individual VAV; in the second phase, they planned model correction and validation based on spatial location, because heat transfer differences between core and perimeter zones affect how outdoor temperature influences indoor zone air temperature.

Why This Matters

This work contributes to research on hybrid building energy modeling and long-horizon indoor temperature forecasting, showing how a data-driven model with selected features can be combined with thermodynamic principles for intelligent HVAC control and demand flexibility.

Real-world applications include:

• HVAC demand flexibility and energy cost savings under hourly flexible tariffs, using building thermal mass for load shifting.
• Data-driven model predictive control for zone air temperature and pressure setpoints, with optimized actions transmitted in real time using industrial protocols such as Modbus or BACNet.
• Scalable cloud solutions for data-driven model predictive control in commercial and industrial buildings that include devices such as batteries, thermal storage, and PV panels in the demand-supply energy system.
• Support for hybrid building energy modeling and downstream tasks such as determining thermal comfort and reducing cooling and heating energy consumption in smart buildings.

Industry relevance: the work is relevant to smart building operators, HVAC control vendors, building energy modelers, energy management companies, and machine learning practitioners working on time-series forecasting for building systems. The author affiliations include the Bosch Center for Artificial Intelligence, Iowa State University, and Tongji University, indicating direct relevance to industrial HVAC and building automation contexts.

Future Directions

• Incorporate the spatial location and orientation of each thermal zone to account for solar radiation and differences between perimeter and core zones.
• Account for different device functions across VAV boxes, which may explain performance differences among the three selected VAVs.
• Further optimize the model for accuracy, as stated in the impact statement.
• Scale and validate the data-driven model predictive control system with cloud solutions across all commercial and industrial buildings and related devices, including batteries, thermal storage, and PV panels.
• Explore relationships between parameters to simplify data collection, reduce data sampling cost, and improve transferability of research results.

Target Audience

This paper benefits HVAC engineers, building energy researchers, smart building control engineers, facility managers, energy flexibility analysts, and machine learning or data science practitioners working on time-series forecasting for indoor environmental control.

Authors’ abstract

With the press of global climate change, extreme weather and sudden weather changes are becoming increasingly common. To maintain a comfortable indoor environment and minimize the contribution of the building to climate change as much as possible, higher requirements are placed on the operation and control of HVAC systems, e.g., more energy-efficient and flexible to response to the rapid change of weather. This places demands on the rapid modeling and prediction of zone air temperatures of buildings. Compared to the traditional simulation-based approach such as EnergyPlus and DOE2, a hybrid approach combined physics and data-driven is more suitable. Recently, the availability of high-quality datasets and algorithmic breakthroughs have driven a considerable amount of work in this field. However, in the niche of short- and long-term predictions, there are still some gaps in existing research. This paper aims to develop a time series forecast model to predict the zone air temperature in a building located in America on a 2-week horizon. The findings could be further improved to support intelligent control and operation of HVAC systems (i.e. demand flexibility) and could also be used as hybrid building energy modeling.