TY - JOUR

T1 - HVAC System Performance Modeling Using Component-Based Machine Learning

AU - Nabavi, Seyed Azad

AU - Saluz, Ueli

AU - Wolf, Marco

AU - Geyer, Philipp

N1 - Funding Information: In the zone components, there are a wide variety of zone configurations that provides a high generalizability of the trained ML model. Similarly, in the secondary HVAC components, the UA value and the maximum design flow rate is representing most of the water-based secondary HVAC systems. Finally, the primary HVAC systems are represented by the capacity that represents most of the primary HVAC systems. Hereby, the proposed ML models for the components have high generalizability as the components have a limited variety of configurations, unlike monolithic HVAC system models. In future works, we will compare the CBML approach with a monolithic approach in terms of forecasting performance, explainability, and generalizability. Conclusion In this study, the proposed approach for modeling HVAC systems involves breaking HVAC systems down into three components, each of which is modeled using machine learning. This approach allows for better error tracking and explanation of the component models. The proposed ML models for each component have high generalizability due to the limited range of configurations of the zone, secondary HVAC, and primary HVAC components. In this study, each building is composed of a set of zones, secondary HVAC systems, and a primary HVAC system. Accordingly, we developed an ML model for each component. These components are connected in a hierarchical pattern as the zones are modeled then the results are transferred to the next component which is the secondary HVAC system. Finally, all secondary systems ML results are transferred to the primary HVAC system in a building to forecast the primary HVAC systems’ capacity. We used a random forest regression algorithm as a component-based machine learning model to forecast building HVAC systems’ component performance. The random forest regression approach has forecasted the HVAC system components with high accuracy according to the estimated performance metrics for each component. The R-squared value for peak heating demand and annual heating demand in the zone components is 0.99, and 0.97, respectively. Moreover, in the secondary HVAC components, the R-squared values are 0.99 and 0.87 in forecasting the UA value and maximum design flow rate of the secondary HVAC systems. Finally, the ML component has high performance in forecasting primary HVAC systems’ capacity prediction with an R-squared value of 0.98. Acknowledgments We gratefully acknowledge the German Research Foundation’s (DFG) support for funding the project under grant GE 1652/3-2 in the Researcher Unit FOR 2363 and under grant GE 1652/4-1 as a Heisenberg professorship. References Afram, A., & Janabi-Sharifi, F. (2015). Gray-box modeling and validation of residential HVAC system for control system design. Applied Energy, 137, 134–150. https://doi.org/10.1016/j.apenergy.2014.10.026 Chen, X., Guo, T., Kriegel, M., & Geyer, P. (2022). A hybrid-model forecasting framework for reducing the building energy performance gap. Advanced Engineering Informatics, 52.Z hou, X., Hong, T., & Yan, D. (2014). Comparison of HVAC https://doi.org/10.1016/j.aei.2022.101627 system modeling in EnergyPlus, DeST and DOE-2.1E. Publisher Copyright: © 2023 IBPSA.All rights reserved.

PY - 2023

Y1 - 2023

N2 - Heating Ventilation and Air Conditioning (HVAC) systems are responsible for a significant portion of building energy consumption, accounting for up to 38% and 12% of global energy consumption. Predicting energy consumption for HVAC systems is highly important in the early design phases due to their significant impact on energy use and user comfort. However, it is a challenging task due to the complex and dynamic nature of these systems requiring the effort of building simulation. The current state-of-the-art methods for modeling HVAC systems use one data-driven model created by machine learning for the whole HVAC system. In this way, the behavior of the individual HVAC components is neglected and the developed model will have limited explainability and generalizability. The novelty of this study is breaking down HVAC systems into three component categories, which are zone components, secondary HVAC components, and primary HVAC components to not only predict the energy performance of HVAC systems but also the dependencies among them. Then, we apply a component-based machine learning approach to create data-driven models for the HVAC components’ performance. A random forest regression algorithm as a machine learning component serves to predict the performance of HVAC system components in buildings. Machine learning models developed for each component is performing predictions in a hierarchy of receiving/delivering information from other components. In this hierarchical component model, the zone components model informs the secondary HVAC components model of heat distribution components, and the secondary HVAC components model informs the primary model of the heat supply. The random forest approach was highly accurate in predicting component performance, with R2 values of 0.99 and 0.97 for peak heating demand and annual heating demand in the zones, and 0.99 and 0.87 for the maximum design flow rate and UA value of Secondary-HVAC systems. The primary model also had high performance, with an R2 value of 0.98. Compared to conventional data-driven models generated by machine learning, this component-based approach allows for better error tracking and offers explainability. Furthermore, the ML models have high generalizability due to the limited number of parameters and their reusability in further systems.

AB - Heating Ventilation and Air Conditioning (HVAC) systems are responsible for a significant portion of building energy consumption, accounting for up to 38% and 12% of global energy consumption. Predicting energy consumption for HVAC systems is highly important in the early design phases due to their significant impact on energy use and user comfort. However, it is a challenging task due to the complex and dynamic nature of these systems requiring the effort of building simulation. The current state-of-the-art methods for modeling HVAC systems use one data-driven model created by machine learning for the whole HVAC system. In this way, the behavior of the individual HVAC components is neglected and the developed model will have limited explainability and generalizability. The novelty of this study is breaking down HVAC systems into three component categories, which are zone components, secondary HVAC components, and primary HVAC components to not only predict the energy performance of HVAC systems but also the dependencies among them. Then, we apply a component-based machine learning approach to create data-driven models for the HVAC components’ performance. A random forest regression algorithm as a machine learning component serves to predict the performance of HVAC system components in buildings. Machine learning models developed for each component is performing predictions in a hierarchy of receiving/delivering information from other components. In this hierarchical component model, the zone components model informs the secondary HVAC components model of heat distribution components, and the secondary HVAC components model informs the primary model of the heat supply. The random forest approach was highly accurate in predicting component performance, with R2 values of 0.99 and 0.97 for peak heating demand and annual heating demand in the zones, and 0.99 and 0.87 for the maximum design flow rate and UA value of Secondary-HVAC systems. The primary model also had high performance, with an R2 value of 0.98. Compared to conventional data-driven models generated by machine learning, this component-based approach allows for better error tracking and offers explainability. Furthermore, the ML models have high generalizability due to the limited number of parameters and their reusability in further systems.

UR - http://www.scopus.com/inward/record.url?scp=85179525630&partnerID=8YFLogxK

U2 - 10.26868/25222708.2023.1531

DO - 10.26868/25222708.2023.1531

M3 - Conference article

AN - SCOPUS:85179525630

VL - 18

SP - 2680

EP - 2687

JO - Building Simulation Conference Proceedings

JF - Building Simulation Conference Proceedings

SN - 2522-2708

T2 - 18th IBPSA Conference on Building Simulation, BS 2023

Y2 - 4 September 2023 through 6 September 2023

ER -