Register now After registration you will be able to apply for this opportunity online.
Transfer Learning for Building Thermal Modeling
Buildings appear as significant energy consumers, especially due to the management of heating, ventilation, and air-conditioning (HVAC). Each building has unique characteristics such as varied geometries, floor layouts, construction properties, age, climatic regions, orientation, and service systems. Better control of indoor temperature in buildings seems to be a means of energy savings. Traditional approaches rely on building modeling for this purpose.
While physics-based models may be precise and aligned with expected physical behaviors, their complex design can limit their application and scalability.
An alternative modeling approach based solely on sensor data (temperature, solar irradiance, etc.) aims to be more flexible and is generating increasing interest. However, these approaches require diverse data in sufficient quantity to train the model parameters and might demand more computing power than what buildings can accommodate.
The complexity of models, their instability, or the lack of data poses obstacles when attempting to model a new building.
The primary goal of this project is to leverage the flexibility of data-driven methods to model the thermal behavior of buildings, emphasizing the development of a transferable model.
This approach aims to streamline the modeling process by enabling the initial learning of a model for one building and its subsequent adaptation to other buildings.
This study, associated with the Euthermo Project (https://www.sairop.swiss/projects/liste/scalable-deep-reinforcement-learning-algorithms-for-building-climate-control-and-energy-management), focuses on leveraging data-driven methodologies to model the thermal behavior of buildings.
The primary objective is to explore the transferability of a building temperature simulation model to other buildings. This involves comparing existing models, developing embedding methods and transfer learning techniques, as well as designing relevant evaluation metrics.
Retraining a model from scratch for each new building is unsatisfactory.
This project's innovative approach aims to enhance the scalability of the modeling process, thereby reducing the time and resources required compared to the usual practices of training separate models for each building. The overarching goal is to contribute to improved energy efficiency in buildings within the broader context of the global energy transition.
This study, associated with the Euthermo Project (https://www.sairop.swiss/projects/liste/scalable-deep-reinforcement-learning-algorithms-for-building-climate-control-and-energy-management), focuses on leveraging data-driven methodologies to model the thermal behavior of buildings. The primary objective is to explore the transferability of a building temperature simulation model to other buildings. This involves comparing existing models, developing embedding methods and transfer learning techniques, as well as designing relevant evaluation metrics. Retraining a model from scratch for each new building is unsatisfactory. This project's innovative approach aims to enhance the scalability of the modeling process, thereby reducing the time and resources required compared to the usual practices of training separate models for each building. The overarching goal is to contribute to improved energy efficiency in buildings within the broader context of the global energy transition.
- Conduct a concise comparison of building thermal models, encompassing literature algorithms such as RC model, ARX, PCNN, and other relevant techniques.
- Develop a data-driven algorithm for building modeling that can be effectively transferred to new buildings, eliminating the need for completely retraining the model for each unit.
- Evaluate the proposed methodology using data associated with the NEST building located in Dübendorf, Switzerland.
- Conduct a concise comparison of building thermal models, encompassing literature algorithms such as RC model, ARX, PCNN, and other relevant techniques. - Develop a data-driven algorithm for building modeling that can be effectively transferred to new buildings, eliminating the need for completely retraining the model for each unit. - Evaluate the proposed methodology using data associated with the NEST building located in Dübendorf, Switzerland.