Transfer learning with deep neural networks for model predictive control of HVAC and natural ventilation in smart buildings
2020
Abstract Advanced control strategies are central components of smart buildings. For model-based control algorithms, the quality of the model that represents building systems and dynamics is essential to guarantee satisfactory performance of smart building control and automation. For the model predictive control of the heating, ventilation, and air conditioning systems in buildings coupled with natural ventilation, a high-fidelity model is necessary to reliably predict the thermal responses of the building under various environmental and operational conditions. This task can be accomplished by using a deep neural network, which can capture the dynamics of complicated physical processes, such as natural ventilation. Training a deep neural network requires the collection of a large amount of data; however, in practice, the target building may not have enough operational data available. This study demonstrates how transfer learning could help with this dilemma. By freezing most layers of a deep neural network model with 42,902 parameters that are pre-trained on multi-year data from a source room in Beijing, the model can be re-trained with only 200 trainable parameters on only 15 days of data from the target room in Shanghai that has entirely different floor area, building material, and window size. The proposed transfer learning model achieves high accuracy predicting both indoor air temperature and relative humidity for a time horizon from 10 minutes to 2 hours, showing the mean squared error almost one magnitude smaller than the comparison model that is only trained on source data or target data. This methodology can be applied to the design of the control system in a new building which reduces the required amount of data for the training of the model, thus saving costs in control system design and commissioning.
Keywords:
- Correction
- Source
- Cite
- Save
- Machine Reading By IdeaReader
33
References
66
Citations
NaN
KQI