Deep Transfer Learning for Thermal Dynamics Modeling in Smart Buildings

Zhanhong Jiang, Young M. Lee

Abstract

Modeling building thermal dynamics is a prerequisite for advanced HVAC control and energy management. However, the high variability between buildings and the lack of historical data in newly commissioned structures—the "cold start" problem—hinder the deployment of data-driven models. This paper introduces a Deep Supervised Domain Adaptation (DSDA) framework based on Sequence-to-Sequence (S2S) Long Short-Term Memory (LSTM) networks. By transferring knowledge from a data-rich source building to a data-poor target building, the proposed method significantly enhances prediction performance for indoor temperature and energy consumption.

Introduction

Traditional physics-based models for building thermal dynamics are often difficult to generalize due to the unique characteristics of every building, such as geometry and materials. Conversely, deep learning models require large datasets that are typically unavailable for new buildings. This research addresses these challenges by exploring transfer learning: the ability of a model to adapt its understanding of thermal behavior from one environment to another using minimal local data.

Methodology

The core of the proposed approach is a Deep Supervised Domain Adaptation (DSDA) framework. The underlying architecture utilizes an LSTM Sequence-to-Sequence (S2S) model, consisting of an encoder that compresses input time-series features (e.g., ambient temperature, humidity) into a latent vector, and a decoder that generates future state predictions.

The transfer process involves two main stages:

  1. Offline Pre-training: The S2S model is trained on a large-scale dataset from a "Source Building" to capture general thermal dependencies.
  2. Domain Adaptation (Fine-tuning): The model parameters are initialized with the source weights and then fine-tuned using a small amount of data from the "Target Building."

To ensure the model learns robust dynamics rather than simple temporal correlations, the authors employ "non-teacher forcing" during training, requiring the model to rely on its own previous predictions.

The DSDA framework workflow, illustrating the transition from source building offline learning to target building adaptation.

Experimental Setup

The framework was validated using four public datasets across two primary case studies:

  • Temperature Evolution: Transferring from the AHU dataset (35,098 samples at 1-min intervals) to the SML dataset (1,373 samples at 15-min intervals).
  • Energy Consumption: Transferring from Building 2 (34,940 samples) to Building 1 (2,000 samples), both at 15-min intervals.

Crucially, the method proved effective even when the source and target buildings had different data sampling frequencies or were used for different prediction tasks.

Results and Discussion

The DSDA approach consistently outperformed models trained from scratch across all evaluated metrics, including CVRMSE, NMBE, MAPE, and RMSE.

In the temperature prediction task, the RMSE for a 15-minute horizon on the SML dataset was reduced from 1.274 (scratch) to 0.355 (DSDA). Similar improvements were observed for energy consumption, where the 6-hour ahead RMSE for Building 1 dropped from 26.600 to 16.646.

Performance metrics for temperature prediction on the SML dataset, comparing the scratch model vs. the DSDA (AHU -> SML) approach.

A notable finding was cross-task transferability: a model pre-trained on energy consumption data could be successfully adapted to predict indoor temperature (Building 2 → SML), achieving an RMSE of 0.403. This suggests the model captures fundamental physical dependencies that are invariant to the specific target variable.

Conclusion

Deep transfer learning via DSDA provides a viable solution to the data scarcity problem in smart building management. By reusing learned representations, buildings can achieve high-fidelity thermal modeling shortly after commissioning. Future research will focus on multi-source transfer learning and unsupervised adaptation to further reduce the reliance on labeled target data.

Summary updated

BibTeX
@inproceedings{jiang2019deep,
  author = {Zhanhong Jiang and Young M. Lee},
  title = {{Deep Transfer Learning for Thermal Dynamics Modeling in Smart Buildings}},
  booktitle = {IEEE BigData},
  year = {2019},
  doi = {10.1109/bigdata47090.2019.9006306},
}
Figure