Adapting to Change: A Comparison of Continual and Transfer Learning for Modeling Building Thermal Dynamics under Concept Drifts

Transfer Learning (TL) is currently the most effective approach for modeling building thermal dynamics when only limited data are available. TL uses a pretrained model that is fine-tuned to a specific target building. However, it remains unclear how to proceed after initial fine-tuning, as more operational measurement data are collected over time. This challenge becomes even more complex when the dynamics of the building change, for example, after a retrofit or a change in occupancy. In Machine Learning literature, Continual Learning (CL) methods are used to update models of changing systems. TL approaches can also address this challenge by reusing the pretrained model at each update step and fine-tuning it with new measurement data. A comprehensive study on how to incorporate new measurement data over time to improve prediction accuracy and address the challenges of concept drifts (changes in dynamics) for building thermal dynamics is still missing. Therefore, this study compares several CL and TL strategies, as well as a model trained from scratch, for thermal dynamics modeling during building operation. The methods are evaluated using 5–7 years of simulated data representative of single-family houses in Central Europe, including scenarios with concept drifts from retrofits and changes in occupancy. We propose a CL strategy (Seasonal Memory Learning) that provides greater accuracy improvements than existing CL and TL methods, while maintaining low computational effort. SML outperformed the benchmark of initial fine-tuning by 28.1% without concept drifts and 34.9% with concept drifts.

💡 Research Summary

This paper tackles the practical problem of keeping data‑driven building thermal dynamics models accurate over the long term as new operational measurements become available. While Transfer Learning (TL) has become the de‑facto method for initializing such models with limited data, the authors ask how to update the model once the building is in service, especially when the underlying dynamics change due to retrofits or occupancy shifts (concept drifts).

Four update strategies are compared: (1) training a model from scratch each time enough new data have been collected; (2) repeatedly fine‑tuning the pretrained TL model with all accumulated data (TL‑based adaptive learning); (3) applying various Continual Learning (CL) algorithms that aim to preserve past knowledge while learning from new data; and (4) a novel CL approach called Seasonal Memory Learning (SML). An additional hybrid method, event‑based Accumulative Learning on General model (eALG), is also introduced but performs slightly below SML.

The experimental platform consists of high‑fidelity simulations generated with the BuilDa framework (Modelica FMU). Eight single‑family houses typical of Central Europe are simulated for 5–7 years, providing 15‑minute resolution data on indoor/outdoor temperature, HVAC operation, solar radiation, wind, and occupancy. Two types of concept drifts are injected: (a) retrofits (e.g., insulation upgrades, window replacements) and (b) occupancy changes (different number of occupants and schedules). A large‑scale scenario with 40 houses over seven years is also evaluated to test generality.

For TL, a generic model is pretrained on 450 simulated houses; the remaining houses serve as test targets. The adaptive learning loop accumulates data over time, and updates are performed at different frequencies (monthly, quarterly, yearly). Performance metrics include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), update‑time, and memory consumption.

Key findings:

• Accuracy – In drift‑free scenarios SML reduces MAE by 28 % relative to the initial TL model; with concept drifts the reduction reaches 34.9 %. Existing CL methods (EWC, GEM, AR, IL) improve accuracy by 15–20 %, while TL‑based adaptive learning yields 22–25 % improvement but at a much higher computational cost. Training from scratch catches up after about one year of data but suffers large errors in the early months.

• Computational efficiency – SML stores a small seasonal replay buffer (one buffer per season) and updates only with the relevant buffer plus the newest data. This results in update times roughly 30 % lower than standard CL and far lower than TL‑based fine‑tuning, which must process the entire accumulated dataset each step (5–10× longer). Training from scratch is the most expensive.

• Update frequency – Monthly updates give the best accuracy but are computationally demanding. Quarterly updates strike a practical balance, delivering most of the accuracy gain with modest resource use. Annual updates are cheap but recover accuracy slowly after a drift.

• Scalability – In the 40‑house large‑scale experiment SML consistently outperforms all baselines, and its performance is robust to variations in building size, climate zone, and HVAC system. It quickly adapts after a retrofit, demonstrating its suitability for real‑world deployments where drifts are abrupt and unpredictable.

The authors conclude that, for long‑term operation of building thermal models on resource‑constrained hardware (e.g., edge controllers), a CL strategy that leverages seasonal memory is superior to both naïve TL re‑training and conventional CL methods. SML offers a practical recipe: maintain a modest replay buffer per season, update the model with the buffer corresponding to the current season together with the latest measurements, and perform updates on a quarterly schedule.

Contributions of the paper are: (1) the first systematic comparison of TL, CL, and scratch learning for building thermal dynamics under both feature drifts and concept drifts; (2) the introduction of Seasonal Memory Learning, a lightweight CL method that attains the highest prediction accuracy while keeping computational demands low; (3) quantitative analysis of how retrofits and occupancy changes affect model performance; and (4) practical guidelines for choosing update frequency and method based on accuracy‑cost trade‑offs.

Future work suggested includes validation on real‑world datasets, extension to multi‑zone buildings with multiple HVAC subsystems, and integration of the adaptive model into model‑predictive control or reinforcement‑learning based control loops.