Thursday 13 March 2025
A new approach to predicting building energy consumption has been developed, using a technique called transfer learning. This method allows researchers to fine-tune a general model on a small amount of data from a target building, rather than starting from scratch.
Traditionally, building energy simulations rely on complex models that require vast amounts of data and computing power. However, this approach can be time-consuming and expensive. The new technique, developed by researchers at the Technical University of Munich, uses a general model trained on data from multiple buildings to predict energy consumption in a target building.
The team used a dataset of 450 simulated buildings with different characteristics, such as insulation properties, size, and location. They then fine-tuned this general model on small amounts of data from individual buildings, achieving significant improvements in prediction accuracy.
One of the key advantages of transfer learning is its ability to reduce the amount of data required for accurate predictions. In traditional methods, a large dataset is needed to train a building-specific model, which can be challenging and expensive to obtain. With transfer learning, a general model can be fine-tuned on just 10-30 days’ worth of data from a target building.
The researchers tested their approach on 144 buildings with varying characteristics and found that the general model outperformed traditional single-source models in terms of prediction accuracy and consistency. The results showed an average reduction of 42% in mean absolute scaled error (MASE) compared to fine-tuning single-source models.
This new approach has significant implications for building energy management, particularly in the context of smart buildings. By reducing the data requirements and improving prediction accuracy, transfer learning can enable more efficient and effective energy consumption control and fault detection.
The team’s findings suggest that a general model trained on multiple buildings can serve as a universal source for fine-tuning, eliminating the need for source-building selection. This approach could be particularly useful in scenarios where data is limited or unreliable, such as in retrofitting existing buildings or predicting energy consumption in new constructions.
While there are still challenges to overcome before widespread adoption, the potential benefits of transfer learning in building energy simulations are significant. As researchers continue to refine and develop this technique, it may play a crucial role in shaping the future of sustainable building design and management.
Cite this article: “Predicting Building Energy Consumption with Transfer Learning”, The Science Archive, 2025.
Building Energy Consumption, Transfer Learning, Machine Learning, Building Simulation, Energy Prediction, Smart Buildings, Data Requirements, Fault Detection, Energy Management, Universal Source
