The transition toward smart grids is accelerating, driven by the growing deployment of advanced metering infrastructure and smart meters that provide detailed, building-level energy data. However, a persistent challenge remains: many buildings, both residential and commercial, lack sufficient historical load data, limiting the performance of short-term load forecasting models that are essential for efficient energy management.
A new study conducted by the National Technical University of Athens (NTUA) within the framework of the CRETE VALLEY project, titled Transfer learning techniques on temporal fusion transformers for short-term building load forecasting under limited data conditions, addresses this gap by leveraging transfer learning (TL) to improve forecasting accuracy in data-scarce environments.
The research tests several forecasting models, including the advanced Temporal Fusion Transformer (TFT), by first training them on data-rich buildings and then adapting them to buildings with limited data. This approach allows models to transfer knowledge from one setting to another.
Results based on real-world datasets from student dormitories across three locations show that transfer learning consistently improves forecasting accuracy, especially in data-scarce conditions. In such cases, forecasting errors were reduced by 10–20% compared to standard models.
The Temporal Fusion Transformer demonstrated the strongest performance, achieving around a 20% reduction in error and significant improvements in prediction accuracy compared to other approaches.
These findings highlight the potential of combining transfer learning with advanced AI models to improve energy forecasting, supporting more efficient energy management and smart grid operations even when data are limited.
—
Photo by Isaac Wolff on Unsplash
