Advancing Smarter Building Energy Forecasting with AI, by Felix Lau Pangestu

EDPS is pleased to recognise Felix Lau Pangestu, a final-year Computer Science student at Hong Kong Baptist University, as one of the joint highest-scoring awardees of the EDPS Innovation Scholarship 2025/26.

Felix’s project, Few-Shot Transfer Learning for Building Energy Forecasting Using LSTM Neural Networks, addresses a practical sustainability challenge. Newly commissioned, renovated or recently instrumented buildings often have only a limited amount of reliable energy data. This makes it difficult to train accurate deep-learning models for energy forecasting.

His research explores whether transfer learning can overcome this limitation. Rather than building every forecasting model from the beginning, transfer learning allows a model to reuse knowledge gained from buildings with richer historical data. Felix compared four approaches: training from scratch, fully fine-tuning a pre-trained model, freezing the model’s main backbone, and adding compact adapter layers.

The project evaluated these methods across different building types and data periods. For well-matched education buildings, transfer learning reduced forecasting error by up to 29% when eight weeks of target-building data were available. In another case, it delivered a 67.5% improvement with only one week of data. However, the research also found that full fine-tuning could perform poorly when the source and target buildings had very different energy-use patterns.

This discovery became an important part of the project. Felix found that freezing the model’s backbone provided a more reliable option for difficult targets. He also proposed an automated switching rule that compares scratch and transfer models using validation data before selecting the stronger approach. In the experiments, this rule achieved 99.9% of the performance of an ideal selection made with hindsight.

The project stands out because it examines not only when AI performs well, but also when and why it may fail. Its findings could help building owners, facilities managers and smart-building technology providers introduce energy forecasting earlier, without waiting years to collect sufficient data.

Through this project, Felix demonstrates how practical and carefully tested AI can support smarter building operations, improved energy efficiency and a more sustainable built environment.