A meta energy model-based transfer learning methodology in cooling load predictions: Office building case study

Building energy consumption constitutes over 40% of the total primary energy consumption, and buildings play an essential role in energy efficiency and sustainability. To unlock their potential, an accurate building energy model is crucial for optimal building controls. However, developing such a model remains challenging due to the high complexity and computational cost of physics-based models, as well as limited generalization and poor interpretability of purely data-driven models in contexts of insufficient data. To address these challenges, this study proposes a novel transfer learning (TL) methodology based on meta energy model (MEM) to enhance predictive accuracy and model generalizability under data scarcity. The study tests the methodology of using 31 office buildings from the open-source Building Data Genome Project 2 dataset. First, a million-scale synthetic dataset is generated by EnergyPlus simulations covering diverse climate zones and building design conditions to develop the MEM. Second, the MEM produces source data for target buildings using their building design and meteorological parameters to implement TL. Finally, two TL strategies, weight initialization (WI) and feature extraction (FE), are systematically evaluated across five models: lightGBM, lightGBM-TL, LSTM, LSTM-WI, and LSTM-FE. Results demonstrate that effective TL strategies (e.g., WI) enabled by the MEM framework can improve prediction accuracy by 20%–40% compared to the standalone lightGBM and LSTM models, without requiring real historical source data. The insights obtained can help the building industry fully utilize basic building information to develop reliable surrogate models in real building applications with or without historical data.