A VRF zonal thermal load prediction method based on transfer learning

In the context of global efforts to enhance building energy efficiency, variable refrigerant flow (VRF) systems are recognized for their high performance and flexible control, achieving widespread adoption, particularly in East Asia. This proliferation necessitates accurate thermal load prediction, which is essential for right-sizing systems and meeting performance guarantees. However, conventional whole-building forecasting fails at the outdoor unit (ODU) zone level because it overlooks the distinct zonal characteristics and the granular, user-driven operational dynamics that primarily govern the load. This paper addresses this critical gap by introducing a physics-guided transfer learning framework centered on a novel methodology for creating a high-fidelity, physics-based source domain. The methodology first empirically establishes the cooling-capacity-weighted indoor unit (IDU) activation ratio as a key determinant, and then develops a dynamic psychrometric blending method to integrate this metric into EnergyPlus. This physics-guided simulation approach enables the creation of a large-scale simulation database. Building on this foundation, a Long short-term memory (LSTM) network is pre-trained to learn general thermal principles, and a transfer learning strategy is then used to adapt this knowledge to data-scarce, real-world scenarios. The framework's efficacy was demonstrated through three distinct transfer strategies that systematically evaluated its performance using non-target data (for zero-shot prediction), limited target-specific data, and a hybrid of both. All strategies markedly outperformed a model pre-trained solely on simulation data, with the optimal hybrid strategy achieving a final R² of 0.866 and reducing the mean absolute error (MAE) by 18.9%. This approach demonstrates a promising pathway toward reliable prediction for ODU zones, offering valuable support for more efficient VRF system design and operation.