Air Conditioning Load Prediction Method Based on Cloud-Edge Collaborative Architecture

As a typical demand-side flexible resource, air conditioning load prediction plays a significant role in the safe and economic operation of power systems. With the increasing proportion of renewable energy and fluctuating loads, accurate prediction has become increasingly critical. However, existing methods face contradictions between model accuracy and edge-side resource constraints, while lacking lightweight deployment solutions. To address these challenges, this paper proposes an air conditioning load prediction method based on cloud-edge collaborative architecture. The method incorporates an improved bidirectional LSTM model with a multi-layer stacked structure and optimized feature extraction strategy to enhance prediction accuracy. By integrating a Maximum Mean Discrepancy (MMD) transfer learning mechanism, the approach enables effective knowledge transfer between different scenarios, solving the data scarcity problem for newly connected users. Furthermore, through ONNX conversion technology for model quantization and Docker container deployment strategies, a complete technical pathway from cloud-side training to edge-side inference is established. Experimental results demonstrate that the improved BiLSTM model substantially outperforms traditional models; the MMD-based transfer mechanism shows clear advantages over DANN models across all evaluation metrics; and the lightweight deployment solution achieves significant model size reduction and inference speed improvement. This method provides valuable reference for demand response decisions in power grid dispatch centers and contributes to renewable energy integration and carbon reduction goals.