Research on Lightweight Edge Intelligent Service User Portrait Construction and Inference for Precision Power Marketing

Against the background of the “dual carbon” goals and new power system construction, power marketing is shifting toward intelligent services like demand identification and anomaly early warning. However, current user portrait methods rely on static features, failing to capture dynamic electricity use patterns, while cloud-centric inference suffers from high latency and large data volumes. This paper proposes a lightweight edge intelligent portrait construction and inference method for precision power marketing. First, multi-source data are integrated to build a multi-dimensional feature system. Second, a lightweight feature enhancement method converts 1D time-series signals into low-dimensional sparse trajectory matrices via color encoding and binary mapping. A GA-CNN-SENet model is then constructed, with ResNet18-SqueezeNext knowledge distillation enabling edge-side lightweight inference. Finally, a cloud-edge collaboration architecture using dynamic PAA compression and weighted DTW-DBSCAN supports low-confidence sample backhaul and model iteration. Experiments show the method achieves 92.84% portrait accuracy, 90.73% anomaly F1-score, and 31.6 ms edge latency, reducing daily data upload for 8000 users from 51.20 MB to 6.56 MB. It cuts edge computing and bandwidth costs while maintaining accuracy, supporting customer segmentation, personalized services, anomaly warning, and demand response in precision power marketing.