A Unified Lightweight Edge Pattern Recognition Framework for Heterogeneous Streaming Power Data

With the continuous development of smart distribution networks and electricity consumption information collection systems, data such as voltage, current, and power are being continuously generated through real-time acquisition, forming streaming power data characterized by multi-source heterogeneity, real-time arrival, and continuous evolution. Most existing relevant studies focus on single tasks and generally suffer from problems including insufficient cross-modal generalization, large overhead for edge deployment, and inadequate coordination between privacy preservation and real-time model updating. To address these limitations, this paper proposes a unified lightweight edge pattern recognition framework for heterogeneous streaming power data. The framework constructs a unified technical pipeline centered on unified preprocessing, modality-adaptive feature extraction, shared feature mapping, and task-adaptive output, and integrates lightweight recognition, knowledge distillation, and federated learning to achieve efficient modeling, collaborative optimization of privacy protection, and continuous updating for various types of streaming power data in resource-constrained edge environments. For different input modalities, this paper designs colored V-I trajectory enhancement and lightweight siamese recognition, CNN-Transformer power time-series modeling, as well as improved APAA and DDTW-OPTICS clustering for voltage sequences, all of which are incorporated into the unified framework via a shared representation mechanism. Experimental results demonstrate that the proposed method achieves favorable recognition performance and deployment capability on the REDD, UK-DALE, WHITED datasets and in embedded real-world scenarios. Compared with the teacher model, the student model reduces the number of parameters by 90.0% and model size by 88.9%. The federated collaborative communication traffic is reduced to 6.56 MB, and the single inference latency on the STM32MP1 platform is approximately 0.2 s. The results show that the proposed framework achieves a good balance among recognition accuracy, model lightweightness, privacy protection, and edge real-time performance, and can provide a unified and practical technical solution for multiscenario edge intelligent analysis of streaming power data. In the revised framework, the shared representation is further constrained by a unified feature-space objective, cross-task shared parameters, and modality-branch collaborative training, and extended ablation, streaming robustness, and sensitivity analyses are added to verify that the method is a unified framework rather than a loose combination of independent models.