A Lightweight Load Identification Method Based on Poincaré Mapping and Multivariate Feature Fusion Composition

With the improvement of demand-side resource regulation requirements of new power system, load identification as a key technology for rigorous load management has received more and more attention. However, the existing methods can further enrich the perspectives of periodic characterization of load electrical quantity, enhance the ability of multivariate feature fusion to improve the model performance, and reduce the complexity of the identification model to enhance the practicality. Therefore, a lightweight load identification method is proposed based on Poincaré mapping and multivariate feature fusion composition. Firstly, Poincaré mapping is used to characterize the periodic trajectory of load operation from the phase space perspective. The U-I trajectories of raw and reactive currents are constructed to portray the multidimensional dynamic characteristics. Multiple electrical expert features are embedded based on binary coding. Fuse the elements above to generate a color feature image with strongly distinguishable properties. MobileOne, the lightweight load identification model with strong feature learning capability, is then constructed. The computational complexity of the model is further reduced, and the inference latency is shortened by a reparameterization strategy to improve the availability of edge deployment significantly. Experiments are conducted in the public datasets PLAID and WHITED, and the proposed method achieves higher accuracy and F1-macro than the existing methods and reduces up to 99.23% of the number of parameters, 98.27% of the floating-point operations, and 89.15% of the inference latency compared with the existing load identification models based on image-based features. © 2024 Power System Technology Press. All rights reserved.