Non-Intrusive Load Monitoring Based on Multi-Feature Fusion and Combinatorial Optimization Networks

Abstract

To address the limitations of traditional Voltage-Current (VI) trajectory features in appliance load identification—such as the difficulty in distinguishing similar appliances, weakened amplitude information, and the absence of dynamic characteristics—this paper proposes a dual-stage cyclic training method for load identification that integrates multi-feature reconstruction with Particle Swarm Optimization (PSO). First, to overcome the high similarity of original VI trajectories, a PSO-based threshold optimization algorithm is designed to reconstruct VI trajectories through reflection operations and normalization, thereby enhancing the geometric morphological differences among similar appliances. Second, to supplement dynamic impedance information and energy level features, conductance-time trajectories and mean-square current color-block maps are introduced to characterize dynamic impedance variations and energy level differences, respectively. Finally, a three-channel classification network based on ResNet18 is constructed, where the reconstructed VI trajectories, conductance-time trajectories, and mean-square current color-block maps are fused via RGB channels as inputs, forming a closed-loop “threshold optimization—feature reconstruction—cyclic training” framework. Experimental results on the PLAID dataset demonstrate that the proposed method achieves an identification accuracy of 98.29% and a macro-averaged F1-score of 97.93%. Comparative experiments verify the effectiveness of the reconstructed VI trajectories, the complementarity of multi-feature fusion, and the superiority of the combinatorial optimization network, significantly improving the identification of multi-state and similar-condition appliances.