Enhanced NILM Based on an Improved RGB Color-Encoded <i>V–I</i> Trajectory Representation

The existing color-encoded V-I trajectory methods for Non-Intrusive Load Monitoring (NILM) often suffer from limited physical interpretability and sparse information representation, which restricts their effectiveness in distinguishing between complex load types. To address these limitations, we propose an enhanced NILM framework that introduces an improved RGB color-encoded V-I trajectory image representation. This new visual representation fused the raw current values, per-pixel data point density, and current amplitude into the V-I trajectory through RGB channels. A pretrained ResNet18 model is then fine-tuned on these images to perform the load classification task. Experimental results on two widely used benchmark datasets, PLAID and WHITED, demonstrate that the proposed approach consistently outperforms existing state-of-the-art methods regarding accuracy and generalization. Moreover, the model maintains a favorable trade-off between performance and computational cost, highlighting its practicality for future NILM applications.