Non-Intrusive Load Identification Based on Timefrequency Image Coding

Non-Intrusive Load Monitoring (NILM) technology is central to achieving refined energy management in smart grids. However, existing methods often rely solely on single-domain temporal or frequency features, or suffer from parameter sensitivity, resulting in insufficient recognition accuracy and generalization capabilities. To address this, this paper proposes a novel load recognition approach that integrates multi-domain features with multi-stream convolutional neural networks. By synergistically mining static, dynamic, and spatiotemporal information, it significantly enhances the robustness of load recognition. Specifically, this method combines the Gramian Angular Difference Field to extract sequential static features and the Markov Transition Field to capture state transition dynamic features. By integrating frequency-domain information, it constructs high-distinctiveness multidimensional image features. During training, hierarchical clustering is introduced to preclassify multi-state appliance subtypes, significantly reducing recognition confusion caused by intra-class variations. The proposed multi-stream convolutional neural network method effectively extracts information from multi-class feature maps. Comparative experiments on the PLAID and LILACD public datasets demonstrate superiority over multiple existing methods, validating both recognition performance and generalization capability. Ablation studies confirm the necessity of the proposed architecture.