Intelligent harmonic diagnosis and tracing method for urban power grids via cross-modal fusion of time-frequency features and unstructured equipment records

Against the backdrop of the accelerated construction of new electricity systems and the continuous increase in the proportion of power electronicization, urban power grid harmonics exhibit characteristics such as multi-source, time-varying, and wide-area propagation. A large amount of semi-structured and unstructured data (ledgers, governance records, waveforms, images, etc.) has been accumulated during operation and maintenance. However, existing power knowledge graphs primarily rely on single-modal textual data (equipment ledgers and fault records) and cannot effectively integrate waveform information into the knowledge representation framework, resulting in fragmented knowledge, semantic inconsistency, and difficulty in cross-modal association. To address these problems, this article proposes a method for constructing a knowledge graph of urban power grid harmonic analysis based on multimodal feature fusion. A top-down construction approach is adopted, forming a three-part framework to support topological correlation analysis, node harmonic level characterization, and historical governance scheme retrieval. For multimodal knowledge extraction, a bidirectional encoder representations from Transformers (BERT)–bidirectional long short-term memory (BiLSTM)–conditional random field (CRF) model is employed on the text side for entity and relationship extraction from unstructured equipment ledgers and governance records. On the waveform side, the short-time Fourier transform is used to generate time-frequency spectrograms, followed by convolutional neural network (CNN)-based harmonic identification. Cross-modal knowledge alignment is achieved through cosine similarity computation between text entity vectors and image feature vectors. Experimental results demonstrate that the BERT–BiLSTM–CRF model achieves 87.5% precision, 88.1% recall, and 87.8% F1-score in text knowledge extraction. The CNN-based harmonic identification achieves 93% test accuracy, surpassing other methods. Threshold sensitivity analysis confirms that 0.85 provides optimal cross-modal alignment with a 91.2% correct association rate and only a 3.8% false association rate. Finally, the constructed knowledge graph, stored and visualized in Neo4j, enables intelligent harmonic problem localization and governance decision support, demonstrating significant practical value for urban power grid harmonic management.