Research on AI-based anomaly detection methods for big energy consumption data in smart cities

Artificial intelligence algorithms demonstrate significant advantages in processing big data on energy consumption in smart cities, exhibiting strong adaptability particularly in anomaly detection tasks involving high-frequency, heterogeneous, and multimodal data scenarios. To enhance the accuracy and response efficiency of energy consumption anomaly identification, this study constructs a multi-source detection model integrating the Transformer architecture with a gated convolutional mechanism. A comprehensive methodological framework encompassing sliding-window preprocessing and dynamic label generation was designed. The model employs a spatio-temporal joint tensor modeling strategy to effectively resolve structural inconsistencies among electricity consumption, ambient temperature, and thermal imaging data, establishing a unified input channel for cross-source consistent representations. Multiple comparative experiments using real-world data from core smart urban districts demonstrate that this model significantly outperforms existing CNN-LSTM, ST-GCN, and autoencoder approaches in metrics such as precision, recall, and F1-score, while exhibiting robust transferability and efficient inference capabilities. This approach not only enhances the accuracy of energy consumption anomaly detection but also demonstrates robust engineering deployment capabilities, providing a theoretical foundation and methodological support for building interpretable, highly responsive smart city energy consumption monitoring systems.