A High-Dimensional Heterogeneous Data Attack Detection Method for Power Network Based on Machine Learning Technology

In the digital transformation of smart grids, attack detection of high-dimensional heterogeneous data faces challenges such as curse of dimensionality, inefficient modal fusion, and small sample class imbalance. This article proposes a machine learning based high-dimensional heterogeneous data attack detection method for power networks (BHG-AD). By constructing a block based distributed framework to achieve localized processing of regional traffic characteristics, a hierarchical dimensionality reduction strategy combining principal component analysis and BERT encoding is adopted to solve the fusion problem of structured and unstructured data. A dual graph convolutional neural network is introduced to capture topological structure anomalies, and a Wasserstein distance based generative adversarial network is improved to alleviate the small sample problem. Finally, a dynamic attention mechanism is used to integrate temporal, topological, and reconstruction features to achieve multimodal collaborative decision-making. The experiment was validated on the IEEE123 node distribution network and CIC-IDS2017 dataset, and the detection accuracy of this method reached 98.7 %, an improvement of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 5. 2 \%}$</tex> compared to a single SVM model. The false alarm rate was reduced to 2.3 %, and the edge device inference delay was 12.5 ms, meeting the real-time requirements of the power grid. The research results provide a technical framework for smart grid security protection that balances detection accuracy, real-time performance, and scalability. Its cross modal feature engineering and distributed collaboration mechanism provide a reference paradigm for solving similar problems in industrial IoT.