基于优化卷积长短期记忆模型的智能电网异常检测

The integration of digital technologies into traditional power systems has increased the efficiency and sustainability of power grids, transforming traditional grids into smart grids. However, this transformation has also introduced new vulnerabilities, such as susceptibility to false data injection (FDI) attacks, which can lead to significant energy theft. Recent reports estimate that these attacks cost utility providers approximately 101 billion dollars annually. This study presents an approach for anomaly detection in smart grids through energy consumption readings from smart meters on the customer side using an optimized lightweight convolutional long short-term memory (ConvLSTM) model. This study benchmarks and evaluates different machine learning models against seven FDI attacks, which are multi-class labeled. The evaluated machine learning models include traditional shallow detectors, deep learning-based detectors, and hybrid models that employ both horizontal and vertical detection strategies. Through extensive experimentation, the optimized ConvLSTM model is shown to demonstrate superior performance in detecting attacks; it achieves a high accuracy of 91.3% compared with other models in classifying these attacks. The results indicate that the proposed model provides a robust solution for improving the security and reliability of smart grids, and it offers significant benefits to utility providers who seek to mitigate energy theft and enhance grid resilience.