A Spatiotemporal Information Extraction-Based FDI Attack Detection and Data Recovery Architecture for Wind Power Prediction Systems

Abstract

With the increasing penetration of wind energy in power systems, forecasting wind power has emerged as an effective method to enhance system stability. However, the accuracy of wind power prediction is significantly influenced by the quality of data collected by the wind farm’s communication system, which is vulnerable to cyberattacks. The inherent strong volatility of wind energy provides natural cover for false data injection (FDI) attacks, making their detection in wind power prediction systems a very challenging problem. This paper brings new insights into FDI attack detection and data recovery through spatiotemporal information extraction for wind power prediction systems. First, wind power time series data from multiple wind turbines are decomposed into trend, seasonal, and residual terms to provide additional spatiotemporal information for feature extraction. Subsequently, a novel extended long short-term memory autoencoder framework, incorporating $k$ -fold cross-validation and $z$ -score-based dual-threshold strategies, is developed to reconstruct these decomposition terms for accurately and iteratively identifying FDI attacks. Further, the compromised data is recovered using a spatiotemporal graph neural network with a dynamic trust weight mechanism. Experimental results demonstrate that the proposed method can achieve more accurate performance than existing methods.