SLAFF: Shapley Local Attribution Feature Fusion with Bayesian-Optimized Ensemble Learning for Detecting Intrusions in Renewable Energy Power Systems

The incorporation of the internet into modern power systems has improved real-time communication, automation, and regulation but also heightened the risk of cyberattacks. Due to the continual generation of massive amounts of high-dimensional and heterogeneous data by these systems, it is difficult to differentiate between the many types of cyberattacks. Thus, this paper proposes a novel Shapley Local Attribution Feature Fusion (SLAFF) algorithm to select the subset of features based on Shapley values and the local attribution values. Initially, the training data are balanced using the oversampling technique. The SLAFF algorithm is implemented on the balanced dataset using only the training set to avoid data leakage. The different feature subsets are formed and trained using machine learning methods with the balanced and unbalanced datasets. The ensemble model is developed using the different ML algorithms and trained using the Bayesian optimization method. The highest accuracy is achieved by the ensemble model with 95.87% accuracy on the Kaggle Power System Intrusion dataset on 25 feature subsets and 95.36% accuracy on the Figshare Power System dataset on 40 feature subsets. Thus, implementing the proposed method will find discriminative features while preserving classification performance and offering an SLAFF method for power-system intrusion detection.