Hybrid Autoencoder–CNN–BiLSTM–Attention Model for Smart Grid Cybersecurity Intrusion Detection

Abstract

The convergence of communication networks and smart electronic devices makes smart grid infrastructures vulnerable to cyber-attacks. Consequently, we need proper intrusion detection mechanisms to ensure secure and reliable grid operation. The present study proposes two advanced hybrid deep learning-based smart grid intrusion detection frameworks, namely optimized Autoencoder–CNN–BiLSTM–Random Forest and Denoising Autoencoder–CNN–BiLSTM–Attention-Random Forest model for the same. The proposed approach involves data preprocessing, scaling of features and latent feature extraction via autoencoder based representation learning. To capture the key spatial and temporal patterns, CNN (Convolutional Neural Networks) and BiLSTM networks are used to process on sequential latent features along with an attention mechanism that emphasizes significant temporal relations. A Random Forest classifier finally performs multiclass intrusion detection. Experimental results depict superior performance of the optimized AE–CNN–BiLSTM–RF model with 95.23% testing accuracy and the proposed DAE–CNN–BiLSTM–Attention–RF data-based approach, achieving accuracy at 99.24% with high precision, recall and F1-scores (close to 0.99) on baseline models such as LGBM (93%) and One-Class SVM (85%).