Mitigating Cyber Risks in Smart Cyber-Physical Power Systems Through Deep Learning and Hybrid Security Models

Abstract

The increasing integration of smart cyber physical power systems with communication networks has increased exposure to cyber threats, requiring advanced intrusion detection solutions. This work presents a deep learning based framework using the Cybersecurity Intrusion Simulated Network dataset and PSCAD generated cyber threat scenarios. Data preprocessing includes standardization, categorical encoding, and SMOTEENN sampling. Multiple models, namely Convolutional Neural Network, Long Short Term Memory, Transformer, and a hybrid CNN LSTM architecture, are trained with hyperparameter optimization. Experimental evaluation shows that the CNN LSTM hybrid achieves the highest performance, reaching 95.3% accuracy and 94.4% F1 score on the Cyber Threat dataset, and 99.9% accuracy with 99.9% F1 score on PSCAD simulations. Explainable AI techniques, including LIME and SHAP, are integrated to interpret feature contributions and enhance trust. For practical deployment, the optimized model is implemented using a Flask based web application enabling real time monitoring. The system classifies grid traffic into no attack, attack detected, injection attack, MITM, replay attack, and spoofing attack categories. The proposed approach delivers accurate, interpretable, and scalable intrusion detection for resilient smart grid cybersecurity operations worldwide deployments.