REAL-TIME EDGE AI FOR SECURE SMART GRID IOT–CLOUD ECOSYSTEMS: AN INTELLIGENT EDGE COMPUTING FRAMEWORK FOR CYBERATTACK DETECTION, PRIVACY-PRESERVING ENERGY ANALYTICS AND ADAPTIVE THREAT RESPONSE

Abstract

The rapid digitalization of smart grids through Internet of Things (IoT) devices, advanced metering infrastructure, distributed energy resources, and cloud-based energy management has improved operational efficiency while simultaneously expanding the cyberattack surface. Conventional cloud-centered security mechanisms often experience communication delays, privacy risks, bandwidth limitations, and inadequate responsiveness to rapidly evolving threats. This study presents a real-time Edge Artificial Intelligence framework for securing smart-grid IoT–cloud ecosystems through distributed cyberattack detection, privacy-preserving energy analytics, and adaptive threat response. A multi-source dataset containing 285,000 operational and cybersecurity records was constructed from smart meters, phasor measurement units, substation sensors, network traffic, authentication logs, and energy-consumption profiles. The dataset incorporated normal activity and six major attack categories: false-data injection, denial-of-service, energy theft, command injection, device spoofing, and malware-based intrusion. Following data cleaning, temporal synchronization, normalization, class balancing, and feature selection, 48 predictive variables were retained. The proposed architecture combines a lightweight convolutional neural network, bidirectional long short-term memory layers, and an attention mechanism to capture spatial, temporal, and context-sensitive attack patterns at the network edge. Federated learning and differential privacy were incorporated to support collaborative model training without transmitting raw consumer or grid-operational data to the cloud. Experimental evaluation used stratified training, validation, and testing partitions of 70%, 15%, and 15%, respectively. The proposed model achieved 98.1% accuracy, 97.8% precision, 97.6% recall, a 97.7% F1-score, and a receiver operating characteristic area under the curve of 0.993. It outperformed logistic regression, support vector machine, random forest, and conventional CNN models, which achieved accuracies of 84.7%, 89.6%, 93.8%, and 95.4%, respectively. Edge-based inference reduced average threat-detection latency from 186 milliseconds in the cloud-only configuration to 42 milliseconds, representing a 77.4% reduction, while decreasing cloud-bound data transmission by 68.2%. The adaptive response mechanism isolated compromised devices and prioritized security alerts, reducing successful attack propagation by 71.6%. Differentially private federated learning preserved sensitive energy data with only a 1.2-percentage-point reduction in detection accuracy. These findings demonstrate that the framework provides accurate, low-latency, privacy-aware, and scalable cybersecurity for next-generation smart-grid infrastructure.