DEEPGRIDSHIELD: A DEEP LEARNING FRAMEWORK FOR REAL TIME ANOMALY DETECTION IN POWER DISTRIBUTION NETWORKS

Abstract

Recent studies indicate that cyber-physical systems, especially those involving vehicular networks 
and industrial control systems, experience up to 35% of security breaches due to misclassified 
anomalies. Additionally, over 28% of attacks in power distribution and transport networks go 
undetected due to poor feature representation. Traditional classifiers, when used in isolation, 
demonstrate an average detection accuracy of only 85–88% in complex, multi-class intrusion 
scenarios. Existing manual analysis methods suffer from high dependence on rule-based signatures, 
which are often ineffective against novel or evolving attacks. They also require continuous human 
supervision, making real-time implementation impractical, and tend to misclassify low-frequency but 
critical anomaly patterns due to lack of robust learning mechanisms. To address these challenges, this 
work proposes a hybrid classification framework that integrates a Multilayer Perceptron (MLP)-based 
Deep Neural Network (DNN) with a Random Forest Classifier (RFC). The methodology leverages the 
MLP’s ability to learn high-dimensional, non-linear feature representations through dense layers, 
particularly extracting features from an intermediate latent layer termed the "feature layer." These 
features are then used as input for the RFC, which excels at ensemble-based decision-making and 
effectively mitigates overfitting by combining predictions from multiple decision trees. The complete 
model pipeline is optimized using early stopping and persistence strategies to enhance generalization 
and reusability. This hybrid approach not only improves classification accuracy and generalization but 
also ensures robustness across varied and imbalanced datasets, making it highly suitable for real-time 
anomaly detection in intelligent systems.