A Novel Image Classification Approach via MobileNet Model for Security Assessment of Power System

Abstract

The computational complexities of conventional techniques used for security assessment of power system are addressed by proposing a maiden deep learning-based technique capable of boosting security calculations, increasing decision-making efficiency, and reducing hazards for real time operation. This approach uses MobileNet Convolutional Neural Network (MCNN) to evaluate security with reduced model size and computational burden as compared to traditional CNN architectures by leveraging special type of convolution technique. MCNN's automatic feature extraction and strong generalization allow the operator to determine whether the current operating state of power system is secured or not. All of the system’s critical operational scenarios are developed via load flow studies for generation of data while accounting for small signal stability with ℓ − 1 security and converted into RGB images of varied input dimensions for training and testing of the MCNN. Once trained, the MCNN is well-suited for real-time applications like assessing whether current state of the system is secured or not including unforeseen occurrences like topological changes in a computation-free manner. The comparative performance of the proposed method with customized CNN on three typical IEEE test system demonstrates its superior performance in terms of better accuracy and computational efficiency.