Design of a Convolutional Neural Network-Based Power System Stabilizer for Improved Small-Signal Stability

Abstract

The rapid integration of renewable energy sources, advanced power electronic interfaces, and distributed generation has significantly increased the nonlinear dynamics and structural complexity of modern power systems. These developments have intensified small-signal stability challenges, particularly low-frequency electromechanical oscillations that degrade power quality and system reliability. Conventional Power System Stabilizers (PSS), typically based on fixed-parameter lead–lag compensators, often lack adaptability and exhibit reduced damping performance under varying operating conditions. This paper proposes an Efficient Power System Stabilizer (EPSS) based on a Convolutional Neural Network (CNN) architecture to enhance oscillation damping in interconnected power networks. The CNN-based controller extracts spatiotemporal features from dynamic signals, including rotor speed deviation, terminal voltage, and power angle variations. Unlike conventional neural networks that treat inputs independently, the CNN effectively captures localized correlations and nonlinear interdependencies inherent in oscillatory power system behavior. A hybrid training strategy combining supervised learning with adaptive optimization algorithms is employed to ensure fast convergence and robust generalization across diverse disturbance scenarios. The model is trained using simulation data from a Single-Machine Infinite Bus (SMIB) system and validated on a two-area, four-machine benchmark system to assess scalability and robustness. Simulation results under symmetrical faults, load perturbations, and topology variations demonstrate superior damping performance compared to conventional lead–lag PSS and ANFIS-based controllers, achieving reduced overshoot, faster settling time, and improved stability margins. The proposed CNN-PSS also shows resilience to parameter uncertainties and measurement noise while maintaining real-time feasibility on embedded control platforms. This work demonstrates the effectiveness of deep learning–based stabilizers for enhancing small-signal stability in modern smart grids.