DEEP LEARNING FOR POWER SYSTEM STABILITY

Abstract

Power system stability has emerged as a critical challenge in modern smart grids owing to increasing penetration of renewable energy sources, complex load dynamics, and the growing scale of interconnected networks. Traditional methods such as time-domain simulation, continuation power flow, and energy function analysis, while theoretically sound, suffer from computational bottlenecks that preclude real-time deployment. This paper proposes a novel hybrid deep learning framework that combines Long Short-Term Memory (LSTM) networks with one-dimensional Convolutional Neural Networks (1D-CNN) to assess voltage stability in real time. The model simultaneously captures temporal correlations in power system measurements through LSTM layers and extracts local feature patterns through convolutional filters. Trained on simulated datasets derived from IEEE 39-bus, 118-bus, and 300-bus test systems, as well as real SCADA measurements from the Northern Indian grid, the proposed LSTM-CNN achieves a classification accuracy of 98.8%, an AUC-ROC of 0.9891, and an average inference time of 4.9 ms per sample. These results represent significant improvements over Support Vector Machines (SVM), Random Forests, Deep Neural Networks (DNN), and stand-alone LSTM baselines. Ablation studies, sensitivity analyses, and feature importance rankings confirm the robustness and interpretability of the proposed framework. The findings strongly support the deployment of deep learning-based tools in energy management systems (EMS) for automated, near-instant stability monitoring.