Ultra-short sequence-augmented power system transient stability assessment using transformer-based deep learning

Abstract

The increasing penetration of renewable energy into power grids has intensified concerns regarding transient stability. Traditional transient stability assessment methodologies, which rely on time domain simulations and energy function-based approaches, face significant limitations in computational efficiency and real-time applicability. To address these challenges, this paper proposes a novel dual-phase framework that integrates Long Short-Term Memory (LSTM) networks and a modified Transformer architecture for enhanced transient stability evaluation. In the first phase, an LSTM-based trajectory reconstruction module is employed to extrapolate complete 5-second dynamic trajectories from initial measurements of 0.5 s post-fault. This addresses the critical issue of temporal data insufficiency during the early stages of system perturbations, enabling accurate representation of electromechanical transients. The second phase employs a Transformer classifier, which processes the reconstructed trajectories to assess system stability. The Transformer’s architecture incorporates tailored positional encoding to align with the frequency and timescale characteristics of system dynamics, while its self-attention mechanism facilitates the extraction of global temporal patterns essential for stability classification. A case study on the IEEE 39-bus system demonstrates the efficacy of the proposed framework. Within the critical post-fault assessment window, the method achieves an accuracy of 99.6%, significantly outperforming conventional methods in both computational efficiency and classification accuracy.