A Transfer Learning-Based Prior VGG Model for Sub-Synchronous Oscillation Recognition in Wind Turbines

Abstract

: Sub-synchronous oscillation (SSO) is a critical stability threat in wind farms connected to power grids through series-compensated transmission lines, where delayed or inaccurate recognition may lead to converter overcurrent, turbine disconnection, shaft torsional vibration, and large-scale power fluctuations. Existing model-based approaches depend heavily on complete system parameters, while conventional signal-processing methods often require long observation windows and are less suitable for rapid online warning. To address these limitations, this paper proposes a physics-aware data-to-image recognition framework and a transfer-learning-based prior VGG model for identifying SSO hazard levels in series-compensated doubly fed induction generator wind farms. First, multi-source power-quality variables, including voltage, current, frequency, active and reactive power, voltage total harmonic distortion, current total harmonic distortion, and rotor-speed-related information, are organized into a two-dimensional feature matrix according to the actual topology of the DFIG grid-connected system. This matrix preserves the electrical relationships defined by the DFIG grid structure and effectively enhances discriminative feature representation by aligning heterogeneous measurements with grid-side and wind-farm-cluster connectivity, enabling more structured feature extraction for subsequent learning models. The resulting matrix is then normalized and mapped into RGB images, allowing neural networks to learn both local physical consistency and cross-cluster spatial coupling patterns. Second, feedforward neural network, capsule neural network, and VGG-style convolutional models are developed and compared for three SSO states: attenuating oscillation, constant-amplitude oscillation, and diverging oscillation. To alleviate limited SSO samples and deep network convergence difficulty, simplified prior SSO images are generated from physically interpretable class prototypes and used to pre-train the VGG feature extractor before transfer to simulated SSO images. The proposed framework achieves strong recognition performance under the random sample-level protocol and shows the best independent case-level performance among the compared methods, providing a useful reference for SSO analysis of wind turbines.