| Wind Turbine Power Classification using Adaptive Transfer Learning | |
|
MSRDG International Journal of Computer Scientific Technology & Electronics Engineering
© 2025 by MSRDG IJCSTEE Journal Volume 1 Issue 3
Year of Publication: 2025 |
Paper Download Article ID MSRDG-IJCSTEE-V1I3P103 |
|
Abstract: Accurate classification of wind turbine output power is essential for optimising grid integration, predictive maintenance scheduling, and energy forecasting. Conventional machine learning methods frequently suffer from domain shift when deployed across turbines manufactured by different vendors or operating under varying climatic regimes. This paper introduces an Adaptive Transfer Learning (ATL) framework that combines a convolutional feature extractor pre-trained on large-scale industrial vibration and power-curve datasets with a domain alignment mechanism based on Maximum Mean Discrepancy (MMD) loss and an attention-gated classifier head. The framework employs layer-wise learning-rate decay to selectively fine-tune deeper representations while preserving low-level spectral features learned during pre-training. Experiments are conducted on three publicly available SCADA datasets—the ENGIE La Haute Borne open dataset, the NREL Western Wind dataset, and an in-house dataset collected from a 2 MW onshore turbine array—encompassing four operational classes: Low Power, Medium Power, High Power, and Fault. The proposed ATL model achieves an overall classification accuracy of 95.7%, an F1-score of 94.9%, and a precision of 95.2%, outperforming SVM, Random Forest, CNN-from-scratch, LSTM-from-scratch, and ResNet-50 fine-tuning baselines by margins of 7.3 to 17.3 percentage points in accuracy. SHAP-based explainability analysis confirms that wind speed, rotor RPM, and generator temperature are the dominant discriminative features. The results demonstrate that the ATL framework generalises robustly across turbine types without requiring complete retraining, offering a practical pathway for scalable condition monitoring in commercial wind farms. |
|
| Keywords: Wind turbine power classification · Adaptive transfer learning · Maximum Mean Discrepancy · SCADA data · Convolutional neural network · Domain adaptation · Condition monitoring | |
