Power Forecasting of Regional Wind Farms via Variational Auto-Encoder and Deep Hybrid Transfer Learning

Abstract

Wind power is a sustainable green energy source. Power forecasting via deep learning is essential due to diverse wind behavior and uncertainty in geological and climatic conditions. However, the volatile, nonlinear and intermittent behavior of wind makes it difficult to design reliable forecasting models. This paper introduces a new approach using variational auto-encoding and hybrid transfer learning to forecast wind power for large-scale regional windfarms. Transfer learning is applied to windfarm data collections to boost model training. However, multiregional windfarms consist of different wind and weather conditions, which makes it difficult to apply transfer learning. Therefore, we propose a hybrid transfer learning method consisting of two feature spaces; the first was obtained from an already trained model, while the second, small feature set was obtained from a current windfarm for retraining. Finally, the hybrid transferred neural networks were fine-tuned for different windfarms to achieve precise power forecasting. A comparison with other state-of-the-art approaches revealed that the proposed method outperforms previous techniques, achieving a lower mean absolute error (MAE), i.e., between 0.010 to 0.044, and a lowest root mean square error (RMSE), i.e., between 0.085 to 0.159. The normalized MAE and RMSE was 0.020, and the accuracy losses were less than 5%. The overall performance showed that the proposed hybrid model offers maximum wind power forecasting accuracy with minimal error.