Multi-source deep transfer learning with stacked denoising autoencoder and Wasserstein distance for wind power prediction in new wind farm

High-precision wind power forecasting is essential for ensuring the secure and stable integration of wind energy into power grids. However, newly commissioned wind farms typically face data scarcity due to their limited operational history, making accurate power output prediction particularly challenging. To overcome the challenge of insufficient historical data in newly established wind farms, this paper proposes a transfer learning-based deep neural network for high-precision wind power point forecasting, integrating multi-source data assimilation. Firstly, a stacked denoising autoencoder is used to establish feature correlations between the source and target domain input data. Then, the parameters of a well-trained long short-term memory (LSTM) prediction model from a source wind farm are transferred to the target farm’s prediction model. Finally, multiple prediction models are integrated by calculating the Wasserstein distance between each source domain and the target domain to form the final wind power forecasting model. Experimental results demonstrate that the proposed transfer model outperforms other comparison models in prediction accuracy, offering strong adaptability and broad applicability for wind power forecasting in newly established wind farms.