A novel approach to ultra-short-term wind power prediction based on feature engineering and informer

Abstract

Wind power is prone to dramatic fluctuations in the short term, posing a threat to the safety and stability of the grid, so accurate forecasting of ultra-short-term wind power is important to ensure the stability and economy of the power system. The historical data of wind power is an enormous and nonlinear time series. It is expected to mine the independent features related to wind power from the original data through feature engineering, and then use the Informer model to solve the prediction problem of long-time series of wind power, thus reducing the space complexity and improving the prediction accuracy. In this paper, Turkey wind farm data are selected to predict ultra-short-term wind power of 10 min based on feature engineering and the Informer model. First, factor data with a high correlation with wind power is formed after feature engineering. Then, train the Informer model and conduct multiple experiments to obtain the optimal parameters. Finally, the prediction results are compared with the recurrent neural network (RNN) model, the long-short-term memory (LSTM) model, and the Transformer model. The experimental results show that the Informer model has high prediction accuracy and operation efficiency. Four evaluation metrics of mean absolute error, mean square error, symmetric mean absolute percentage error, and runtime decreased by at least 32.849, 8495.193, 5.544%, and 92, which proves the approach based on feature engineering and Informer has prominent advantages in ultra-short-term wind power prediction. Its prediction results can provide a reference for the coordinated dispatching, risk analysis, and scientific decision-making of wind power systems.

Keywords

• Long time series
• Feature engineering
• Informer model
• Ultra-short-term wind power
• Prediction