Feature Dimensionality Reduction Based on Deep Lasso for Wind Power Forecasting

Abstract

Wind power forecasting considering spatio-temporal correlations can effectively improve the forecasting accuracy. However, this will lead to a complicated structure in the forecasting model, making it difficult to solve due to dimensional catastrophe. To this end, a neural network framework called Deep Lasso is applied, which achieves feature selection by adding the regularisation of Lasso to the input gradients. Primarily, a forecasting model based on Deep Lasso, considering the features of all wind farms (i.e., global variables), is constructed. Subsequently, the coefficients of Deep Lasso can directly represent the contribution of input features to wind power forecasts. Therefore, to construct a more efficient forecasting model, secondary modelling is performed by filtering the features with small coefficients. Experiments including 20 wind farms demonstrate that Deep Lasso exhibits remarkable suitability and effectiveness in ultra-short-term wind power forecasting compared with six feature selection methods. Moreover, to test the effectiveness of feature dimensionality reduction, the secondary modelling forecasting model is verified by comparing it with principal component analysis (PCA) and factor analysis (FA). The results obtained show that the overall performance of the proposed method outperforms PCA and FA while improving the computational efficiency to a certain extent.