A Transfer Learning Strategy for Short-term Wind Power Forecasting

Abstract

Newly-constructed wind farms often lack collections of historical wind power data and it is changeling to forecast their future wind power accurately. A novel transfer learning strategy for short-term wind power forecasting is proposed to tackle this issue in this paper. To more accurately forecast the wind power output of the newly constructed wind farm, a nearest-neighbors approach is employed to select highly relevant historical data from other wind farms. Thus, the wind power dataset of the target wind farm is significantly enriched and it allows to better train the forecasting models. A hybrid Jaya Extreme Gradient Boosting (Jaya-XGBoost) algorithm is employed to generate the forecasting results and wind power data collected in China is utilized. The Jaya-XGBoost algorithm is compared with Support Vector Machines (SVM), and Least Absolute Shrinkage and Selection Operator(LASSO),Neural Networks in wind power forecasting with different time horizons. Computational results demonstrate that the forecasting results of the four algorithms are all improved by leveraging information from other wind farms while the Jaya-XGBoost algorithm yields the best results over the four algorithms.