An adaptive wind power forecasting method based on wind speed-power trend enhancement and ensemble learning strategy

Abstract

Accurate wind power forecasting (WPF) is essential for power system planning, operation, and management. However, the high uncertainty and stochastic behavior of natural wind brings great challenges to high performance WPF. In this context, an adaptive WPF model based on wind speed-power trend enhancement and an ensemble learning strategy is proposed in this study. For wind speed-power trend enhancement, abnormal data are detected and removed by the combined local outlier factor algorithm and quartile method. The artificial power data are interpolated using a neural network based on the normal wind speed-power distribution. In the ensemble learning strategy, a total of eight individual learners are involved as the candidate base learners. The principle of selecting base learners with low correlation and high accuracy is provided to achieve high performance forecasting, and thus, four base learners with different internal mechanisms are finally selected. The partial least squares regression is utilized for outputs weighting, and the K-fold cross-validation is adopted for dataset division. Collected data from a real wind turbine system are used for performance investigation. Forecasting tests with three time horizons (10, 30, and 60 min) and three seasons (Spring, Summer, and Autumn) are carried out. The results reveal that the proposed model is more accurate and adaptive compared with individual learners and other ensemble models.