ADAPTIVE WIND POWER FORECASTING BASED ON ONLINE SELECTIVE ENSEMBLE JUST-IN-TIME LEARNING

Wind power forecasting can provide effective guidance information for grid connection and optimal scheduling of wind power, and plays an important role in the development and utilization of wind energy. However,accurate wind power forecasting often encounters great challenges due to the inherent intermittency and randomness of wind power. Moreover,the characteristics of wind power data changes over time due to the factors such as seasonality,climate and equipment aging,which causes performance degradation of offline wind power forecasting models. To address these issues,an adaptive wind power forecasting method based on online selective ensemble just-in-time learning(OSEJIT)is proposed. Firstly,we construct a JIT base model library,incorporating similarity and learner perturbation techniques to effectively handle wind power’s nonlinearity and time-varying behavior,ensuring reliable forecasting. Secondly,we establish metrics for ensemble effectiveness,utilizing the Friedman test for diversity and prediction accuracy for model selection during online prediction. Subsequently,the final prediction is obtained through adaptive weighted ensemble based on the recent prediction performance of the individual models. To update the base model library while minimizing frequent model reconstruction and resource consumption,a state identification method based on KL divergence is employed. The effectiveness and superiority of the proposed method are validated through a real wind power data set. © 2024 Science Press. All rights reserved.