Research and application of a hybrid model based on Meta learning strategy for wind power deterministic and probabilistic forecasting

Wind power forecasting is becoming an increasingly significant part in the operation and programming of electric force and power systems. However, highly precise wind power forecasting is still a difficult and challenging issue owing to the randomisation and transience of the wind series. In this paper, a novel Meta-learning strategy is proposed for adaptively combining heterogeneous forecasting models that were selected from a constructed candidate model bank. This study first implements the Box-Cox transformation to the wind speed and wind power sequence. Subsequently, the wind power as well as wind speed, which are decomposed by adopting the wind direction, are regarded as the inputs of the individual models. They are used to train a base-level forecasting learner to model the forecasting values of the wind power series. Finally, models with poor performances are dynamically trimmed and combining the remaining individual models are combined by adopting the random forest algorithm for the subsequent deterministic and probabilistic forecasting task. The wind power data from a wind farm located in northwestern of China are adopted to illustrate the forecasting effectiveness of the developed approach. The simulation in three experiments demonstrated the following: (a) the proposed Meta-learning based model is suitable for providing accurate wind power forecasting; (b) the proposed Meta-learning based hybrid model exhibits a more competitive forecasting performance than the individual models by extract advantage of each models; (c) the proposed model not only improves the accuracy of the deterministic forecasts but also provides more probabilistic information for wind power forecasting.