期刊
APPLIED ENERGY
卷 228, 期 -, 页码 1783-1800出版社
ELSEVIER SCI LTD
DOI: 10.1016/j.apenergy.2018.07.050
关键词
Wind speed forecasting; Wavelet packet decomposition; Mother wavelet; Vanishing moment; AdaBoost.MRT; Outlier-robust extreme learning machine
资金
- National Natural Science Foundation of China [U1534210, 51308553]
- Changsha Science & Technology Project [KQ1707017]
- Shenghua Yu-ying Talents Program of the Central South University
- innovation driven project of the Central South University [502501002]
Accurate wind speed forecasting is essential for smart wind power conversion and integration. In the study, a novel ensemble model, using four novel hybrid models as base predictors to obtain high prediction accuracy, is proposed for the multi-step wind speed forecasting. The hybrid base predictors consist of the Wavelet Packet Decomposition (WPD), the Multi-Objective Grey Wolf Optimizer (MOGWO), the Adaptive Boosting.MRT (AdaBoost.MRT) and the Outlier-Robust Extreme Learning Machine (ORELM). The proposed ensemble model is named as the MOGWO-WPD -AdaBoost.MRT-ORELM model. The accuracy and diversity of the base predictors have significant positive influences on the performance of the proposed ensemble model. To guarantee the diversity of the base predictors, one of the most important hyper-parameters in the WPD computation (i.e., the mother wavelet) for every base predictor is investigated. In addition, the MOGWO is used to assemble the base predictors. By combining various models with different hyper-parameters, the ensemble structure can be used to improve the forecasting performance of the hybrid model with single hyper-parameter. To investigate the performance of the proposed forecasting architecture, four sets of experiments were conducted in the study. The results show that: (a) the proposed ensemble model has good convergence and forecasting performance; (b) the forecasting accuracy of the base predictor increases as the vanishing moment increases; and (c) the proposed ensemble model outperforms other benchmark models significantly.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据