【作者】 赵强；

【导师】 杨正瓴；

【作者基本信息】 天津大学 ， 控制科学与工程， 2019， 硕士

【摘要】 由于能源供应紧张和环境污染加剧,对清洁能源的研究受到了越来越多的重视。风是一种高品质的清洁能源,但因其不稳定的性质,风电的应用受到一定程度的限制。而风电功率预测是解决风电并网问题的重要途径。为了提高超短期风速预测的准确率和可靠性,根据“机理+辨识”预测策略将时间序列的k近邻预测算法扩展到空间相关性风速预测领域。以风向为标准优选被预测地点(本地)的上游地点,随后计算各个上游地点风速对本地风速的最优延迟时间。将本地最近的风速历史观测值,与按照最优延迟时间提前的上游风速观测值相结合,形成k近邻空间相关性预测的参考矢量。并以Pearson相关系数为标准,从风速的历史观测值中优选出该参考矢量的k个最相似的近邻。采用线性回归、偏最小二乘回归、最小二乘支持向量回归、BP神经网络、RBF神经网络、广义回归神经网络和随机森林共7种回归模型,对本地的未来风速进行预测。以荷兰的Huibertgat为被预测地点,对其冬季时期风速预测的仿真表明:线性回归、偏最小二乘回归和最小二乘支持向量回归是比较优化的回归模型,优化的k近邻数量选择在100左右,并且采用近10年的历史风速预测效果较好。算例分析结果表明:k近邻空间相关性风速预测,能够有效利用历史数据的相似性,给出可靠的超短期风速预测结果。更多还原

【Abstract】 Due to the tight supply of energy and the increasing of environmental pollution,more and more attention has been paid to the research of clean energy.Wind energy is a kind of high-quality clean energy,however,because of its instability,the application of wind power is subject to certain restrictions.Wind power prediction is an important way to solve the problem that wind power is difficult to parallel in the grid.In order to improve the accuracy and reliability of ultra-short term wind speed prediction,the k-nearest neighbor prediction of time series is generalized to the spatial correlation wind speed prediction according to the “mechanism model + identification model” strategy.Firstly,the wind speed upstream sites are sorted out through their wind directions to the predicted site(the local site),and their optimal lag time of wind speed to the local site is calculated.The reference vector of k-nearest neighbor prediction based on spatial correlation is a combination of the latest local wind speed historical observations and the upstream wind speed observations adjusted by its lag time.The k most similar neighbors of the reference vector are sought out from the wind speed historical observations by the Pearson product-moment correlation coefficient.The future wind speed of the local site is regressed by 7 models,include linear regression,partial least squares regression,least squares support vector machine regression,back propagation neural network,radial basis function neural network,generalized regression neural network and random forest.The numerical experiments of the prediction of wind speed of Huibertgat in the Netherlands in winter show that the linear regression,partial least squares regression and least squares support vector machine regression are more optimized regression models,and the optimal quantity of k-nearest neighbors is around 100.Using historical wind speed observations of 10 years gives a better result.The case studies support that the spatial correlation based on k-nearest neighbor can effectively use the similarities in historical data,then predict the ultra-short term wind speeds reliably.更多还原