基于MKL方法的短期风电功率预测研究

发布时间：2018-11-22 13:31

【摘要】：支持向量机(Support Vector Machine,SVM)等核学习方法是解决非线性问题的一种有效方法,在短期风电功率预测中已有成功的应用。多核学习(Multiple Kernel Learning,MKL)作为一种的新型核学习方法,通过核权值系数将具有不同特性的核函数进行组合,其核权值系数使得核函数的选择问题转化为核权值系数的分布问题,且核权值系数的稀疏性能增强决策函数可解释性,其不同核函数组合形成的再生希尔伯特空间使模型具有更强的泛化能力与鲁棒性。为了进一步提高短期风电功率预测模型的性能,以MKL方法为主线,研究其在短期风电功率直接预测与间接预测方面的应用。本文的主要研究内容如下:(1)分析了用于数据预处理的经验模态分解方法(Empirical Mode Decomposition,EMD)和经验小波变换方法(Empirical Wavelet Transform,EWT)的基本原理及其实现步骤,并通过ECG(Electrocardiograph,心电图)标准数据集对其进行对比分析,实验结果表明,EWT分解得到模态信号分量数量明显少于EMD得到的模态信号分量数量且EMD分解得到的模态分量存在明显的模态混叠现象。在SVM理论的基础上,对基于半无限线性规划的多核学习及MKL-wrapper算法和MKL-chunking算法进行了深入研究,并简要阐述了Simple MKL方法的基本原理及其具体实现步骤。(2)分析了某大型风电场输出功率不同季节中的季节周期性和时间连续性的特点,并从不同季节中随机选取四个具有不同特性测试周的风电功率数据作为测试集;将自适应分解预处理方法EWT与由MKL-wrapper、MKL-chunking、Simple MKL算法实现的MKL方法结合,形成一种新的组合预测方法,即EWT-MKL方法;将不同MKL方法应用于不同季节的短期风电功率直接预测实例中,在同等条件下,并与SVM方法进行对比。实验结果表明,MKL预测模型的精度优于SVM方法,而不同算法实现的EWT-MKL组合预测模型的效果最好,不同季节测试集中MKL模型的核参数及惩罚函数在取值范围内的随机取值及其实验结果表明,MKL具有较强的泛化能力且其对参数的选择具有较强的鲁棒性。(3)分析了不同“风速-功率”特性曲线求解方法对风速-功率转换精度的影响;将不同算法实现的MKL预测方法及EWT-MKL组合预测方法应用于某风电场平均风速的短期预测;结合“风速-功率”特性曲线实现短期风电功率间接预测,并在同等条件下与小波支持向量机(Wavelet Support Vector Machines,WSVM)方法进行对比。实验结果表明,在短期风电功率间接预测中,不同算法实现的EWT-MKL组合预测模型的精度明显高于MKL、SVM及WSVM等方法,而MKL预测模型的精度优于SVM方法建立的预测模型。
[Abstract]:Support Vector Machine (Support Vector Machine,SVM) is an effective method to solve nonlinear problems and has been successfully applied in short-term wind power prediction. As a new kernel learning method, multi-kernel learning (Multiple Kernel Learning,MKL) combines kernel functions with different characteristics through kernel weight coefficients, and the kernel weight coefficients transform the selection of kernel functions into the distribution of kernel weight coefficients. The sparse property of the kernel weight coefficient enhances the interpretability of the decision function, and the reproducing Hilbert space formed by the combination of different kernel functions makes the model have stronger generalization ability and robustness. In order to further improve the performance of short-term wind power prediction model, the application of MKL method in direct and indirect prediction of short-term wind power is studied. The main contents of this paper are as follows: (1) the basic principle and implementation steps of empirical mode decomposition (Empirical Mode Decomposition,EMD) and empirical wavelet transform (Empirical Wavelet Transform,EWT) for data preprocessing are analyzed. The experimental results show that the number of modal signal components obtained by EWT decomposition is obviously less than that obtained by EMD, and the modal components obtained by EMD decomposition have obvious modal aliasing phenomenon. On the basis of SVM theory, the multi-core learning based on semi-infinite linear programming, MKL-wrapper algorithm and MKL-chunking algorithm are studied. The basic principle of Simple MKL method and its realization steps are briefly described. (2) the characteristics of seasonal periodicity and time continuity in different seasons of output power of a large wind farm are analyzed. Four wind power data with different characteristic test weeks were randomly selected from different seasons as the test set. Combining the adaptive decomposition preprocessing method (EWT) with the MKL method realized by MKL-wrapper,MKL-chunking,Simple MKL algorithm, a new combined prediction method, EWT-MKL method, is formed. The different MKL method is applied to the direct prediction of short-term wind power in different seasons. Under the same conditions, the method is compared with the SVM method. The experimental results show that the precision of MKL prediction model is better than that of SVM method, and the effect of EWT-MKL combination prediction model realized by different algorithms is the best. The random values of the kernel parameters and penalty functions of the MKL model in different season test sets are obtained in the range of values and the experimental results show that, MKL has strong generalization ability and strong robustness to parameter selection. (3) the influence of different "wind speed power" characteristic curve solving method on the precision of wind speed power conversion is analyzed. The MKL forecasting method and the EWT-MKL combination forecasting method realized by different algorithms are applied to the short-term prediction of the average wind speed of a wind farm. Combined with the characteristic curve of "wind speed and power", indirect prediction of short-term wind power is realized, and compared with wavelet support vector machine (Wavelet Support Vector Machines,WSVM) method under the same conditions. The experimental results show that the accuracy of EWT-MKL combination prediction model implemented by different algorithms is obviously higher than that of MKL,SVM and WSVM methods in indirect prediction of short-term wind power, while the accuracy of MKL prediction model is better than that of SVM method.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM614

【参考文献】