【作者】 毛宇清；

【导师】 王咏青； 王革丽；

【作者基本信息】 南京信息工程大学 ， 气象学， 2007， 硕士

【摘要】 本文研究了一种新的机器学习方法——支持向量机(SVM)方法对非平稳时间序列的预测能力，并对其在气象领域中的应用进行了试验研究。全文分为以下三个部分：在论文的第一部分中，我们系统地介绍了支持向量机方法的基本思想、特点、回归方法的内容、预测建模思路以及支持向量机学习建模软件平台等。在论文的第二部分中，以33模Lorenz系统和虫口模型作为“理想”时空序列的“发生器”，利用SVM回归方法建立预测模型，并与人工神经网络(ANN)进行了比较。结果表明：(1)SVM方法不仅对平稳时间序列有较好的预报能力，也适用于非平稳时间序列，预报值与真实值的相关系数均能达到0.99以上；(2)SVM回归模型预报的准确率和效率均显著优于人工神经网络，平均相对误差优于ANN模型0.3％—0.5％。我们可以理解为SVM通过非线性映射，将低维空间中的非平稳过程映射到高维空间，一定程度上降低了系统的非平稳程度。论文的第三部分，我们选取实际资料，利用上述方法分别对北京密云县的温度和印度新德里地区的臭氧浓度进行了试预报。得到初步结论：(1)温度预报值与真实值吻合较好，二者的相关系数达到0.98以上；臭氧浓度预报值较真实值提前，相关系数为0.63，显示了SVM对实际序列有一定的预测能力；(2)随着预报因子的不断增加，预报值与真实值的相关系数略有上升，预报误差明显下降。这表明，训练样本中包含的信息越多，SVM方法建立的预报模型就越稳定；(3)采用相同的温度资料，ANN的预报误差大于SVM的预报误差，说明在实际应用中SVM回归方法的预报效果也优于ANN方法；(4)对两个个例的预测都显示SVM回归方法对一些拐点的预报误差较大，原因可能是建模受到资料的限制，在参与训练的预报因子中缺少与预报对象密切相关的物理量场。更多还原

【Abstract】 A new machine learning method—support vector machine (SVM) is used tobuild forecast models on the nonstationary time series in this paper, and theapplication in weather prediction field by this method is tested and analyzed aswell. Three parts are given as follows:In the first part, the SVM regression principal and basic ideas based on thestatistical learning theory, the main idea of the forecasting model and the CMSVMsoftware are introduced systematically.In the second part, the 33 modes Lorenz system and the logistic map are used asgenerators for chaotic spatio-temporal series. We build SVM regression forecastmodels and compare them with the artificial neural network (ANN) method. Thepreliminary results are: (1) The SVM method is available for both stationary timeseries and nonstationary ones, and the correlation coefficient between the predictedvalue and the actual value can reach above 0.99. (2) The SVM regression modelsgain an advantage over the ANN method on both the forecasting accuracy and thecomputing speed, and the average relative error is about 0.3%—0.5% less than thatof the ANN method. We can consider that when we mapped a nonstationary processin the low-dimension sample space to the high-dimension (infinite-dimension)feather space by a nonlinear mapping, the nonstationarity of the system is reduced.In the third part, we use the method mentioned before to try to predict thetemperature in Miyun country in Beijing and the ozone concentration in NewDelhi. The main results are: (1) The temperature predicted value match well with theactual value and the correlation coefficient between them can reach above 0.98. Theozone concentration predicted value are a little ahead that of the actual ones and the correlation between them is 0.63. This shows that the SVM method can be used inreal data prediction. (2) With the increase of the sample numbers, the correlationcoefficient between the predicted value and the actual value are raised a bit and theprediction errors are descended obviously. That means the more informationincluded in the training sample, the more stable the model built by SVM method willbe. (3) Comparing with the SVM method, the errors produced by the ANN methodare larger when using the same temperature data to predict. It turns out that the SVMregression method also has advantages in real data prediction. (4) Both of theexamples show that there exist high errors when predicting some inflexions by SVMregression method. That maybe ascribes to the limitation of the real data and thereare little physical quality fields which have close correlation with prediction objectsin the prediction factors when training.更多还原