【作者】 黄静华；

【导师】 王振翀；

【作者基本信息】 中国矿业大学（北京） ， 控制理论与控制工程， 2011， 博士

【摘要】 针对气象数据具有时空属性、多维、复杂、相关度高等问题,提出基于SVM解决时序多维气象预测的方法。在进行气象数据挖掘时,预报国子的选择和处理是决定预测精度和有效性的关键所在。本文针对基于精度的分类算法在不平衡数据挖掘中表现出有偏性问题,提出解决类不平衡挖掘的补偿方式和重构样本集的SVM算法,将噪声代价嵌入SVR模型中,有效减小噪音数据引起的回归模型的过学习问题。本文给出气象数据挖掘中,选择预报国子的基本方法和处理原则,同时对大规模数据的处理和分类问题、高维特征数据的特征约简和特征优选、SVM核参数的选择和优化、以及采用增量学习算法增强模型的自适应能力等问题进行了探讨和研究。实验结果表明,基于SVM的气象数据挖掘具有较高的预测正确性。更多还原

【Abstract】 This paper proposes a method for weather forecasting of multi-dimensional, time-series meteorological data based on SVM algorithm, which could effectively solve the troublesome problems of meteorological data mining, as the attributes of meteorological data have the characters of multi-dimension, complex, high dependency and correlation, also concerning with time and space factors. The key point to obtain the satisfactory accuracy and effectiveness of meteorological data mining is to effectively select and correctly process the forecasting factors. This paper proposes an approach to compensate the imbalance of classes and to reconstruct training sample set for SVM, which could solve the bias problem of the classification algorithms. With the cost factor of noise data embedded into SVR model, the SVR algorithm obtains the effective regression result in solving problem of overfitting. This paper researches and discusses on the basic methods and principles of selecting factors, processing and classification of large-scale data, reduction and optimization of high dimensional quantization factors, selection and optimization of SVM kernel function’s parameters, and the incremental learning SVM algorithm. The experiments have showed that meteorological data mining based on SVM has satisfactory performance and efficiency.更多还原