| 基于wavelet-SVM的PM10浓度时序数据预测 |
| 摘要点击 1625 全文点击 653 投稿时间：2016-12-22 修订日期：2017-03-18 |
| 查看HTML全文
查看全文 查看/发表评论 下载PDF阅读器 |
| 中文关键词 支持向量机 小波变换 大气污染物浓度预测 输入向量 预测模型 |
| 英文关键词 SVM wavelet transform air pollutant concentration forecasting input variables forecasting model |
|
| 中文摘要 |
| 太原是以煤炭为主要能源的重工业城市，PM10（particulate matter）是太原市的主要大气污染物，因此研究其变化趋势，并给出污染物浓度预测结果，为相关部门进行大气污染防治，为突发污染事件应急提供理论支持是一项非常重要的工作. 支持向量机（support vector machine，SVM）应用于PM10污染物浓度时序数据预测时，表现出良好的泛化能力. 在预测模型建立过程中通常选择历史数据作为学习模型的输入特征，然而这样的数据表示形式，结构单一，信息表达不完备，在很大程度上将影响预测模型的泛化能力. 本文以山西省太原市城区4个监测站点的PM10日浓度数据为研究数据，通过小波变换（wavelet transform）将一维输入数据转化为由低频信息和高频信息构成的高维数据，并以该数据为输入数据建立wavelet-SVM预测模型. 结果表明，相较于传统SVM模型预测，wavelet-SVM模型预测结果具有更高的精度，尤其能更加准确捕捉到PM10浓度突变点，为大气污染预警提供有效信息支持，并且wavelet-SVM模型对于PM10浓度时序数据变化趋势的预测精度有明显提升，能更好地预测PM10浓度变化趋势，揭示PM10浓度时序数据内在规律. |
| 英文摘要 |
| PM10 is the main air pollutant in Taiyuan, as the city is a heavy industrial center with coal as its main energy source. Therefore, research on the prediction of this pollutant's variation and concentration is of great theoretical significance for air pollution prevention and emergency solutions. The source of PM10 is very complex, as it is affected by industrial emissions, vehicle exhaust, fugitive dust, and many other factors. The emission sources of PM10 are difficult to determine accurately. The goal of our research was to give accurate forecasting results efficiently when only time-series PM10 concentrations, and no other exogenous information, is available. A support vector machine (SVM) enjoys good generalization performance in the PM10 concentration forecasting area. Traditionally, an SVM chooses historical data as the input features in the process of dealing with the time-series data of air pollutant concentrations. However, data with simple structure and incomplete information have become the fetter of generalization ability improvement. In this study, the data for simulation experiments was the PM10 concentration dataset collected from four monitoring stations in Taiyuan. The PM10 concentration time-series one-dimension data was decomposed into high dimension, constructed by low frequency and high frequency series using a wavelet transform. The wavelet-SVM forecasting model can be established by introducing the high-dimension data as the input features. The experiment results indicate that, contrasted with the traditional SVM, the wavelet-SVM model boasts higher accuracy for PM10 concentration prediction. In particular, it captures the concentration mutational points more accurately and provides information support that is more effective for atmospheric pollution warning. In addition, with the wavelet-SVM model, prediction accuracy for the concentration variations was significantly improved and laws that were more inherent in the PM10 concentration time series were revealed. |
|
|
|
