摘要
通过对大同市2011-2012年PM10质量浓度、有关气象要素和参数进行随机抽样、分组,建立单隐含层BP神经网络、多隐含层BP神经网络以及RBF网络对以上数据进行调试和训练,得出:就2011-2012年预测PM10日均质量浓度样本而言,按预测效果好坏排序,多隐含层最佳BP神经网络﹥单隐含层最佳BP神经网络﹥RBF神经网络;从网络最小误差总和来看,三种网络对夏秋两季的预报效果最好;从预测值和实测值的拟合效果来看,RBF网络对春季和冬季PM10质量浓度的预测效果最好;多隐含层BP网络对秋季PM10质量浓度的预测效果最好;三种神经网络对夏季PM10质量浓度的预测效果都很好,优劣性差异不大。
Based on the random sampling and grouping of PM10 mass concentration from 2011-2012, meteorological elements and parameters in Datong, a single hidden layer of BP neural network and hidden muhilayer BP neural network and RBF network were established and debugged and trained. It is concluded that： for predicted samples of PM10 daily average mass concentration during 2011-2012, according to the forecast effect, hidden muhilayer optimum BP neural network 〉 RBF neural network; optimum BP neural network 〉 single hidden layer from the view of network minimum error sum, three kinds of network to forecast effect is best for summer and autumn; from the point of the fitting effect in forecast values and measure values, RBF network for the spring and winter PMIO mass concentration prediction effect is the best; hidden muhilayer BP network to fall prediction effect of PMIO mass concentration is the best; Three kinds of neural networks for summer PMIO mass concentration prediction effect is very good and the difference is not big.
出处
《气象研究与应用》
2014年第1期50-55,90,共7页
Journal of Meteorological Research and Application
基金
山西省气象局2013年青年基金课题(SXKQNDQ20138764)资助
关键词
神经网络
气溶胶质量浓度
预测研究
neural network
aerosol mass concentration
prediction research
