论文部分内容阅读
O3是近地面大气中一种重要的二次污染物.本研究采用神经网络多层感知器(Multi-Layer Perceptron)和多元线性回归2种模型,以广州万顷沙站2006年的气象观测数据为输入,对该站O3的1 h平均峰值浓度进行提前1 d的预测,并比较了2种模型的预测效果.模型的输入参数为前1d O3的最高1h平均浓度和第二天的气象参数(温度、湿度、风速、风向、气压和光照).为了降低神经网络的复杂度以提高模型的泛化能力,采用了OBS(Optimal brain surgeon)方法对神经网络模型进行了修剪.结果表明,经过修剪后的神经网络预测结果的准确指数(agreement index)为92.3%,RMSE为0.042 8 mg/m3,R-square为0.737,重污染事件(1 d中O3峰值浓度超过0.20 mg/m3)的预报准确率为77.0%.为了进一步提高重污染事件发生概率大小的预报效果,采用了神经网络分类器对臭氧的污染级别进行预测,该处理后重污染事件预报准确率可以达到83.6%.综合比较神经网络模型和多元线性回归模型的拟合效果后发现,神经网络模型在O3峰值预报中具有明显优势,本研究建立的神经网络模型具有臭氧污染预测预警的实用价值.
O3 is an important secondary pollutant in the near-surface atmosphere.In this study, two models of neural network Multi-Layer Perceptron and multiple linear regression were used to study the meteorological observation data of 2006 at Wanqingsha Station in Guangzhou , The 1-h average peak concentration of O3 in this station was predicted 1 day in advance and the prediction results of the two models were compared.The input parameters of the model were the highest 1-hour mean concentration of O3 and the meteorological parameter of the next day Temperature, humidity, wind speed, wind direction, air pressure and light.) In order to reduce the complexity of the neural network and improve the generalization ability of the model, the neural network model was pruned by using Optimal brain surgeon (OBS) The accuracy of the post-neural network prediction was 92.3%, RMSE was 0.042 8 mg / m3, R-square was 0.737, and forecast of heavy pollution events (O3 peak concentration over 0.20 mg / m3 in 1 day) was accurate The rate of 77.0% .In order to further improve the prediction of the probability of occurrence of heavy pollution incidents, the neural network classifier is used to predict the pollution level of ozone. After the treatment, the prediction accuracy of heavy pollution incidents can reach 83.6% By comprehensively comparing the fitting results of neural network model and multivariate linear regression model, it is found that the neural network model has obvious advantages in the O3 peak forecasting. The neural network model established in this study possesses the practical value of forecasting the ozone pollution.