基于主成分分析的 RBF 神经网络预测 SO2 浓度

利用基于主成分分析 (PCA) 算法的径向基 (RBF) 神经网络对大气中 SO2 浓度进行滚动预测。以北京大兴地 区 2019 年 9 月 1 日至 2020 年 10 月 31 日的气象数据和空气质量参数为基础, 结合逐步回归法筛选出与 SO2 线性相 关的参数作为输入样本, 构建 PCA-RBF 预测模型。利用该模型预测北京大兴地区某天的 SO2 浓度, 将预测值保留并 作为下一天预测模型的输入参数。以此将预测值不断地向前延伸并进行分析和预测, 从而实现 SO2 浓度的滚动预测。 对比 RBF 网络和 PCA-RBF 网络两种模型的预测结果, 其中 PCA-RBF 模型期望值和预测值的误差及相关系数分别为 0.03 µg·m?3 和 0.9989。表明 PCA-RBF 网络模型能精准预测 SO2 浓度变化趋势, 为进一步解决大气污染问题提供技术 支持。

Prediction of SO2 concentration by RBF neural network based on principal component analysis

The rolling prediction of atmospheric SO2 concentration using radial basis function (RBF) neural network based on principal component analysis (PCA) algorithm is presented. Based on the meteorological data and air quality parameters from September 1, 2019 to October 31, 2020 in Daxing district of Beijing, the PCA-RBF prediction model is constructed by combining the stepwise regression method to select the high correlation parameters between SO2 and meteorological factors as input samples. Then the PCA-RBF prediction model is used to predict the SO2 concentration in Daxing area on a certain day, and the prediction results are retained as the input parameters of the prediction model for the next day. In this way, the predicted value is continuously extended forward for the following analyzation and prediction, so as to realize the rolling prediction of SO2 concentration. Comparing the predicted results of RBF network and PCA-RBF network, the error and correlation coefficient of expected value and predicted value of PCA-RBF model are 0.03 µg·m?3 and 0.9989. It is shown that the PCA-RBF network model can accurately predict the variation trend of SO2 concentration, and provide a new technical support for further solving the air pollution problem.