【摘要】 提出一种基于卷积神经网络（CNN）和支持向量回归机（SVR）的多轴差分光学吸收光谱（MAX-DOAS）对流层NO₂垂直分布预测方法。将2019年南京站点采集的原始MAX-DOAS数据通过QDOAS软件拟合获取O4和NO₂差分斜柱浓度，结合基于最优估算的气溶胶和痕量气体廓线反演算法——PriAM算法反演了对流层NO₂廓线，并将其作为预测模型的输出。此外，通过平均影响值方法进行预测模型输入变量的选择，确定了MAX-DOAS数据、温度、气溶胶光学厚度和低云覆盖率为模型的最佳输入变量。通过实验优化网络结构和参数，最终建立预测模型在测试集与PriAM的平均百分比误差仅为9.14%，与单独建立的CNN、SVR、反向传播模型相比，平均百分比误差分别降低了8.22%、6.00%、32.28%。因此，CNN-SVR能够利用MAX-DOAS数据对对流层NO₂廓线进行有效预测。更多还原

【Abstract】 This study proposes a method based on a convolutional neural network（CNN） and support vector regression machine（SVR） for predicting the vertical distribution of NO₂ in the troposphere by multi-axis differential optical absorption spectroscopy（MAX-DOAS） technology. Taking the Nanjing site as an example, we obtain the O4 and NO₂ differential slant column density（dSCD） according to the raw MAX-DOAS data collected by QDOAS fitting in 2019, invert the tropospheric NO₂ profile by combining the optimal estimation-based aerosol and trace gas profile inversion algorithmPriAM, and use the profile as the output of the prediction model. In addition, the input variables of the prediction model are selected by the mean impact value method, with MAX-DOAS data, temperature, aerosol optical thickness, and low cloud coverage finally identified as the optimal input variables for the model. Furthermore, the network structure and parameters are optimized through experiments, and the average percentage error of the final CNN-SVR prediction model in the test set with PriAM is only 9. 14%, which is 8. 22%, 6. 00%, and 32. 28% lower than that of the separately constructed CNN, SVR, and backpropagation models, respectively. Therefore, CNN-SVR can effectively predict tropospheric NO₂ profiles by using MAX-DOAS data.更多还原