基于Stacking集成学习模型的气态亚硝酸预测

建立了基于Stacking集成学习下气态亚硝酸（HONO）预测模型.利用非相干宽带腔增强吸收光谱（IBBCEAS）系统获得的北京城区HONO的浓度，结合HONO的来源，选取了O₃、CO、SO₂、NO、NO₂、NO_y、温度（T）、相对湿度（RH）、风速（WS）、j（HONO）、j（NO₂）、j（O¹D）作为特征数据，通过对HONO的平均日变化分析，将测量时间按小时转换为新特征.分别以极端梯度提升（XGBoost）、轻量化梯度促进机（LightGBM）以及随机森林（RF）算法构建基模型，采用5折交叉验证的方式划分训练集，将基模型输出的结果作为新特征集，并将新特征集作为第二层线性回归模型的输入，通过对这两层中的模型进行训练，最终得到Stacking集成学习HONO预测模型.通过对模型的特征重要度分析和计算夜间交通直接排放所占的贡献，表明CO是模型预测中重要的影响因子，说明机动车的直接排放是该区域冬季时期HONO的重要来源.利用测试集分别对单模型和融合后模型的预测性能进行评估，3个单模型的预测结果与测量值的相关系数都达到了0.91以上，其中Stacking融合后的模型性能最好，相关系数达到了0.94，平均绝对误差和均方根误差分别为0.307×10^-9和0.453×10^-9，结果表明基于Stacking集成学习方式下HONO预测模型的可解释性和推广性.

Prediction of gaseous nitrous acid based on Stacking ensemble learning model

A gaseous nitrous acid (HONO) prediction model based on Stacking ensemble learning was proposed. The concentrations of HONO in Beijing urban area were obtained using incoherent broadband cavity enhanced absorption spectroscopy (IBBCEAS). Combined with the HONO sources, O₃, CO, SO₂, NO, NO₂, NO_y, temperature (T), relative humidity (RH), wind speed (WS), j(HONO), j(NO₂), j(O¹D) were selected as characteristic data. By analyzing the average diurnal variation of HONO, the measurement time was converted into a new feature hour by hour. The base model was constructed by utilizing Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM) and Random Forest (RF) algorithm. The training set was partitioned by 5-fold cross-validation method. The output of the base model was taken as a new feature set and as the input of second-level linear regression model. HONO prediction model was finally obtained via training the models in these two layers. Through the feature importance analysis and calculating the contribution of direct emission of vehicles at night, it showed that CO was an important impact factor in the prediction model, and that the direct emission of vehicles was a major source of HONO in the winter period at the region. The prediction performance of the base model and the Stacking ensemble model were evaluated by the test set respectively. The correlation coefficients between forecast results and measured values for the three base models were above 0.91. The performance of the Stacking ensemble model was the best, whose correlation coefficients reached 0.94. The average absolute error and root mean square error were 0.307×10^-9 and 0.453×10^-9, respectively. Explanability and applicability of the HONO prediction model based on Stacking ensemble learning.