京津冀大气能见度特征分析及影响因素研究

Analysis of characteristics and influencing factors of atmospheric visibility in Beijing-Tianjin-Hebei region

为研究京津冀能见度状况和分析影响能见度的特征贡献模式,基于2019年京津冀气象站点和空气质量监测站点数据研究能见度时序变化特征,运用随机森林算法建立能见度估算模型分析影响因子整体解释度,并基于SHAP框架结合随机森林模型构建能见度影响因子可解释模型,对特征因子贡献大小、方向以及单变量贡献情况进行了详细解释和分析:(1)能见度状况在早晚高峰时较差,每日15时左右最好,工作日和非工作日无明显差别,从季节上看冬季能见度最差;(2)随机森林模型拟合系数解释方差为0.8973,R2为0.8978,拟合结果良好;(3)根据SHAP可解释模型分析结果可得,PM2.5是影响能见度的最重要因子,呈负向相关,且贡献度变化率以浓度100μ/m3为转折点由急促转向平缓。实验证明,基于SHAP框架的能见度解释模型不仅能反映贡献度的大小以及影响效应的方向,而且可以对单个变量的贡献进行详细分析,提高了特征贡献分析的精细度和准确性。

In order to study the visibility status of Beijing-Tianjin-Hebei and analyze the characteristic contribution model that affects visibility,this paper analyzed the time-series characteristics of visibility using the data of the Beijing-Tianjin-Hebei meteorological stations and air quality monitoring stations in 2019,then established a visibility estimation model with random forest to analyze the overall interpretation of impact factors,thus constructed an interpretable model of the visibility impact factors combined with the random forest based on SHAP framework,providing the explanation of the contribution of impact factors and carried out the univariate contribution analysis detailedly.The results showed that:(1)The visibility is poor during the morning peak and evening peak,which becomes the best at 3 and 4 pm,and there is no significant difference between workdays and nonworkdays,and it becomes the worst in winter seasonally.(2)The explained variance is 0.8973 and the fitting coefficient R2 is0.8978,showing that the model fits well.(3)According to the SHAP interpretable,PM2.5 is the most important factor affecting visibility,which is negatively correlated,and the change rate of contribution degree changes from rapid to gentle with 100μ/m3 as the inflection point.It can be shown that the SHAP interpretation framework can not only reflect the magnitude of the contribution,but also the direction of the influence effect,and it can conduct a detailed analysis of the single variable on the model,improved the precision and accuracy of feature contribution analysis.