基于机器学习算法对二次无机气溶胶模拟进行偏差订正

Bias correction of the secondary inorganic aerosol modeling based on machine learning algorithm

采用机器学习算法随机森林结合空气质量模型NAQPMS,利用华北地区模式模拟结果及硫酸盐、硝酸盐和铵盐的观测结果,构建了二次无机气溶胶模拟的偏差订正模型.结果表明:基于随机森林算法融合空气质量模型建立的二次无机气溶胶模拟偏差订正模型可以显著改善二次无机气溶胶的模拟效果,更细致地再现出二次无机气溶胶的时空分布特征.对于训练站点,硫酸盐、硝酸盐和铵盐的模拟偏差可降低89.6%、16.7%和98.0%;对于验证站点,硫酸盐、硝酸盐和铵盐的模拟偏差可降低68.3%、60.0%和81.3%,相关系数均有显著提高.特征因子敏感性试验表明,硫酸盐、硝酸盐和铵盐的模拟浓度是构建二次无机气溶胶模拟的偏差订正模型的重要特征因子,潜在地考虑了二次无机气溶胶物理化学生成过程.本文揭示了融合了深度学习方法和传统数值模式方法在改善区域二次无机气溶胶模拟上的巨大潜力.

Based on the simulation of meteorological,emission and air pollutants in North China and the observation of sulfate,nitrate and ammonium,a bias correction model of secondary inorganic aerosol simulation was established by machine learning algorithm random forest combined with the Nested Grid Air Quality Forecast Model System(NAQPMS).The results show that the bias correction model of secondary inorganic aerosols based on random forest algorithm combined with air quality model can significantly improve the simulation of secondary inorganic aerosols,and reproduce the spatial and temporal distribution characteristics of secondary inorganic aerosols in more detail.For the training station,the model bias of sulfate,nitrate and ammonium decreased by 89.6%,16.7%and 98.0%.For the verification station,the model bias of sulfate,nitrate and ammonium decreased by 68.3%,60.0%and 81.3%,and the correlation coefficients were significantly improved.The sensitivity test of characteristic factors showed that the simulation of sulfate,nitrate and ammonium were important characteristic factors for constructing the bias correction model of secondary inorganic aerosol simulation,which potentially considered the physicochemical generation process of secondary inorganic aerosol.This paper revealed the great potential of combining deep learning with traditional numerical model to improve the regional simulation of secondary inorganic aerosol.