Separating emissions and meteorological impacts on peak ozone concentrations in Southern California using generalized additive modeling

Emerging investigator series: a machine learning approach to quantify the impact of meteorology on tropospheric ozone in the inland southern California

“…Since the concentration of ozone largely depends on precursor emissions and surface meteorology, ML was performed on a predetermined set of meteorological and air quality data to better capture the interactions of meteorology and emissions in an empirical model. 26…”

Multidecadal Analysis of Meteorological and Emissions Regimes for PM2.5 Across California

“…The modeling tool we employ in this study is the generalized additive model, introduced by Hastie and Tibshirani in 1986 . We use GAMs because of their ability to capture nonlinear relationships between covariates and the target variable to be modeled and for how they facilitate a covariate-by-covariate analysis of the resulting models. , A GAM is analogous to a generalized linear model (GLM) of the form g false( μ false) = β 0 + ∑ i β i x i + ϵ In eq , g is the link function that represents the relationship between the covariates x i and the expected value μ of the target variable, where the β i are the fitted model coefficients, β 0 is the intercept, and ϵ are the residuals. When g is the identity function, we have multiple linear regression, whereas we have logistic regression in the case where g is the logit link g false( p false) = log true( p 1 − p true) .…”

“…Our base model is .25ex2ex ŷ = log ( μ ) = β 0 + s ( TMAX ) + s ( WINDS ) infix+ s ( WINDD ) + s ( RH ) infix+ s ( SR ) + s ( RH 850 ) infix+ s ( WS 850 ) + s ( WD 850 ) infix+ s ( NOX ) + s ( HCHO ) infix+ s ( O 3 ) + ϵ In eq , ŷ is the model fit and the argument of the log link, μ, is the expected value of the target variable, which is the 24-hour average PM 2.5 mass or one of its speciated components. We use the log link in our models following precedent. ,, Moreover, we found the log link to reduce bias in our models’ fit against annual 98th percentile values as compared with the identity link. As before, β 0 is the model intercept and ϵ is the residual.…”

Spatiotemporal Dynamics of Surface Ozone and Its Relationship with Meteorological Factors over the Beijing–Tianjin–Tangshan Region, China, from 2016 to 2019

“…Since the dynamics of surface ozone are mutually influenced via interference coupling with many factors, including the emissions of nitrogen oxides (NO x ) and volatile organic compounds (VOCs) of ozone [28], the levels of fine particulate matter (e.g., PM 2.5 and PM 10 ) in the atmosphere [27], and meteorological conditions (such as atmospheric circulation patterns) [42,43], the relationships between surface ozone and its influencing factors are quite complex, making it very difficult to examine individual factors and independently analyze their effects on ozone. Therefore, correlation analyses between ozone and meteorological factors are also difficult, and there is unavoidable uncertainty in the analytical results [28].…”