Valuation of GDAS atmospheric stability data usage on PM2.5 predictions

Air pollution prediction is an appropriate tool for minimizing the adverse impact of pollutants on public health. The use of independent variables (e.g., meteorological factors) on the prediction of local pollution levels is a principal research challenge in that area. The limited availability of local meteorological data measurements encourages researchers to find alternative and competitive data sources (e.g., forecast data, satellite data, and mesoscale model products). The main objective of this study is to investigate the value of Global Data Assimilation System (GDAS) atmospheric stability forecast data (e.g., boundary layer depth, friction velocity, vertical mixing, and horizontal mixing coefficients) to employ in local air pollution prediction model. In addition to GDAS data, conventional surface meteorological measurements (e.g., surface temperature, surface pressure, wind speed, humidity, precipitation) were also used in the model to compare their prediction performances with the forecast data. A case study was performed to test the level of the explanatory power of the GDAS data in a local application in Nur-Sultan, which is one of the youngest capitals of the world, was constructed 20 years ago, and has become one of the most polluted cities in Kazakhstan. Three-hourly concentrations of inhalable fine particle (PM2.5) concentrations measured in the city from June 2018 to June 2019 were selected as the experimental dataset. PM2.5 measurements were obtained from the air quality monitoring network of the U.S. Embassy & Consulate in Kazakhstan; located downtown Nur-Sultan. All local meteorological data were acquired from the "KazHydroMet" (Kazakh National Hydro-Meteorological Center), and atmospheric stability data were obtained from the GDAS operational system, which is initially retrieved from the National Centers for Environmental Prediction (NCEP) reanalysis model datasets. A benchmark Multivariate Regression Analysis (MRA) model was constructed to test the study assumptions. Firstly, missing data handling was performed before the analysis stage. A logarithmic transformation was performed on the data to achieve a higher correlation, followed by a stepwise MRA method to select the most appropriate predictors. A lag analysis was also considered in the prediction model and utilized by using techniques such as autocorrelation and cross-correlation between predictors and dependent variables. The models were chosen based on the highest adjusted coefficient R² and a final MRA model was developed. Correlation analysis of the selected parameters with the pollution levels showed that some of the forecast data (e.g., boundary layer depth) have a higher correlation with PM2.5 concentrations than surface met data. The results indicate that PM2.5 concentrations have a significant positive relationship with pressure and a negative correlation with temperature, boundary layer depth, and vertical mixing coefficient. The regression model based on shifted transformation was effective, achieving a high level of goodness-of-fit. The model performance was entirely satisfactory (up to 74%) when preceding 3-hr average PM2.5 data was employed as an independent parameter. Our research findings suggest that GDAS atmospheric stability time-series data may help authorities obtain early information for preserving the air quality. However, the increase in performance after using those parameters is not high, and it should be taken into consideration before implementation.