Comparing the Performance of Statistical Models for Predicting PM10 Concentrations

Abstract

The ability to accurately model and predict the ambient concentration of Particulate Matter (PM) is essential for effective air quality management and policies development. Various statistical approaches exist for modelling air pollutant levels. In this paper, several approaches including linear, non-linear, and machine learning methods are evaluated for the prediction of urban PM 10 concentrations in the City of Makkah, Saudi Arabia. The models employed are Multiple Linear Regression Model (MLRM), Quantile Regression Model (QRM), Generalised Additive Model (GAM), and Boosted Regression Trees1-way (BRT1) and 2-way (BRT2). Several meteorological parameters and chemical species measured during 2012 are used as covariates in the models. Various statistical metrics, including the Mean Bias Error (MBE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), the fraction of prediction within a Factor of Two (FACT2), correlation coefficient (R), and Index of Agreement (IA) are calculated to compare the predictive performance of the models. Results show that both MLRM and QRM captured the mean PM 10 levels. However, QRM topped the other models in capturing the variations in PM 10 concentrations. Based on the values of error indices, QRM showed better performance in predicting hourly PM 10 concentrations. Superiority over the other models is explained by the ability of QRM to model the contribution of covariates at different quantiles of the modelled variable (here PM 10 ). In this way QRM provides a better approximation procedure compared to the other modelling approaches, which consider a single central tendency response to a set of independent variables. Numerous recent studies have used these modelling approaches, however this is the first study that compares their performance for predicting PM 10 concentrations.

Particulate Matter Sampling Techniques and Data Modelling Methods

“…Litle diference was observed between the two models, suggesting that the simpler MLR method was a beter option than GLM in that case. Studies in other locales have reported that PM 10 was normally distributed [26] or that PM 10 concentrations were right skewed [14]. Studies have found that the use of curvilinear transformations of input variables e.g.…”

“…QR has been used very litle for modelling pollutants. One study that compared models developed using QR with MLR found that QR was beter at predicting hourly PM 10 concentrations [14].…”

“…MLR models are considered to be limited models of PM 10 concentration due to the inability to extend the response to non-central locations of the explanatory variables and to meet the other assumptions of the model [14,33]. Despite possible nonconformity with one or more of the assumptions, MLR has been used extensively for predicting PM 10 and is often used as a benchmark to which other methods are compared.…”

“…The outputs of numerical weather forecast models can also be used as input variables in PM 10 models. However, this is not common because of the uncertainties such variables introduce to PM 10 predictions [14,15].…”

“…This is in part due to the fact that in some countries, there is no rain for long periods of time or litle rainfall in summer. The lack of rain data means that it is not often included in PM 10 models [14].…”