EWA-GBDT: A Novel Air Quality Prediction Model Combining Exponentially Weighted Averages and Gradient Boosting Decision Tree

Abstract

Air quality prediction is a hot topic in the field of meteorology. Challenges still exist following consideration of the uncertainty of atmospheric pollutant emission sources, as well as the multi-dimensional, multi-scale, and non-stationary characteristics of meteorological environment data. For example, traditional statistical forecasting methods usually fit a nonlinear relationship between meteorological features and pollutants, which is why it is extremely difficult to learn their models. To address these challenges, we propose a novel air quality prediction model combining Exponentially Weighted Averages and Gradient Boosting Decision Tree (EWA-GBDT). More specifically, we first collected two real-world datasets, including 1) the daily concentration data of six pollutants from 01/01/2014 to 31/12/2016 and 2) the daily concentration data of meteorological features in the cities over Shijiazhuang and Xingtai from 01/01/2014 to 02/28/2017. Then, we extracted 13 types of meteorological features using the Support Vector Machine Recursive Feature Elimination method. From the respective of pollutant concentration, these features are the highest correlated with each other. Next, we applied the EWA principle to compute the above features and pollutant concentration for obtaining meteorological expectation values. Finally, considering the excellent overall prediction performance of ensemble learning, we utilized its GBDT algorithm to predict the concentration value of pollutants and thus output the air quality level. We conducted experiments on the two datasets, and the results demonstrate that EWA-GBDT outperforms other baseline methods in terms of root-mean-square deviation, mean absolute error, and the square of R.