Abstract
Air quality prediction is a critical task in meteorology, complicated by the uncertainty of pollutant sources and the multi-dimensional, non-stationary nature of meteorological data. Traditional statistical methods often struggle to model the complex nonlinear relationships between weather features and pollutants. This paper introduces EWA-GBDT, a hybrid model that integrates Support Vector Machine Recursive Feature Elimination (SVM-RFE), Exponentially Weighted Averages (EWA), and Gradient Boosting Decision Trees (GBDT) to enhance prediction accuracy.
Introduction
Predicting air quality is notoriously difficult due to the "messy" nature of atmospheric data. Meteorological variables like wind speed, humidity, and temperature exhibit highly non-linear relationships with pollutant concentrations. Furthermore, raw sensor data is frequently plagued by noise and missing values, hindering the generalization of standard machine learning models. The EWA-GBDT framework addresses these challenges by filtering irrelevant data noise and focusing on underlying trends that drive pollution levels.
Methodology
The EWA-GBDT model employs a three-stage pipeline designed for robust feature selection and noise reduction.
- Feature Selection (SVM-RFE): To avoid overfitting and reduce dimensionality, the model uses SVM-RFE. This algorithm iteratively trains a model and prunes the least significant features, ultimately identifying the 13 most influential meteorological factors, such as the static stability index and mixed layer height.
- Data Smoothing (EWA): To mitigate the volatility of daily weather data, Exponentially Weighted Averages are applied with a weight value ($\beta$) of 0.9. This technique prioritizes recent data while maintaining a 10-day historical context, effectively smoothing the input and highlighting long-term trends.
- Ensemble Prediction (GBDT): The smoothed data is processed by a Gradient Boosting Decision Tree. GBDT builds an ensemble of decision trees sequentially, with each subsequent tree correcting the residual errors of its predecessors, making it highly effective for complex regression tasks.
Data and Preprocessing
The study utilized real-world datasets from Shijiazhuang and Xingtai, China, spanning from January 2014 to early 2017.
- Pollutant Data: Daily concentrations of $PM_{2.5}, PM_{10}, SO_2, NO_2, CO,$ and $O_3$.
- Meteorological Data: Daily weather features.
- Preprocessing: Missing values were handled using Lagrangian quadratic polynomials. Data was segmented by city and quarter to account for significant seasonal variations in air quality patterns.
Results and Discussion
The EWA-GBDT model demonstrated superior performance compared to standard SVM, Random Forest (RF), and basic GBDT models. In Xingtai, the model achieved an $R^2$ value of 0.9244 and a Root-Mean-Square Error (RMSE) of 13.6248.
Key insights from the analysis include:
- Feature Dominance: The static stability index emerged as a primary driver for $CO, NO_2, O_3,$ and $SO_2$ concentrations.
- Effectiveness of Smoothing: Applying EWA across all meteorological features significantly boosted performance, confirming that weather trends are more predictive than isolated daily snapshots.
- Seasonal Variability: Accuracy was highest in spring and summer (Q2 and Q3). Performance decreases in autumn and winter (Q1 and Q4) due to external factors like coal heating emissions, which are not captured by meteorological data alone.
Conclusion
EWA-GBDT provides a robust framework for air quality forecasting by effectively managing data noise and identifying critical meteorological drivers. Future work will focus on integrating formal time-series analysis to better handle seasonal instability and improving spatial granularity to enable neighborhood-level predictions.
Summary updated