Abstract
To accurately predict PM2.5 concentrations in Nanchang City, this study proposes a hybrid model combining Support Vector Regression (SVR) with a modified Whale Optimization Algorithm (mWOA). Feature selection was performed using Pearson correlation coefficients, identifying PM10, SO2, CO, temperature, and wind power as critical inputs. The mWOA was employed to optimize SVR hyperparameters, effectively avoiding local optima. Results demonstrate that the mWOA-SVR model, particularly when incorporating meteorological factors, significantly outperforms traditional single-pollutant models in forecasting accuracy.
Introduction
Fine particulate matter (PM2.5) is a critical air quality indicator with severe implications for public health. Due to the non-linear and complex nature of atmospheric chemistry, predicting PM2.5 levels remains a significant challenge. Machine learning techniques like Support Vector Regression (SVR) have shown promise, but their performance is highly sensitive to hyperparameter settings. This paper addresses the "Goldilocks" problem of SVR tuning by introducing a modified Whale Optimization Algorithm (mWOA) designed to find global optimal parameters more reliably than standard metaheuristic approaches.
Methodology
Feature Selection and Data Acquisition
The study focused on Nanchang City, China, utilizing a comprehensive dataset from 2022. Air pollutant data (PM2.5, PM10, NO2, SO2, CO, and O3) were collected from monitoring sites, while meteorological data (temperature and wind power) were sourced from the ECMWF. Pearson Correlation Coefficient (PCC) analysis revealed that PM10, SO2, and CO had the strongest relationships with PM2.5. Additionally, weather factors like temperature and wind power were found to be essential for capturing the seasonal and daily variability of particulate matter.
The mWOA-SVR Model
The proposed mWOA-SVR model enhances the standard Whale Optimization Algorithm to prevent premature convergence. The modification includes a multi-population strategy that divides the search space into bounding and searching populations. This allows the algorithm to simultaneously exploit known high-performing regions and explore new areas of the parameter space.
The algorithm alternates its update logic based on the iteration count (odd vs. even) and incorporates reverse learning strategies, ensuring a dynamic search process that can "jump out" of local traps. This optimized algorithm is then used to determine the ideal penalty factor ($C$), loss function ($\epsilon$), and kernel width ($\gamma$) for the SVR model.
Results and Discussion
The performance of the mWOA-SVR model was evaluated using Mean Absolute Percentage Error (MAPE) and the coefficient of determination (R²).
- Meteorological Synergy: A key insight was that models using both pollutants and weather data achieved much higher accuracy. Weather factors act as multipliers or dispersants for pollutants, making them indispensable for accurate forecasting.
- PM10 Correlation: The extremely high correlation (>0.8) between PM10 and PM2.5 confirms their shared origins and similar physical behavior in the Nanchang atmosphere.
- Predictive Precision: The mWOA-SVR model successfully tracked the volatile "peaks and valleys" of PM2.5 concentrations, providing a much tighter fit to real-world observations than models using manual or standard optimization.
Conclusion
The mWOA-SVR hybrid model offers a significant advancement in air quality forecasting. By automating the complex task of hyperparameter tuning and integrating diverse environmental features, it provides a more reliable tool for environmental management. Future research will focus on incorporating spatial data and reducing the computational overhead associated with the multi-population optimization strategy.
Summary updated