Scholarlink

<p>This study provides a detailed evaluation of eleven machine learning regression models—linear models (like linear regression, ridge, lasso, and elastic net), nonlinear methods (such as support vector regressor and K nearest neighbors), and ensemble techniques (including decision tree, random forest, gradient boosting, AdaBoost, and XGBoost)—to predict the daily Air Quality Index (AQI) in Hapur, Uttar Pradesh, for 2024. Daily averages of PM₂.₅, PM₁₀, NO₂, SO₂, CO, and O₃ were examined to calculate the AQI and assess the contributions of pollutants. Monthly, seasonal, and annual analyses indicated significant winter surges, with pollution levels markedly beyond WHO and CPCB guidelines. The data was split into two parts: 80% for training the models and 20% for testing them, and all the features were adjusted to a standard scale before training, while the best settings for the models were found using a method that checks different options to reduce errors. The evaluation of model performance was conducted utilizing R², RMSE, MSE, MAE, and computational duration. The Random Forest Regressor demonstrated superior generalization (test R² = 0.9987, RMSE = 3.25), with Gradient Boosting and XGBoost closely trailing. In contrast, SVR and K Nearest Neighbors exhibited subpar performance, while linear models reached a plateau (test R² ≈ 0.96). A detailed sensitivity analysis of the Random Forest model—by removing one, two, or three of the six predictors—showed that PM₂.₅ was the most important factor (removing it dropped R² to 0.7374 and raised RMSE to 44.74), with PM₁₀ being the next most important (R² = 0.9397, RMSE = 21.43), while the gaseous species were mostly unnecessary (∆R² &lt; 0.001, ∆RMSE &lt; 1 when removed). SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs) confirmed that PM₂.₅ and PM₁₀ are the main factors affecting the Air Quality Index (AQI), while CO and NO₂ are less important. The research illustrates that tree-based models facilitate precise and transparent air quality forecasts, while dependence on fewer particulate-based inputs decreases monitoring expenses and computing demands without sacrificing accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative benchmarking of eleven machine learning regressors for daily air quality index (AQI) forecasting: Sensitivity analysis and SHapley Additive exPlanations (SHAP)-Based interpretability framework

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Asif Ansari,
  • Abdur Rahman Quaff

摘要

This study provides a detailed evaluation of eleven machine learning regression models—linear models (like linear regression, ridge, lasso, and elastic net), nonlinear methods (such as support vector regressor and K nearest neighbors), and ensemble techniques (including decision tree, random forest, gradient boosting, AdaBoost, and XGBoost)—to predict the daily Air Quality Index (AQI) in Hapur, Uttar Pradesh, for 2024. Daily averages of PM₂.₅, PM₁₀, NO₂, SO₂, CO, and O₃ were examined to calculate the AQI and assess the contributions of pollutants. Monthly, seasonal, and annual analyses indicated significant winter surges, with pollution levels markedly beyond WHO and CPCB guidelines. The data was split into two parts: 80% for training the models and 20% for testing them, and all the features were adjusted to a standard scale before training, while the best settings for the models were found using a method that checks different options to reduce errors. The evaluation of model performance was conducted utilizing R², RMSE, MSE, MAE, and computational duration. The Random Forest Regressor demonstrated superior generalization (test R² = 0.9987, RMSE = 3.25), with Gradient Boosting and XGBoost closely trailing. In contrast, SVR and K Nearest Neighbors exhibited subpar performance, while linear models reached a plateau (test R² ≈ 0.96). A detailed sensitivity analysis of the Random Forest model—by removing one, two, or three of the six predictors—showed that PM₂.₅ was the most important factor (removing it dropped R² to 0.7374 and raised RMSE to 44.74), with PM₁₀ being the next most important (R² = 0.9397, RMSE = 21.43), while the gaseous species were mostly unnecessary (∆R² < 0.001, ∆RMSE < 1 when removed). SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs) confirmed that PM₂.₅ and PM₁₀ are the main factors affecting the Air Quality Index (AQI), while CO and NO₂ are less important. The research illustrates that tree-based models facilitate precise and transparent air quality forecasts, while dependence on fewer particulate-based inputs decreases monitoring expenses and computing demands without sacrificing accuracy.