What does this research mean for the field?

A three-stage stacked ensemble framework integrating LightGBM, a residual neural network, and Ridge regression enables highly accurate, leakage-free day-ahead PM10 forecasting across multiple countries, significantly outperforming traditional baseline models. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

January 5, 2026Discover Atmosphere10 citationsOpen Access

A hierarchical ensemble approach for multi-country PM10 forecasting using LightGBM and residual neural network

SISyed Azeem InamSindh Madressatul Islam University HRHaider RajputSindh Madressatul Islam University SUSaddam UmerSindh Madressatul Islam University

Key Points

Key points are not available for this paper at this time.

Abstract

Accurate day-ahead forecasting of particulate matter (PM10) concentrations is critical for public health interventions, regulatory compliance, and urban air quality management. However, existing approaches suffer from temporal leakage, single-city limitations, inadequate hierarchical modeling of geographic dependencies, and reliance on single-model architectures that fail to capture complex nonlinear pollution dynamics. This study presents a novel three-stage leakage-free stacked ensemble framework for city-level PM10 prediction across 25 countries and 380 cities using the World Air Quality Index (WAQI) dataset comprising 1,798,600 records. The framework integrates Light Gradient Boosting Machine (LightGBM) as the base learner, a residual neural network to model higher-order nonlinearities, and a Ridge regression meta-learner trained exclusively on out-of-fold predictions. Rigorous leakage prevention is achieved through chronological per-city train-test splits (70:30), leakage-safe target encoding of hierarchical city and country features fitted only on training data, and expanding-window rolling-origin cross-validation. The proposed model achieves exceptional performance on the held-out test set with R2 = 0.9983, RMSE = 1.01 µg/m3, MAE = 1.01 µg/m3, and NSE = 0.9983, representing 96.7% RMSE improvement over the persistence baseline and 76% improvement over XGBoost. Cross-validation results demonstrate consistent R2 values exceeding 0.99 across large sample countries and above 0.98 for smaller datasets. SHAP-based feature importance analysis confirms that lag-1 PM10, rolling means, and hierarchically encoded geographic features are the primary predictors. Temporal shuffle validation confirms the absence of information leakage. The framework significantly outperforms ARIMA, SARIMA, Random Forest, MLP, and traditional baselines while maintaining computational efficiency. This research establishes a reproducible, operationally viable, and theoretically rigorous benchmark for multi-country air quality forecasting with direct applications in public health policy and environmental monitoring systems.

Ask AI

Helpful

Bookmark

View Full Paper