Enhancing Ozone Forecasting Precision: A CEEMDAN-IPSOSE-SVR Approach with Multi-Source Data Integration

Abstract

Accurate forecasting of ground-level ozone (O3) concentrations is critical for mitigating public health risks and guiding air quality management, yet remains challenging due to the pollutant’s nonlinear dynamics, meteorological dependencies, and data noise. This study proposes a hybrid model, CEEMDAN_IPSOSE_SVR, integrating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Improved Particle Swarm Optimization Selective Ensemble (IPSOSE), and Support Vector Regression (SVR), to enhance O3 prediction accuracy. Leveraging multi-source data -- including WRF-CMAQ forecasts and real-time observations from Monitoring Station A -- the framework addresses noise reduction, feature selection, and model optimization. CEEMDAN decomposes raw O3 time-series into interpretable intrinsic mode functions (IMFs), effectively isolating high-frequency noise while preserving critical trends. IPSOSE optimizes SVR hyperparameters and selectively integrates base learners, balancing exploration and exploitation to mitigate overfitting. Experimental results demonstrate the model’s superiority, achieving reductions of 18–24% in MAE and RMSE compared to benchmarks like FNN and standalone SVR. The hybrid framework also attains an R² score of 0.92 for daily maximum 8-hour O3 predictions, outperforming conventional methods. Key contributions include methodological innovation in noise-robust forecasting, practical validation through real-world data, and insights into meteorological drivers of O3 variability. This study advances air quality modeling by synergizing physical and data-driven approaches, offering policymakers a reliable tool for proactive pollution control. Future work may extend the framework to multi-pollutant forecasting and optimize computational efficiency for real-time applications.