Modeling PM2.5 urbane pollution using hybrid models incorporating decomposition and multiple factors

Abstract

PM_2.5 negatively impacts air quality, human health, and the environment. Modeling PM_2.5 concentrations is helpful for understanding pollution dynamics and supporting government emergency responses and preventive measures. This study introduces a novel method to develop hybrid models that enhance PM_2.5 concentration modeling and evaluate source contributions. We applied empirical mode decomposition (EMD)-based models—EMD-LSTM, EMD-Bi-LSTM, EMD-GRU, EMD-CNN, and EMD-CNN-LSTM— to model hourly PM_2.5 concentrations using a 4-year dataset. PM_2.5 concentration data from the target and nine neighboring stations, combined with EMD and time lag functions, as well as other air pollutants and meteorological inputs, were used to develop models. We adopted a Shapley additive explanations analyzer-based LSTM model to identify pivotal features. Among all models, EMD-Bi-LSTM emerged as the top performer, achieving up to 89.5 % model accuracy (R²). PM_2.5 concentration at the target station from the previous 1 h was identified as a key contributor in the model. Other influencing factors included PM_2.5 concentrations at neighboring stations, PM₁₀, CO, O₃, total hydrocarbon compounds, and wind direction.