Decomposition and forecasting of concentrations of highly volatile atmospheric pollutants in plateau cities based on the VMD-CNN-BiLSTM framework

In recent years, multiple factors including climate change, accelerated industrialization, and rising population density have led to severe air pollution in numerous cities worldwide. For cities situated in plateau regions, complex topography and unique climatic characteristics further exacerbate the dispersion and accumulation of PM_2.5 and PM₁₀. This results in highly nonlinear and uncertain air quality data, posing significant challenges for accurate pollution concentration forecasting. Therefore, this paper proposes a “Decomposition-Prediction-Integration” framework. This method first employs variational modal decomposition to adaptively decompose raw, chaotic pollutant sequences into a series of more stable, physically meaningful intrinsic modal functions, effectively separating noise from fluctuating trends across different timescales. Subsequently, a CNN-BiLSTM model is constructed: CNN extracts local fine-grained features from multiple IMFs, while BiLSTM learns their long-term bidirectional dependencies. Experimental validation across four monitoring stations in Kunming demonstrates that this framework significantly outperforms eight baseline models in PM_2.5 forecasting. This study not only establishes a new paradigm for air quality prediction in complex topography but, more importantly, identifies and thoroughly analyzes the distinct response mechanisms exhibited by the same decomposition-prediction framework for PM_2.5 and PM₁₀. This reveals the critical importance of model adaptation for different particle sizes in plateau cities.