An Enhanced Air Quality Forecasting Method Integrating Feature Mode Decomposition and Multilayer Plasticity Echo State Network

Air quality prediction plays a pivotal role in environmental science, particularly in mitigating the impacts of rapid urbanization and industrialization. Conventional approaches, including both mechanism-based and data-driven models, often struggle to cope with the pronounced non-stationarity, spatial heterogeneity, and intricate pollutant interactions inherent in atmospheric time series. To address these challenges, we propose a Feature Mode Decomposition-based Multi-layer Plasticity Echo State Network (FMD-MPESN), a novel architecture that integrates advanced environmental data preprocessing with adaptive neural dynamics. Specifically, feature mode decomposition (FMD) is employed to disentangle raw pollutant series into interpretable frequency components, enabling the separation of long-term trends from short-term fluctuations. These components are then processed by an echo state network enhanced with multilayer neuro-plasticity, allowing dynamic adjustment of internal connectivity and hierarchical feature extraction to better capture evolving environmental states. Extensive experiments on China and International air quality datasets demonstrate that FMD-MPESN consistently outperforms competitive deep learning and ESN-based baselines, achieving up to 11.7% lower RMSE and 12.4% lower MAPE in multi-step forecasting and exhibiting stronger robustness and generalization across different regional datasets. These results highlight the potential of FMD-MPESN as a scalable decision-support tool for environmental monitoring, public health protection, and sustainable urban management.