AFSA-optimized VMD-TCN-LSTM hybrid model for forecasting urban air pollutants: a case study of Taiyuan

With the rapid industrialization and urbanization, air pollution has become a global environmental challenge, and accurate prediction of air pollutant concentrations is crucial for public health protection and policy formulation. Conventional numerical strategies often suffer from prohibitive computational costs and data dependency, while statistical and singular deep learning models frequently fail to resolve the stochastic, nonlinear complexities inherent in air quality data. To address these deficiencies, this study introduces a hybrid VMD-TCN-LSTM framework optimized by the Artificial Fish Swarm Algorithm (AFSA), applied to air quality prediction in Taiyuan. Uniquely, AFSA adaptively tunes the Variational Mode Decomposition (VMD) parameters, specifically the mode number (K) and penalty factor (α), to optimizing the decomposed components for subsequent processing by Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) layers. Validation against a five-year dataset comprising pollutant and meteorological variables reveals that the proposed architecture significantly outperforms standalone LSTM, TCN, and unoptimized TCN-LSTM models across R², RMSE and MAE metrics for PM_2.5 and PM₁₀. The model exhibits enhanced adaptability and predictive performance within the study scope, offering a technical foundation for urban pollution control strategies and presenting a feasible innovation in forecasting methodology.