Abstract:

With the acceleration of global urbanization, the severe PM_2.5 pollution in arid zone cities, owing to their unique geographical and climatic conditions, exhibits strong non-stationarity and complex spatiotemporal characteristics, making it difficult for traditional prediction models to effectively capture its dynamic patterns. To address this challenge, this study proposes a hybrid prediction framework of an “adaptive noise complete ensemble empirical mode decomposition-Ivy optimization algorithm-Kolmogorov Arnold network-bidirectional long short-term memory neural network” (CEEMDAN-IVY-KAN-BiLSTM), aiming to enhance the prediction accuracy of PM_2.5 concentrations. This framework jointly extracts multi-scale features through noise reduction decomposition as well as parameter optimization and integrates the strong nonlinear fitting and bidirectional time series modeling capabilities of the KAN-BiLSTM model, effectively improving the prediction performance. The results reveal that the PM_2.5 concentration in Urumqi City from 2021 to 2024 shows significant seasonal fluctuations, with an average of 41.97 μg·m^-3 in winter due to coal heating and the influence of the inversion layer and drops to 14.04 μg·m^-3 in summer due to enhanced atmospheric convection. Moreover, it shows an overall decreasing trend annually. Moreover, the importance ranking of the data indicates that PM_2.5 is significantly positively correlated with air quality index, PM₁₀, CO, and NO₂, and negatively correlated with temperature and dew point temperature, suggesting that coal emissions, vehicle exhaust, and meteorological diffusion conditions are the main influencing factors. Moreover, the model effectively separates the high-frequency fluctuations (such as sandstorm events) and low-frequency trends (seasonal changes) in the PM_2.5 sequence, reducing the impact of data non-stationarity. Finally, the experiments were based on daily air quality data in Urumqi City from 2021 to 2024, results of which demonstrate that this model achieves the coefficient of determination, mean absolute error, and root mean square error values of 0.991, 1.391, and 1.881, respectively, significantly outperforming conventional machine learning and common deep learning models. This verifies the applicability of the “decomposition-optimization-integration” deep learning framework in the prediction of arid zone cities.

Key words: PM_2.5 concentration prediction, CEEMDAN decomposition, IVY optimisation algorithm, KAN-BiLSTM model, deep learning, urban PM_2.5 concentration in arid zones