An enhanced machine learning model for urban air quality forecasting under intense human activities

Abstract

With industrialization and urbanization, intense human activities are intensifying the complexity and dynamics of air quality variations, presenting significant challenges to prediction efforts. In this research, an enhanced machine learning model was proposed for forecasting urban air quality, based on integrating data de-noising, an optimized decomposition method, and error adaptive reduction into a hybrid framework. The Complete Ensemble Empirical Mode Decomposition with Adaptive Noise and Threshold (CEEMDANT) was developed by introducing statistical methods into the decomposition process, which enhanced the capability of extracting abrupt variations. At the same time, an error adaptive reduction strategy was designed to enhance the model's robustness and minimize forecasting risks. The model was demonstrated through a real-world case study of air quality prediction in four megacities of China, including Beijing, Shanghai, Guangzhou, and Shenzhen. The results indicated that CEEMDANT decreased the loss ratio of valid information by 16.01 %. Compared to traditional hybrid models, the error adaptive reduction strategy improved forecasting accuracy and stability by 9.96 % and 3.41 %, respectively. The proposed model provided precise benchmarks for residents to avoid health risks.