Machine Learning for Real-Time Urban Air Quality Forecasting Model Performance Comparison and Field Deployment Framework

Abstract

Urban air pollution has become a critical global public health threat, with real-time and accurate air quality forecasting serving as the core foundation for pollution prevention, environmental management, and public health early warning. Traditional numerical prediction models suffer from high computational cost and poor real-time performance, while classical statistical models fail to capture the non-linear and non-stationary characteristics of air quality time series. To address these gaps, this study systematically compares the performance of four mainstream models (Seasonal Auto-Regressive Integrated Moving Average, Random Forest, Long Short-Term Memory, and Time Series Transformer) for multi-time-scale (1h, 24h, 72h) and multi-pollutant forecasting. We use hourly air quality monitoring data and synchronous meteorological data from 6 national monitoring stations in the Beijing-Tianjin-Hebei region of China from 2019 to 2023, covering six core pollutants including PM2.5, PM10, NO2, and O3. Results show that the Time Series Transformer (TST) model achieves the optimal performance across all forecasting horizons and pollutants: its 1h PM2.5 forecasting MAE is 61.6% lower than the SARIMA benchmark and 18.7% lower than the LSTM model, and it maintains an R² of 0.81 even for 72h long-term forecasting. Furthermore, we propose a lightweight edge deployment framework, which compresses the model size by 65.2% through quantization and pruning, achieving a single-sample inference latency of only 427ms on edge devices with less than 1% accuracy loss. This study provides a complete technical solution for real-time urban air quality forecasting, with significant theoretical value and practical application potential for urban environmental governance and public health protection