Air Quality Prediction in Shanghai: Evaluating Traditional and Machine Learning Models for PM2.5 and PM10

Air pollution, especially fine particulate matter (PM2.5 and PM10), poses severe risks to public health and urban sustainability. Accurate forecasting is essential for early warning and effective policy-making. This study compares four models for predicting winter air quality in Shanghai: Linear Regression (LR), Seasonal Autoregressive Integrated Moving Average with Exogenous Variables (SARIMAX), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). The dataset (20142021) includes hourly pollutant concentrations and meteorological variables such as temperature, humidity, wind speed, and precipitation. After preprocessing and feature selection, models were implemented in R and evaluated using R, MAE, RMSE, and MAPE. Results indicate that deep learning models significantly outperform traditional statistical approaches. LSTM achieved the best performance (R = 0.83, RMSE = 16.3 g/m), followed by GRU, which offered comparable accuracy with lower computational demand. SARIMAX captured seasonal patterns but underestimated pollution peaks, while Linear Regression performed the weakest overall. In conclusion, deep learning methods, particularly LSTM and GRU, provide more accurate and robust forecasts of PM2.5 and PM10 in Shanghai. These findings highlight the potential of advanced machine learning techniques to support air quality management and public health protection.