Deep Learning Method for Urban Air Pollution Prediction: Empirical Study Based on PM2.5

A systematic comparison was conducted between traditional ARIMA models and deep learning Long Short-Term Memory (LSTM) models in predicting PM₂.₅ in the Shanghai air quality dataset from 2014 to 2025. The first step of the research is to perform missing value imputation, outlier detection, and standardization on the original dataset. Afterwards, analyze the temporal variation characteristics and correlation of the main pollutants. Then, create and train Autoregressive Integrated Moving Average Model (ARIMA) and LSTM models and perform model assessment employing MSE, MAE, and R². The paper will compare ARIMA with LSTM for prediction results. The results showed that LSTM reduced MSE by nearly 56% and MAE by about 31%. The paper can say that LSTM is more versatile than ARIMA in capturing nonlinear features and handling fast fluctuations. This study provides a methodology for predicting urban air pollution in meteorology and lays the foundation for building more complex prediction systems.