Research on spatiotemporal sequence prediction of air quality parameters in Hefei area based on multi-model comparative analysis

This study is based on the hourly PM2.5 monitoring data of Hefei from January 2015 to March 2025 provided by the National Urban Air Quality Real - time Release Platform. Aiming at the problem of spatiotemporal sequence prediction of air quality parameters, the performances of four models, namely ARIMA, LSTM, Prophet, and Random Forest, were systematically compared. Through data pre - processing methods such as forward - filling of missing values, construction of time features (hour, week, month), and sliding window division (for LSTM), the difficulties of data fragmentation and non-linear feature extraction were solved. Experiments show that the LSTM model performs best in long - term dependence modeling (R²=0.98). The Prophet model has outstanding decomposition capabilities and can clearly present annual and seasonal variation characteristics. The Random Forest model has an R² of 0.97, good prediction performance, and a training time of only 24.28 seconds, with high efficiency. However, the ARIMA model has an R² of - 0.12, with a large long - term prediction deviation, but it can quickly respond to short - term mutations and has a low computational cost, making it suitable for short - term (1 - 3 days) prediction. The research results provide methodological support for multi - scale prediction of regional air quality and offer scenario - based suggestions for model selection.