Abstract: Objective　PM2.5 pollution is a prominent environmental issue in urban areas of China, and accurate prediction of PM2.5 concentration is the key to formulating scientific air pollution control measures. Traditional PM2.5 prediction methods mostly rely on single meteorological or atmospheric pollutant data, which have the limitations of low prediction accuracy and inability to capture complex coupling relationships between multiple influencing factors. In view of the close correlation between atmospheric water vapor, atmospheric pollutants, meteorological elements and PM2.5 formation, this study takes Xi'an City as the research area, and by integrating GNSS precipitable water vapor (GNSS-PWV) retrieved from groundbased GNSS observation data, atmospheric pollutant data and meteorological element data, constructs high-precision PM2.5 801 concentration prediction models to realize accurate prediction of hourly scale PM2.5 concentration in haze weather and provide technical support for urban PM2.5 pollution monitoring and early warning. Methods　Firstly, the GNSS observation data of Xi'an city, China, in January 2021 were processed by GAMIT software, and combined with long baseline IGS station data, high-precision GNSS-PWV was retrieved. The accuracy of retrieved water vapor was verified using radiosonde data. Then, the atmospheric pollutant data (AQI, PM10, SO2, NO2, O3, CO), ERA5 meteorological element data (temperature, air pressure, dew point temperature, wind speed, etc.) and GNSS-PWV were collected as the research data set. The correlation characteristics between PM2.5 concentration and various influencing factors were analyzed at daily and hourly scales respectively, and the significance tests of correlation coefficients were carried out. On this basis, two deep learning models were constructed for PM2.5 concentration prediction: the traditional LSTM model and the T-LSTM model combining Transformer self-attention mechanism with LSTM. The multi-source fusion data were normalized to eliminate the influence of magnitude deviation, then the data sets were divided into training sets and test sets at a ratio of 4:1, and the model parameters were optimized according to the data scale and computer performance. Root mean square error (ERMS), mean absolute error (EMA) and correlation coefficient (R²) were used as evaluation indexes to compare and analyze the prediction performance of the two models. Results and Discussion　The correlation analysis showed that there were obvious differences in the correlation between PM2.5 and each influencing factor at different time scales. At the daily scale, PM2.5 was significantly positively correlated with AQI, PM10, NO2, CO, dew point temperature, vapor pressure, relative humidity and GNSS-PWV, and weakly negatively correlated with O3. However, the correlation of PM2.5 with SO2, surface temperature, air pressure and wind speed failed the significance test. At the hourly scale (for the three days of severe haze in January 2021), PM2.5 was significantly correlated with all studied influencing factors, with positive correlation with AQI, PM10, O3, CO, most meteorological elements, and GNSS-PWV, and negative correlation with SO2, NO2 and air pressure. This indicated that the hourly scale data could more fully reflect the real-time response relationship between PM2.5 and various factors in haze weather. The model prediction results showed that both LSTM and T-LSTM models had good prediction performance for PM2.5 concentration. The LSTM model achieved R² of 0.896, ERMS of 7.887 μg/m³ and EMA of 6.913 μg/m³. While the T-LSTM model, combined with selfattention mechanism, further improved the prediction accuracy, with R² increasing to 0.981 (an increase of 9.5% compared with LSTM), ERMS decreasing to 4.311 μg/m³ and EMA decreasing to 3.644 μg/m³. The introduction of self-attention mechanism enables the T-LSTM model to capture the long-term dependence of PM2.5 concentration time series and the global correlation between multiple influencing factors, while retaining the advantage of LSTM in mining fine characteristics of short-term time series, thus achieving better simulation results. In addition, the positive correlation between GNSS-PWV and PM2.5 at the hourly scale (R = 0.876) indicated that the hygroscopic growth of PM2.5 particles under high water vapor conditions was an important reason for the increase of PM2.5 concentration in haze weather of Xi'an. Conclusions　This study realized the effective fusion of GNSS-PWV, atmospheric pollutants and meteorological element data, and clarified the time-scale difference of the correlation between various factors and PM2.5 concentration in Xi'an city. Both the constructed LSTM and T-LSTM models can effectively predict hourly scale PM2.5 concentrations in haze weather. The T-LSTM model, due to the introduction of self-attention mechanism, has higher prediction accuracy and better simulation effect, which provides a more accurate technical method for PM2.5 pollution monitoring and early warning in Xi'an. The research results confirm the application value of GNSS water vapor data in PM2.5 prediction and provide a new idea for multi-source data fusion in the field of atmospheric pollution prediction. The limitation of this study is that the research time series is relatively short (only January 2021). Subsequent research can expand the data time scale and combine with seasonal characteristics to further verify the stability and universality of the model and optimize the model structure to improve the prediction ability under different pollution conditions.

Key words: PM2.5, Global navigation satellite system water vapor, atmospheric pollutant, meteorological elements, long short-term memory (LSTM), T-LSTM