[PM_2.5 Prediction Based on EOF Decomposition and CNN-LSTM Neural Network]

Based on the surface meteorological data and ambient air quality data of Taiyuan from 2016 to 2020, the temporal and spatial variation characteristics of PM_2.5 concentration in Taiyuan were analyzed. The temporal and spatial variation characteristics of PM_2.5 concentration in Taiyuan were studied using the EOF decomposition diagnostic analysis method. At the same time, the importance of meteorological factors was analyzed using a random forest model, and a PM_2.5 concentration prediction model based on the CNN-LSTM neural network was established. The results showed that from 2016 to 2020, the annual mean PM_2.5 concentration in the urban area of Taiyuan generally exhibited a decreasing trend from year to year, and the high value mainly appeared in November, December, January, and February. From 18：00 to 02：00 of the next day, the peak value of PM_2.5 concentration was easily reached, and the annual average value of PM_2.5 concentration gradually increased from northwest to southeast. The EOF decomposition of PM_2.5 concentration was as follows： the variance contribution rate of modal 1 eigenvector was 49.4%, and the variance contribution rate of modal 2 eigenvector was 30.8%. Considering Nanzhai-Julun-Jinyuan as the boundary, it was a positive area to the northwest and a negative area to the southeast. The positive center appeared in Jinsheng district, and the negative center appeared in Xiaodian in the southeast. PM_2.5 concentration was positively correlated with relative humidity and dew point temperature. Moreover, it was mainly negatively correlated with wind speed, precipitation, and mixing layer height and generally negatively correlated with ventilation and self-purification capacity, with no significant correlations involving temperature. Relative humidity, dew point temperature, air pressure, humidity, and mixing layer height all played an important role in the ranking of the four seasonal characteristics, followed by wind speed, wind direction, ventilation volume, and self-purification capacity. Using the CNN-LSTM model for modeling, the R² of PM_2.5 concentration prediction was 0.805, 0.826, 0.897, and 0.901 in spring, summer, autumn, and winter, respectively. R² was above 0.8 in all four seasons. The predicted residuals of the CNN-LSTM model in all four seasons were approximately normally distributed, and the absolute error of the model was controlled within 10 μg·m^-3. The prediction results below 10 μg·m^-3 reached a maximum of 81.2% in summer, followed by 75.9% and 62.9% in autumn and spring, respectively. The performance in winter was average, with 51.5% of the prediction results having an absolute error below 10 μg·m^-3.