Page 298 - 《高原气象》2026年第3期
P. 298

高     原      气     象                                 45 卷
              906





                     Visibility Classification Forecasting and Interpretability Analysis in

                              the Guanzhong Region of Shannxi Province Based on
                                                 LightGBM Algorithm



                                               1
                                    HUANG Lei , DA Xuanfang , BAI Qingmei , LIU Ruifang 1
                                                             2
                                                                          1
                                (1. Xi’an Meteorological Bureau of Shaanxi Province, Xi’an  710016, Shaanxi, China;
                                 2. Gansu Provincial Meteorological Service Center, Lanzhou  730010, Gansu, China)
             Abstract: The characteristics of precipitation and cloud of two transverse shear line low visibility (VIS) weath‐
             er, resulting from complex interactions among atmospheric boundary layer processes and human activities, pos‐
             ing  significant  forecasting  challenges  for  geographically  enclosed  and  pollution-prone  regions  like  the  Guan‐
             zhong Basin. Based on 2023 -2025 surface observations from 44 national stations in Shaanxi's Guanzhong re‐
             gion, ERA5 reanalysis data, and environmental monitoring data, this study establishes a visibility classification
             forecasting model using the LightGBM (Light Gradient Boosting Machine) algorithm. By inputting 3 h meteoro‐
             logical and pollutant concentration forecasts from ECMWF and CUACE models, it achieves objective 72 h fore‐
             cast lead time visibility classification forecasts for national stations in the Guanzhong region. The results of quan‐
             titative verification on eight low-visibility events from January to March 2025 shows that the LightGBM outper‐
             forms the ECMWF in all levels of visibility. For low-visibility conditions (VIS≤1 km), the LightGBM model's
             TS and POD are 0. 13 and 0. 16 higher than those of the ECMWF forecasts, respectively. It also exhibits superior
             performance in predicting nocturnal low visibility, with the TS score increasing by nearly 104% compared to the
             ECMWF within the 48 h forecast lead time. SHAP (Shapley Additive Explanations) analysis reveals that meteo‐
             rological factors of the lower atmosphere and PM  concentrations primarily drive model decisions of low visibil‐
                                                       2. 5
             ity prediction. Surface humidity thresholds (the difference of temperature and dew temperature less than 2 ℃;
             relative humidity less than 80%) serve as critical criteria. When 850 hPa temperatures surpass -5 ℃, the com‐
             bined effects of warm-moist advection and sustained temperature inversion significantly heighten the model's re‐
                                                                                -3
             sponsiveness to visibility decline. For PM  concentrations below 130 μg·m , meteorological factors dominate
                                                  2. 5
             visibility reduction; however, when PM  exceeds this threshold, aerosol extinction becomes the predominant
                                                 2. 5
             factor influencing visibility.
             Key words: visibility prediction; machine learning; LightGBM; SHAP
   293   294   295   296   297   298   299   300   301   302   303