Research on Persistent Low Visibility Processes along Rivers in Autumn and Winter Based on Deep Learning

To figure out the association between visibility and meteorological elements in addition to forecasting visibility through LSTM multivariate neural network model, this research work shortlisted long-term meteorological records of Zhangjiagang weather station in overdue November 2018. The prediction results were assessed by R2, MAE, RMSE and MSE. The analysis was carried out on low visibility processes. Key conclusions include: 1. 1. The event had an extensive duration, coverage, and intensity, with a minimum visibility of <50 m; and occurred during the mid-rainy season along the river. The occurrence and duration of the low-visibility process was affected by atmospheric composite pollutants, in which PM2. 5, PM10, and SO2 concentrations played important roles in maintaining and enhancing low visibility. 2. Graph 1: Key meteorological conditions with visibility highlights that there are significant relationships between meteorological conditions and visibility. The correlation coefficient between relative humidity and the temperature-dew point difference was higher than 0.8 during the low-visibility process. The correlation coefficient close to 0.8 between temperature and visibility was also positive. 2.2 Radar wind speed at 2 meters positively correlated with visibility less than 1000 meters and 500 meters occurred under use of below 1.6 m·s−1 wind speed. Moreover, a specific wind speed was ultimately responsible for the transport of near-surface water vapor and pollutants, which were essential for maintaining the low-visibility process. The atmospheric pressure was strongly and negatively correlated with visibility (correlation coefficient of -0.411), which explained the critical role of several key meteorological conditions (e.g. warm and humid near-surface air, low pressure and low wind speed) in the formation and maintenance of low-visibility weather. 3. Low-visibility events frequently appeared from midnight until the early morning, when the boundary layer height dropped considerably. Under noticeable water vapor advection into the river area and nighttime radiative cooling (clear sky condition) along with the evaporated water vapor of the Yangtze River,the near surface relative humidity tend to be saturation. Meanwhile, because of the inversion effect, the vertical mixing and dispersion of water vapor and pollutants were not easy, contributing to the persistence and escalation of the low-visibility process along the river. 4. The multivariable LSTM model's forecast fit (R2) was 0.985 for 12 hours-ahead prediction for short- to medium-term visibility forecasting, which surpassed that of single-variable LSTM prediction model in previous studies. The fit (R2) of the forecast was above 0.9 for 3 days’ forecasts. The prediction performance slightly decreases to 72 hours or less than 6 hours, while the fit (R2) is still at the level of 0.641 to 0.880, significantly higher than the traditional subjective and objective visibility forecasting methods.