基于PhyDNet-ATT的能见度预报方法

Visibility Forecast Based on PhyDNet-ATT Deep Learning Algorithm

本文采用PhyDNet-ATT深度学习方法建立江苏省能见度预报模型PhyDNet-ATT-VIS,该模型融合了高时空分辨率地面观测数据和模式产品,实现了空间分辨率为3 km、时间分辨率为1 h、预报时效为6~18 h的能见度预报,并且对模型结果进行了检验评估。与ECMWF能见度产品相比,PhyDNet-ATT-VIS预报的均方根误差和平均绝对偏差分别降低201%和310%;对于不同能见度等级,命中率显著提高,空报率显著降低,TS评分显示预报技巧优势明显,但15~18 h低能见度预报仍存在很大提升空间。PhyDNet-ATT-VIS在观测站点密集区域的预报误差显著低于观测站点稀疏区域。在区域性雾过程和局地雾过程预报方面,PhyDNet-ATT-VIS均能较准确地预报雾的落区、强度、生消等关键特征参数。研究为能见度短时临近预报技巧的提升提供了新思路。

NWP(numerical weather prediction)and statistical methods still have limitations in forecasting low visibility.Therefore,enhancing nowcasting techniques is crucial for ensuring the safety of daily life and industrial activities.A short-term visibility forecast model in Jiangsu Province PhyDNet-ATT-VIS is established based on PhyDNet-ATT(physical dynamics network with attention)deep learning method,using high spatial and temporal resolution ground observations and NWP data.Ground observations contain 1 min average horizontal visibility from automatic weather stations in Jiangsu Province,as well as its upstream regions of Anhui Province,Henan Province,and Shandong Province.NWP data are derived from PWAFS(Precision Weather Analysis and Forecast System)gridded meteorological forecast data.Temporal resolution and spatial resolutions of PWAFS forecast products are 1 h and 3 km,respectively.Due to PhyDNet-ATT-VIS’s outstanding ability to handle nonlinear problems,visibility forecast from 6 h to 18 h with a spatial resolution of 3 km and a temporal resolution of 1 h is achieved,and model results are tested and evaluated.The visibility forecasting product of ECMWF(European Centre for Medium-Range Weather Forecasts)is selected to evaluate and compare the forecasting skills with that of PhyDNet-ATT-VIS.The initial forecast time of ECMWF is 0800 BT and 2000 BT every day,with the time interval of 3 h and the spatial resolution of 0.125°×0.125°.Compared to ECMWF visibility product,the rootmean square error(RMSE)and mean absolute error(MAE)of PhyDNet-ATT-VIS are reduced by 201%and 310%,respectively.Across different visibility levels,the probability of detection(POD)significantly improves,while the false alarm ratio(FAR)substantially decreases.Threat score(TS)of the forecast demonstrates a clear advantage,although the model’s ability to predict low visibility(defined as visibility less than 0.2 km)in 15 h to 18 h range still requires further improvement.In terms of spatial distribution,the forecast error for visibility in lake and coastal areas is significantly lower than that in other regions for NWP.For PhyDNet-ATT-VIS,the error in areas with dense observation sites is significantly lower than that in regions with sparse observation sites.Compared to ECMWF model,PhyDNet-ATT-VIS can more accurately predict key characteristic parameters,such as affected areas,intensity,onset,and dissipation,in both regional and local fog processes.This study can provide a reliable and operationally replicable reference for improving short-term forecasting skills in visibility.