-
To identify thunderstorm gusts in North China between 2021 and 2023, this study utilizes wind field information from automatic weather stations (AWSs), as well as data obtained from lightning locators and radars.
Each piece of thunderstorm gust data contains the following elements: AWS data, lightning data, radar data, RMAPS-RISE data, and rapidly updated high-resolution NWP products of every thunderstorm gust process. The AWS data, lightning data, and radar data are used to make the ground truth of thunderstorm gusts, and the radar data, RMAPS-RISE data, and rapidly updated high-resolution NWP products are used to produce the training, validation, and testing datasets, respectively. Among them, the AWS data come from more than 3000 national and regional surface AWSs in North China, where extreme wind speed in 1 h (WSX) is used. The lightning data are cloud-to-ground lightning data from the National Lightning Monitoring Network, which mainly contain two-dimensional spatial information (LOCATION) and the time of occurrence (TIME). North China is covered by the entire lightning monitoring network, the positioning accuracy is better than 300 m, the detection efficiency is greater than or equal to 80%, and the detection range of a single station is greater than or equal to 300 km (Zhou et al., 2021). The weather radar data are obtained from the composite reflectivity factor mosaic of eight S-band radars (Beijing radar, Cangzhou radar, Qinhuangdao radar, Shijiazhuang radar, Tianjin radar, Handan radar, Haituoshan radar, and Kangbao radar) and two C-band radars (Chengde radar and Zhangbei radar) in North China (Yang et al., 2023b) (Fig. 1).
Figure 1. The topographic characteristics of the study area and the location of radar stations (red circles represent the coverage of radar stations; color shading represents terrain height; units: m).
RMAPS-RISE was developed by the IUM. With a lead time of 24 h, the system is updated every 10 min and there is a 1-h forecast gap. The spatial resolution is 500 m. The fundamental principle of RMAPS-RISE is to improve upon NWP model output by utilizing the most recent surface observational and high-resolution topographical data. Through the assimilation of multi-source observation data, the error of RMAPS-RISE has been greatly reduced, especially in nowcasting and short-range forecasts (Cheng et al., 2019; Song et al., 2019; Yang et al., 2019, 2021; Chen et al., 2020a). Therefore, one of the main input datasets used in this study is from RMAPS-RISE. There are 1221 and 1521 grid points in the x- and y- directions, respectively. The dataset mainly includes a 10-m height analysis of current gusts (UVana), a 10-m height prediction of future gusts (UVpred_RISE), a 1-h cumulative precipitation forecast (RRpred), and the difference between the current and predicted 2-m surface temperature (TQdiff).
The high-resolution NWP product is derived from the Rapid-Refresh Multi-Scale Analysis and Prediction System for Short-Term weather (RMAPS-ST) (Chen et al., 2020b, 2021c; Feng et al., 2021), which is a fast-updating cyclic mesoscale numerical prediction system with a spatial grid spacing of 3 km. There are 320 and 265 grid points in the x- and y- directions, respectively. It mainly includes air pressure (PRES), average wind speed at 80-m height (UVpred_ST), CAPE, composite radar reflectivity of RMAPS-ST (RADARST), vertical wind shear at 0–1 km (SHEAR1) and 0–6 km (SHEAR2), and temperature difference between 850 hPa and 500 hPa (TMPdiff). The parameters used in this study are shown in Table 1.
Product Parameter (s) Specific meaning Unit AWS DATA WSX Extreme wind speed in 1 h m s−1 LIGHTNING DATA LOCATION Two-dimensional spatial information ° TIME Time of occurrence UTC RMAPS-RISE UVana 10-m height analysis of current gusts m s−1 UVpred_RISE 10-m height prediction of future gusts m s−1 RRpred 1-h cumulative precipitation forecast mm TQdiff Difference between current and predicted 2-m surface temperature °C RADAR DATA RADAR Composite radar reflectivity factor dBZ RMAPS-ST PRES Pressure Pa TMPdiff Temperature difference between 850 hPa and 500 hPa K CAPE Convective available potential energy J kg−1 RADARST Composite radar reflectivity dBZ UVpred_ST Average wind speed at 80-m height m s−1 SHEAR1 0–1 km vertical wind shear m s−1 SHEAR2 0–6 km vertical wind shear m s−1 Table 1. Description of physical quantity parameters used in this study.
The study area is located within 35.9°–42.7°N and 113.2°–120.2°E, roughly covering North China. As the input data resolution varies, we standardize all the data into a 500 m grid spacing using an interpolation method. The x- and y- directions in the input data contain 1221 and 1521 grid points, respectively.
-
Thunderstorm gusts are extreme strong convective weather events with limited samples (Zhou et al., 2019). Therefore, an under-sampling method is used to solve the problem of possible unbalanced proportions of positive and negative samples. The specific process for diagnosing the ground truth for thunderstorm gusts is illustrated in Fig. 2. In order to generate a suitable ground truth for model training, we construct the ground truth of thunderstorm gusts (i.e., locations where it is believed that thunderstorm gusts occurred) by utilizing radar data, lightning data, and AWS data, as well as specific selection criteria. Generally, when the reflectivity factor in radar data exceeds 35 dBZ, it is believed that strong convective weather has occurred. In order to enrich the sample data, the reflectivity factor threshold is defined as 30 dBZ. That is, only when the radar reflectivity factor corresponding to a grid point is higher than the threshold is that grid point defined as meeting the radar reflectivity criterion (Fig. 3a).
Figure 3. Five different grid types at 1400 UTC 31 July 2021: (a) grid points meeting the radar reflectivity criterion; (b) grid points meeting the lightning criterion; (c) grid points in the thunderstorm region; (d) grid points in the gust region; (e) ground truth.
Thunderstorms are often accompanied by lightning. Therefore, we focus on the exact location of lightning events. We extract latitude and longitude geographic coordinate information of all lightning events from 10 min before the current moment until now (i.e., if the current moment is 1400 UTC, we pick all lightning events that occurred between 1350 UTC and 1400 UTC). A circular area with a radius of 20 km is made with this coordinate as the center of the circle. All grid points located in this circular region are defined as meeting the lightning criterion (Fig. 3b).
Combining both the radar reflectivity criterion and lightning criterion at the same time, the thunderstorm region is determined. Inspired by the concept of the image connectivity domain in digital image processing, we propose the following approach: if a grid point is identified as meeting both the radar reflectivity criterion and lightning criterion, then all adjacent grid points meeting the radar reflectivity criterion are considered as the thunderstorm region (Fig. 3c).
The selection criteria for gust wind are based on the ground-based meteorological observation specifications compiled and issued by the CMA (Cui, 2011), which classify winds with instantaneous speeds of 17.2 m s−1 or higher (or visually estimated winds of magnitude 8 or higher) as high winds. Therefore, if the gust wind speed from AWS data is greater than or equal to 17.2 m s−1, a circular region with a radius of 2 km is made with the latitude and longitude coordinates of this specific AWS as the center. All grid points located in this circular region are considered as the gust region. (Fig. 3d).
If a grid point is in the thunderstorm region and the gust region, an intersection of the thunderstorm region and the gust region will be taken. If this intersection contains 50 or more grid points, we classify all thunderstorm regions that contain this portion as the ground truth for thunderstorm gusts (Fig. 3e). Taking the moment of 1400 UTC 31 July 2021 as an example, the intuitive meaning of the above five different grid types, explained in detail, can be seen in Fig. 3.
-
This paper selects 34 days on which thunderstorm gusts definitely occurred in North China from 2021 to 2023 during the summer season. The occurrence times of thunderstorm gusts can be found in Table 2. We split the dataset into a training and validation dataset and a testing dataset. The testing dataset consists of data from three days [20210731 (YYYYMMDD): 0800, 0900, 1000, 1100, 1200, 1300, 1400, 1500, 1600, 1700 UTC; 20220612: 1100, 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900 UTC; 20220818: 1300, 1400, 1500, 1600, 1700, 1800, 1900 UTC] on which thunderstorm gusts definitely occurred. The remaining 31 days constitute the data for training (90%) and validation (10%). We process all the data into the RMAPS-RISE format, which, along with the grid spacing, was detailed in section 2.1. In order to enrich the dataset and take into account the continuity of data splicing, the original 1521 × 1221-sized grid point graph is cropped into multiple 48 × 48-sized subgraphs, and some subgraphs located at the edge of the study are discarded (while they are also at the edge of North China). This is done by constructing a square of size 48 × 48 starting at 35.9°N and 113.2°E (lower left corner of the study area) and moving it in non-recombining order over the study area: (1) The square is moved in non-recombining order from left to right up to the right boundary of the study area; (2) All the squares produced by process (1) are moved in non-recombining order from bottom to top, up to the upper boundary of the study area; (3) Processes (1) and (2) produce a total of 775 squares. Therefore, each original graph is divided into 775 subgraphs by sequential cropping. The specific size of the whole dataset is shown in Table 3, and the process of enriching the dataset can be seen in Fig. 4.
Date (YYYYMMDD) Moments of occurrence (UTC) Number of true grid points 20210613 0500 0600 0700 0800 0900 1000 1100 1200 1300 1500 407 792 20210625 0800 1000 1100 1200 1300 1400 1500 1600 1700 392 061 20210626 0400 1100 1200 1300 1400 1500 1600 2000 93 081 20210629 1000 21 772 20210630 0500 0600 0700 0800 0900 1000 1100 1300 1400 182 532 20210701 0700 0800 0900 1000 1100 1200 1300 1400 590 772 20210702 0900 1000 1100 2100 2300 294 398 20210705 0300 0700 0800 0900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1 210 975 20210707 0600 0700 0800 0900 1000 1100 1200 1300 279 714 20210708 0800 0900 1000 1200 1400 2200 76 834 20210710 1600 1700 1800 1900 2000 288 313 20210711 0400 0500 0600 0700 0800 0900 1000 1100 1200 1300 1400 1500 1600 1700
1800 1900 2000 2100 2200 23003 026 709 20210712 0100 0200 0300 0400 0500 0600 0700 1200 1500 1600 904 921 20210731 0800 0900 1000 1100 1200 1300 1400 1500 1600 1700 553 150 20210805 0500 0600 0700 0800 0900 1000 1100 1200 1300 1400 1500 1600 163 705 20210808 1000 1100 1200 1300 1400 1700 1800 1900 2000 2100 2200 392 908 20210809 0900 1000 1100 1200 1300 44 978 20210816 1100 1200 1300 1400 1500 1600 65 088 20210823 0800 0900 1000 1100 1200 1300 1400 1500 531 181 20210907 0900 1000 1100 1200 1300 1400 78 512 20210912 0800 0900 1000 56 768 20210919 1400 1500 1600 161 755 20220524 0600 0700 0800 0900 1200 1300 1400 104 367 20220604 0900 1000 1100 1200 1300 1400 49 878 20220612 1100 1200 1300 1400 1500 1600 1700 1800 1900 626 609 20220726 0400 0600 13 233 20220804 0500 0600 0700 0800 0900 1000 1100 1200 1300 1400 1600 1700 1800 269 587 20220806 0400 1000 1100 1300 1400 69729 20220818 1300 1400 1500 1600 1700 1800 1900 289 698 20220904 1000 1100 1200 32 567 20230428 0700 0800 0900 1000 1100 1200 1300 1800 1900 2000 318 338 20230628 0000 0100 0200 0300 0400 0500 0600 0700 0800 391 332 20230724 0700 0800 0900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 650 974 20230820 0200 0400 0600 0700 0800 0900 1100 1200 1600 1700 1800 276 525 Table 2. Occurrence times of thunderstorm gusts, 2021–23.
Lead time Size of training set Size of validation set Size of testing set 1 h 78300 8700 5394 2 h 77064 8562 4845 3 h 75721 8413 4328 4 h 71862 7984 4058 5 h 68571 7618 3819 6 h 64494 7165 3636 Table 3. Statistics of the training, validation, and testing datasets for six different models in DL (corresponding to 1–6 h forecast timescales).
Figure 4. Process of enriching the datasets: (a) one 48 × 48-sized subgraph (each original graph has 775 subgraphs); (b) the coverage of all subgraphs in each original graph (sequential stitching of 775 subgraphs); (c) the edge of North China (discarded in this study). Color shading represents the terrain height (units: m).
-
The confusion matrix is introduced as an evaluation metric. In order to evaluate the performance of the prediction, the following metrics are used: the POD, false alarm ratio (FAR), CSI, and equitable threat score (ETS), where the values of POD, FAR and CSI are between 0 and 1, and the value of ETS is between negative 1/3 and 1. When the values of POD, CSI and ETS are closer to 1, and the value of FAR is closer to 0, the prediction effect of the model is better. POD, FAR, CSI and ETS are defined as
where TP indicates the number of actual thunderstorm gust samples that are correctly predicted as thunderstorm gusts, FN refers to the number of actual thunderstorm gust samples that are erroneously predicted as non-thunderstorm gusts, FP represents the number of actual non-thunderstorm gust samples that are mistakenly predicted as thunderstorm gusts, and TN signifies the number of actual non-thunderstorm gust samples that are accurately predicted as non-thunderstorm gusts. The confusion matrix can be seen in Table 4.
Ground truth Thunderstorm gusts Non-thunderstorm gusts Predicted results Thunderstorm gusts TP FP Non-thunderstorm gusts FN TN Table 4. Confusion matrix in this study.
-
The specific classification comparison results can be seen in Table 5. Forecasting thunderstorm gusts poses a slightly greater challenge than forecasting other types of strong convective weather. Therefore, our approach is to prioritize achieving higher values of CSI and ETS.
Lead time Models POD FAR CSI ETS 1 h RISEgust 0.150 0.518 0.129 0.121 U-net 0.420 0.468 0.307 0.266 CU-net 0.580 0.582 0.321 0.267 TransU-net 0.571 0.537 0.344 0.293 TG-TransUnet 0.588 0.512 0.364 0.314 2 h RISEgust 0.115 0.571 0.100 0.092 U-net 0.425 0.691 0.218 0.168 CU-net 0.417 0.664 0.229 0.181 TransU-net 0.557 0.693 0.247 0.192 TG-TransUnet 0.525 0.664 0.258 0.206 3 h RISEgust 0.092 0.523 0.083 0.077 U-net 0.374 0.716 0.193 0.150 CU-net 0.378 0.702 0.200 0.158 TransU-net 0.538 0.738 0.214 0.164 TG-TransUnet 0.421 0.678 0.223 0.181 4 h RISEgust 0.066 0.565 0.061 0.056 U-net 0.226 0.761 0.132 0.103 CU-net 0.354 0.771 0.162 0.125 TransU-net 0.387 0.728 0.190 0.156 TG-TransUnet 0.436 0.692 0.220 0.186 5 h RISEgust 0.054 0.609 0.050 0.045 U-net 0.216 0.884 0.082 0.052 CU-net 0.209 0.868 0.088 0.060 TransU-net 0.269 0.872 0.095 0.064 TG-TransUnet 0.316 0.868 0.103 0.071 6 h RISEgust 0.036 0.683 0.034 0.030 U-net 0.136 0.893 0.064 0.044 CU-net 0.185 0.895 0.072 0.049 TransU-net 0.251 0.897 0.079 0.054 TG-TransUnet 0.180 0.868 0.082 0.062 Table 5. Skill scores of RISEgust and four different DL algorithms in forecasting thunderstorm gusts with a lead time from 1 to 6 h. The bolded entries correspond to the highest CSI and ETS scores among the five methods.
RISEgust, as a traditional method, is used in this study. The experimental results clearly show that the values of CSI and ETS with the RISEgust method are the lowest, and the TransU-net family (including TransU-net and TG-TransUnet) yields significantly better values of CSI and ETS than does the U-net family (including U-net and CU-net). This improvement can be largely attributed to the integration of the transformer and CNN mechanisms in the TransU-net family, which effectively compensates for some of the limitations of solely using the CNN. The transformer has strong modeling capability as it can be considered as a graph-based modeling approach. Its data-driven method enables it to learn the relationships between nodes effectively, making it highly adaptable and versatile. A CNN is a local operation where a convolution layer typically captures the relationship between neighboring pixels. A transformer, on the other hand, is a global operation where a transformer layer can effectively model relationships between all pixels. MHSA can produce more interpretable models, as each attention head has the ability to learn and perform distinct tasks. CU-net yields significantly better results than U-net, demonstrating that the sub-pixel convolution module can effectively enhance the forecasting performance for thunderstorm gusts. The experimental findings of Han et al. (2021) support the superiority of the sub-pixel convolution module over traditional methods like bilinear interpolation in upsampling, as it reduces the impact of human factors.
Since the location of the thunderstorm gusts is the focus of our attention, we introduce a method (CA module) by embedding the location information into the channel attention mechanism. In the forecast results of 1–6 h, the values of CSI and ETS with TG-TransUnet are slightly higher throughout than those with TransU-net. This can mainly be attributed to the fact that feature extraction becomes harder with longer forecasting periods, whereby the attention mechanism and sub-pixel convolution exhibit a positive impact on performance. Despite the increased difficulty of forecasting for the 4–6 h timescale, TG-TransUnet outperforms TransU-net with the assistance of attention mechanisms and sub-pixel convolution. These techniques enable TG-TransUnet to achieve better forecasting results. For example, in the case of a 6–h ahead forecast, TG-TransUnet achieves a CSI of 0.082, surpassing TransU-net’s score of 0.079. Overall, as the forecasting time increases, the values of CSI and ETS gradually decline except for TG-TransUnet’s ETS and HSS at 4–h compared to 3–h, which increase slightly.
-
In this section, separate cases within one specific individual process of thunderstorms gusts, which started at 1300 UTC 12 June 2022 and ended at 1500 UTC 12 June 12 2022, are selected to analyze the forecasting effectiveness of RISEgust, U-net, CU-net, TransU-net, and TG-TransUnet at lead times of 1–3 h.
-
Figure 8 shows the results of the thunderstorm gusts forecast at a lead time of 1 h at 1300 UTC 12 June 2022. In the forecast, RISEgust forecasts the existence of a portion of the ground truth of thunderstorm gusts with wind speed at 10-m height greater than or equal to 17.2 m s−1 (or visually estimated winds of magnitude 8 or higher), but the range is much smaller than the ground truth, and the other four DL models produce better forecast results than RISEgust. In addition, RISEgust forecasts areas with winds of magnitude 8 or higher across the southern part of Tianjin as well as the southern part of Hebei, but according to the radar reflectivity these areas cannot be considered as having experienced thunderstorm gusts. However, this affects the forecast results of the other four DL models: both U-net and CU-net forecast thunderstorm in the south of Tianjin, which is a very obvious empty report. In contrast, TransU-net and TG-TransUnet do not produce such a situation. CU-net, TransU-net and TG-TransUnet all forecast thunderstorm gusts to occur in parts of southern Hebei, which corresponds to the areas forecasted by RISEgust. The forecast ranges of CU-net and TransU-net are significantly smaller than those of U-net and TG-TransUnet in the areas similar to the ground truth of thunderstorm gusts, and the results of the CU-net and TransU-net forecasts are not consistent with a whole connected area.
Figure 8. Thunderstorm gust forecasts at a lead time of 1 h at 1300 UTC 12 June 2022: (a) ground truth of thunderstorm gusts; (b) gust forecast results at 10-m height produced by RISEgust (color shading represents wind speed; units: m s−1); (c) thunderstorm gusts forecast results of the U-net model; (d) thunderstorm gusts forecast results of the CU-net model; (e) thunderstorm gusts forecast results of the TransU-net model; (f) thunderstorm gusts forecast results of the TG-TransUnet model.
Figure 9 shows the results of the thunderstorm gusts forecast at a lead time of 2 h at 1300 UTC 12 June 2022. This time, there is basically no cross over between the RISEgust forecast of winds of magnitude 8 or higher and the ground truth of thunderstorm gusts, but the other four DL models show some forecasting capability. Specifically, their forecast results include not only part of the area in and around the ground truth of thunderstorm gusts, but also part of the area with winds of magnitude 8 or higher as forecasted by RISEgust. Among the four DL models, U-net has a large number of empty reports in the southern, central and eastern parts of Hebei, while TG-TransUnet has the fewest empty reports and avoids the discontinuity of the forecast range, which is significantly better than with the other three models.
Figure 10 shows the thunderstorm gusts forecast at a lead time of 3 h at 1300 UTC 12 June 2022. In this forecast, RISEgust basically only forecasts the occurrence of winds of magnitude 8 or higher in some areas of southern Hebei, which is completely incommensurate with the ground truth of thunderstorm gusts, while the other four DL models forecast a range larger than that of RISEgust. Among them, the forecast effect of U-net is the worst, and there is basically no cross-correlation with the ground truth of thunderstorm gusts; the forecast results can basically be regarded as all empty reports. The forecasts of the remaining three DL models are hardly comparable, but it is worth noting that TG-TransUnet has relatively fewer empty reports, especially in the Beijing area and at the junction of Beijing, Tianjin, and Hebei.
Figure 10. As in Fig. 8 but at a lead time of 3 h.
-
Figure 11 shows the thunderstorm gusts forecast at a lead time of 1 h at 1400 UTC 12 June 2022. In the forecast, RISEgust forecasts the area with winds of magnitude 8 or higher mainly in the southern part of Hebei and the northern part of eastern Beijing, as well as the adjacent Hebei area of Tianjin, which is also the main area of the ground truth of thunderstorm gusts. All four DL models forecast the approximate area of thunderstorm gusts, but with some slight differences. Among them, U-net has a large range of empty reports in the western part of Hebei, while the remaining three DL models all forecast thunderstorm gusts in the southern part of Hebei, which is mainly due to the large wind speeds in this part of the region. The range of the TG-TransUnet accurate forecast is larger than that of CU-net and TransU-net, and the forecast results of TransU-net do not form a continuous through area.
Figure 11. Thunderstorm gusts forecast at a lead time of 1 h at 1400 UTC 12 June 2022: (a) ground truth of thunderstorm gusts; (b) gust forecast results at 10-m height produced by RISEgust (color shading represents wind speed; units: m s−1); (c) thunderstorm gusts forecast results of the U-net model; (d) thunderstorm gusts forecast results of the CU-net model; (e) thunderstorm gusts forecast results of the TransU-net model; (f) thunderstorm gusts forecast results of the TG-TransUnet model.
Figure 12 shows the thunderstorm gusts forecast at a lead time of 2 h at 1400 UTC 12 June 2022. This time, RISEgust only forecasts thunderstorm gusts in some areas, but its forecasts of winds of magnitude 8 or higher in southern Tianjin and southern Hebei affect the forecasts of the other four DL models, with all of them forecasting thunderstorm gusts in southern Hebei, which is contrary to the ground truth. However, it is obvious that TG-TransUnet has significantly fewer nulls than the other DL models, and produces better forecasts.
Figure 13 shows the thunderstorm gusts forecast at a lead time of 3 h at 1400 UTC 12 June 2022. In this forecast, RISEgust has essentially no forecasting effect. The remaining four DL models also produce worse forecasts than those at a 2-h lead time. TG-TransUnet forecasts better coverage than U-net and has fewer short reports than CU-net and TransU-net.
-
Figure 14 shows the thunderstorm gusts forecast at a lead time of 1 h at 1500 UTC 12 June 2022. In the forecast, RISEgust forecasts winds of magnitude 8 or higher mainly in the north-central and eastern parts of Hebei Province, which has some overlap with the ground truth. The remaining four DL models all forecast the general area of thunderstorm gust occurrence, but U-net has a certain range of shortcomings in the western part of Hebei, and the TransU-net forecast does not form a continuous through-area. Meanwhile, the CU-net and TG-TransUnet forecasts are relatively good, and the range of the TG-TransUnet forecast is slightly larger.
Figure 14. Thunderstorm gusts forecast at a lead time of 1 h at 1500 UTC 12 June 2022: (a) ground truth of thunderstorm gusts; (b) gust forecast results at 10-m height produced by RISEgust (color shading represents wind speed; units: m s−1); (c) thunderstorm gusts forecast results of the U-net model; (d) thunderstorm gusts forecast results of the CU-net model; (e) thunderstorm gusts forecast results of the TransU-net model; (f) thunderstorm gusts forecast results of the TG-TransUnet model.
Figure 15 shows the thunderstorm gusts forecast at a lead time of 2 h at 1500 UTC 12 June 2022. This time, RISEgust forecasts wind gusts of magnitude 8 and above mainly in the northwestern and eastern regions of Hebei, which overlap somewhat with the ground truth. Among the other four DL models, both U-net and TransU-net have large-scale empty reports in the central and southern regions of Hebei, and the forecasts of CU-net and TG-TransUnet are relatively better. Meanwhile, the number of empty reports produced by TG-TransUnet are fewer compared with those of CU-net, and the overall forecasts are slightly better.
Figure 16 shows the thunderstorm gusts forecast at a lead time of 3 h at 1500 UTC 12 June 2022. In this forecast, RISEgust is essentially ineffective. U-net only forecasts the general area, but this is far from the ground truth. CU-net, TransUnet, and TG-TransUnet all roughly forecast thunderstorm gusts, but all produce empty reports in parts of southern Hebei. On the whole, TG-TransUnet produces a smaller number of empty reports compared to the other DL models.
Product | Parameter (s) | Specific meaning | Unit |
AWS DATA | WSX | Extreme wind speed in 1 h | m s−1 |
LIGHTNING DATA | LOCATION | Two-dimensional spatial information | ° |
TIME | Time of occurrence | UTC | |
RMAPS-RISE | UVana | 10-m height analysis of current gusts | m s−1 |
UVpred_RISE | 10-m height prediction of future gusts | m s−1 | |
RRpred | 1-h cumulative precipitation forecast | mm | |
TQdiff | Difference between current and predicted 2-m surface temperature | °C | |
RADAR DATA | RADAR | Composite radar reflectivity factor | dBZ |
RMAPS-ST | PRES | Pressure | Pa |
TMPdiff | Temperature difference between 850 hPa and 500 hPa | K | |
CAPE | Convective available potential energy | J kg−1 | |
RADARST | Composite radar reflectivity | dBZ | |
UVpred_ST | Average wind speed at 80-m height | m s−1 | |
SHEAR1 | 0–1 km vertical wind shear | m s−1 | |
SHEAR2 | 0–6 km vertical wind shear | m s−1 |