Reaserch On Thunder Cloud Monitoring And Lightning Nowcasting Method Based On Multi-source Data

Lightning disaster is the major natural disaster in many industries such as electric power system.High-resolution lightning nowcasting techniques are mainly based on real-time observation data,such as atmospheric electric field mill data,lightning location data and Doppler weather radar volume scan data.These data contain different levels and types of information in the lightning process,however there still exists some challenges in the comprehensive utilization of them.For example,electric field mills can be greatly affected by their installation environments,resulting in low mutual reference between different sites of the networked electric field mills.The dynamic information of thundercloud in high resolution radar volume scan data is not fully utilized.lacking objective and stable obtaining method of thunder cloud charge model,the inversion products of thunder cloud charge structure which solely basing on electric field data can hardly be applied to lightning nowcasting.And it is difficult to comprehensively analyze the spatiotemporal sequence data of multi-source data to obtain early warning products.To address these challenges,the following work has been done in this study:The temporal and spatial distribution characteristics of lightning in Guangdong region were analyzed using historical lightning location data,and the correlations between the lightning distribution characteristics and factors such as population,topography and land-use types are discussed.Based on the results of lightning distribution characteristics analysis,30 electric field mills and one X-band dual-polarization Doppler radar are deployed at selected sites in and around Guangzhou city,local lightning location system data are also imported to build a thundercloud monitoring and lightning nowcasting system,these provide observation data for subsequent studies.To improve the utilization efficiency of the networked electric field mills data,the influences of the real terrain,buildings and high voltage overhead lines on the ground electric field are simulated based on the finite element method.The results show that large-scale topographic differences and small-scale overhead lines will both have nonnegligible influence on the ground electric field,which makes it impossible to obtain ideal calibration results for networked electric field mills by selecting open fields or using calibration method based on fine weather electric field.Therefore,a new calibration method for networked electric field mills based on electric field jumps in lightning process is proposed,and its feasibility and validity are verified by experiments on simulated and measured data.To obtain higher resolution 3D motion field information for interpolation or extrapolation of radar data in lightning nowcasting applications,a 3D radar data motion field estimation model based on a 3D point cloud deep learning network is built,a dataset creation method based on synthetic motion field is introduced,and the feasibility of this method is verified through experiments on synthetic data and real data,respectively.Furthermore,the proposed method is compared with the traditional 2D optical flow method in a data reconstruction task.The result shows that even if the 3D result is compressed into 2D form,the proposed method can obtain better reconstruction results than the traditional 2D optical flow method.To comprehensively use radar data and networked electric field mills data in thunderstorm charge model inversion task,the problems that need to be considered in obtaining point cloud form thundercloud charge models based on radar volume scan data are analyzed.Then,an extraction method of point cloud charge model based on hydrometeor thresholds and an identification method of thundercloud cells based on density clustering algorithm are proposed.The feasibility of the proposed method is verified by being applied measured data of a recorded lightning process.A by-product of thunderstorm charge model inversion called analytical electric filed is proposed.Through case analysis,it shows that this product combines the advantages of electric field mill data and radar data,has a high spatio-temporal correlation with lightning distribution,and is suitable for being one of the inputs of lightning nowcasting algorithm.On the basis of the above studies,a lightning nowcasting method based on multi-source data and Conv-LSTM(Convolutional Long Short-Term Memory)structure is proposed.First,the pre-processing method of multi-source data and the dataset curation method are introduced.Moreover,the encoding-forecasting model structures,the experimental schemes and control groups are introduced.According to the experimental results,the proposed network outperforms the network in control group,and the results using multisource data as input are better than those using single data as input.Furthermore,an application scheme of the nowcasting products is proposed,and application cases are presented.The scheme can provide lightning nowcasting products with different levels and lead times above 1 km resolution.The average POD(Probability of Detection),FAR(False Alarm Rate)and CSI(Critical Success Index)of the first-level warning products can reach 0.578,0.267 and 0.477 respectively under 0-30 min lead time.