A Lightning Nowcasting Method and System Based on Spatiotemporal Transformer and Uncertainty Awareness

Abstract

This application relates to the field of meteorological forecasting technology, and discloses a method and system for lightning nowcasting based on spatiotemporal Transformer and uncertainty perception. The method includes: acquiring multi-source meteorological observation data; preprocessing and spatiotemporally aligning the multi-source meteorological observation data; inputting the multimodal input sequence into a trained uncertainty-aware spatiotemporal Transformer model to obtain a discounted lightning occurrence probability field; and determining and outputting lightning forecast information based on the lightning occurrence probability field. This application effectively solves the problems of easy over-triggering, high false alarm rate, unstable spatial structure, and lack of confidence expression in traditional deep learning lightning forecasting. Based on the unified alignment and fusion of multi-source observations, the model achieves more reliable and controllable lightning occurrence probability prediction, providing a directly implementable technical path for lightning warning services.

Description

Technical Field

[0001] This application belongs to the field of meteorological forecasting technology, and specifically relates to a method and system for lightning nowcasting based on spatiotemporal Transformer and uncertainty perception. Background Technology

[0002] Lightning, a highly destructive convective weather phenomenon, frequently causes casualties, infrastructure damage, power system failures, and significant economic losses. Its strong locality and rapid evolution pose enormous challenges to disaster prevention and emergency response. With the intensification of climate change and the increasing frequency of extreme convective events, the development of accurate, fine-scale lightning nowcasting models has become increasingly important. This is especially true for regions like Qingdao, located in the eastern coastal monsoon region and influenced by ocean-land circulation and complex topography, where convective processes are diverse and frequent, making lightning forecasting particularly difficult.

[0003] Existing lightning forecasting technologies mainly include extrapolation techniques, numerical weather prediction models, and machine learning methods. Early studies primarily employed extrapolation techniques such as centroid tracking and cross-correlation to estimate the direction and velocity of convective cells by analyzing continuous lightning observations. These methods can capture the general propagation trend of lightning activity, but they rely on a single data source, lack sufficient feature representation, and cannot adequately characterize nonlinear convective evolution, resulting in low prediction accuracy under complex meteorological conditions. Numerical weather prediction models explicitly represent charge separation and discharge mechanisms through convective parameterization schemes and microphysical process simulations; for example, the Price-Rind model establishes an empirical relationship between vertical velocity and lightning frequency. However, traditional numerical prediction models are computationally expensive, sensitive to observational errors, and struggle to provide warnings with lead times of minutes or shorter.

[0004] Despite the progress made by existing deep learning models in lightning prediction, significant limitations remain. First, while these models achieve relatively high detection rates, their false alarm rates are generally high, limiting their reliability in practical applications. This is primarily due to their inability to effectively capture the spatiotemporal evolution of complex convective systems, measurement noise, and heterogeneous environmental conditions. In areas with weak lightning signals or strong background interference, predictions often exhibit overconfidence, leading to frequent false alarms. Second, most existing methods focus on point-based or deterministic predictions, neglecting the probabilistic characteristics of lightning occurrence and the quantification of prediction uncertainty, which are crucial for risk perception and decision-making. Furthermore, traditional convolutional neural networks have limited ability to capture long-range spatial dependencies, making it difficult to model the large-scale evolution patterns of convective systems.

[0005] Regarding multi-source data fusion, although meteorological observation data (including dual-polarization radar, lightning location networks, and automatic weather station data) are accumulating rapidly, existing models have weak spatiotemporal feature extraction capabilities for these multimodal data, making it difficult to fully utilize their inherent spatial topological relationships and temporal dynamic characteristics. How to align and fuse heterogeneous data sources within a unified spatiotemporal framework to provide a physically consistent representation of convective systems remains an unsolved challenge.

[0006] Therefore, there is an urgent need to develop a lightning nowcasting model that can effectively reduce the false alarm rate while maintaining a high detection rate (Transformer), explicitly quantify the prediction uncertainty (UQ), and make full use of multi-source meteorological data. Summary of the Invention

[0007] To address or at least mitigate one or more of the above problems, a lightning nowcasting method and system based on spatiotemporal Transformer and uncertainty perception is provided. This method can provide a lightning nowcasting model that quantifies forecast uncertainty and effectively integrates multi-source data to improve forecast reliability, reduce false alarms, and provide more comprehensive information support for risk prevention and control.

[0008] To achieve the above objectives, according to the first aspect of this application, a lightning nowcasting method based on spatiotemporal Transformer and uncertainty perception is provided, comprising: Acquire multi-source meteorological observation data, which includes at least radar data, lightning location data, and automatic weather station data; The multi-source meteorological observation data are preprocessed and spatiotemporally aligned to generate a multimodal input sequence under a unified spatiotemporal grid; The multimodal input sequence is fed into a trained uncertainty-aware spatiotemporal Transformer model. The model extracts and fuses multimodal features to obtain multimodal fusion features, and then performs spatiotemporal dependency modeling based on these multimodal fusion features using a geographic distance-aware biased Transformer attention mechanism to obtain spatial augmentation features. The spatial augmentation features are then aggregated in the temporal dimension and their spatial resolution restored to generate nominal lightning probability fields under multiple forecast lead times. Uncertainty quantification is performed on the nominal lightning probability fields to obtain the prediction mean and variance for each spatial location. Based on the variance, a confidence discount is applied to the nominal probability of high-uncertainty regions to obtain the discounted lightning probability field. Based on the lightning occurrence probability field, lightning forecast information is determined and output.

[0009] To achieve the above objectives, according to a second aspect of this application, a lightning nowcasting system based on spatiotemporal Transformer and uncertainty perception is provided, comprising: The acquisition module is used to acquire multi-source meteorological observation data, which includes at least radar data, lightning location data, and automatic weather station data. The preprocessing module is used to preprocess and spatiotemporally align the multi-source meteorological observation data to generate a multimodal input sequence under a unified spatiotemporal grid. The data processing module loads a spatiotemporal Transformer model with uncertainty awareness, the model including at least a multimodal encoder, a feature fusion module, a spatiotemporal Transformer module ST-Block, a decoder, and an uncertainty awareness module UQCD; The multimodal input sequence is input into a trained uncertainty-aware spatiotemporal Transformer model. The multimodal encoder and feature fusion module are used to perform multimodal feature extraction and fusion on the multimodal input sequence. The spatiotemporal Transformer module ST-Block is used to perform spatiotemporal dependency modeling on the fused features based on a distance-constrained attention mechanism, and outputs a spatially enhanced feature representation. The temporal decoder is used to perform temporal dimension aggregation and spatial resolution restoration on the spatially enhanced feature representation to generate a nominal lightning occurrence probability field under at least one forecast lead. The uncertainty-aware module UQCD is used to quantize the uncertainty of the nominal lightning occurrence probability field, obtain the prediction mean and variance corresponding to each spatial location, and perform confidence discounting on the nominal probability of high uncertainty areas based on the variance, and output the final lightning occurrence probability field used for early warning. The output module is used to determine and output lightning forecast information based on the lightning occurrence probability field.

[0010] By adopting the above technical solution, this application has the following beneficial effects compared with the prior art: This application addresses the core bottleneck of achieving both high detection rate (POD) and low false alarm rate (FAR) in lightning nowcasting services by proposing an uncertainty-aware spatiotemporal Transformer framework for lightning nowcasting. In terms of overall technical logic, this application does not simply suppress false alarms by increasing thresholds or post-processing. Instead, it employs a paired design of uncertainty quantization to reduce FAR and spatiotemporal Transformer to improve POD: First, uncertainty quantization and confidence discounting are used to proactively suppress overconfidence triggering in highly ambiguous regions, reducing false alarms at the source. Simultaneously, this scheme also considers the known constraints between reducing false alarm rate and maintaining high detection rate. Therefore, it further introduces distance-aware Spatial Transformer Block (ST-Block) and temporal decoding to enhance the long-distance spatiotemporal dependency modeling capability of convective systems. This significantly improves the efficiency of feature extraction from real lightning events while suppressing noise, achieving a better POD-FAR balance and more stable spatial structure output for operational purposes. This framework achieves spatiotemporal alignment and fusion of multi-source observations from radar, lightning location (LIG), and automatic weather station (AWS) on a unified latitude and longitude grid, providing the model with physically consistent input representations and avoiding systematic biases caused by cross-source aliasing.

[0011] In this application, a distance-aware spatiotemporal Transformer module is constructed to overcome the shortcomings of traditional Transformer models and improve the detection rate. Although uncertainty quantification effectively reduces false alarms, relying solely on this mechanism inevitably leads to a conservative model and a reduced detection rate. To improve detection capabilities while maintaining low false alarms, this application introduces a spatiotemporal Transformer architecture. While traditional Transformers possess a global receptive field, they lack awareness of geographic physical space, easily establishing false long-distance associations that do not conform to meteorological patterns, such as erroneous attention interactions between isolated pixels that are extremely far apart. This not only introduces computational redundancy but also limits the accurate capture of true convection structures. To address this, this application proposes an improved spatiotemporal Transformer module (ST-Block). The core innovation of this module lies in the introduction of an attention bias matrix based on Haversine distance. By imposing geographic distance constraints in attention calculation, the model can suppress non-physical long-distance skip connections and strengthen feature interactions within local and mesoscale convection systems. This design enables the model to more accurately capture the morphological evolution and long-range spatiotemporal dependencies of convective systems, thereby improving detection capabilities by leveraging enhanced spatial reasoning while reducing false alarms through the uncertainty module. Ablation experiments show that this module is key to maintaining high POD; removing it causes the model to become unstable in spatial structure predictions over long time periods, resulting in a significant decline in detection capability.

[0012] This application proactively suppresses false alarms in high-risk areas by introducing uncertainty quantification and confidence discounting mechanisms, addressing the problem of overconfidence. To address the issue that existing deterministic models cannot perceive predicted risks and are prone to high false alarm rates in weak echo areas or under noise interference, this application innovatively designs a Bayesian uncertainty quantization head based on a logistic-normal distribution at the decoder end. This module not only outputs the predicted mean of lightning occurrence but also simultaneously predicts the logarithmic variance to characterize the prediction uncertainty. Based on this, this application further proposes a confidence discounting mechanism. This mechanism adaptively penalizes the nominal probability using the predicted standard deviation: in areas with blurred convection boundaries, sparse observation data, or complex background noise, the model automatically increases the uncertainty estimate, thereby significantly reducing the final triggering probability in that area through a discounting algorithm. Experiments demonstrate that this explicit probability penalty mechanism directly addresses the pain point of the sensitivity-accuracy dilemma, sacrificing a very small number of low-confidence detections in exchange for a significant reduction in the false alarm rate, making early warning triggering more prudent and reliable.

[0013] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application. The illustrative embodiments and descriptions of the application are used to explain the application, but do not constitute an undue limitation of the application. Obviously, the drawings described below are merely some embodiments, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0015] In the attached diagram: Figure 1 This is a flowchart illustrating the lightning nowcasting method based on spatiotemporal Transformer and uncertainty perception in this specific embodiment. Figure 2 This is a training architecture diagram of the UQ-STNet model in this specific implementation; Figure 3 This is a flowchart illustrating the training process of the UQ-STNet model in this specific implementation. Figure 4 This is a schematic diagram of the multimodal encoder and decoder of the UQ-STNet model in this specific embodiment; Figure 5 This is a schematic diagram of the feature fusion module and uncertainty perception module of the UQ-STNet model in this specific embodiment; Figure 6This is a structural diagram of the lightning nowcasting system based on spatiotemporal Transformer and uncertainty perception in this specific embodiment; Figure 7 This is a comparison chart of the prediction performance of the UQ-STNet model and several baseline models in this specific implementation. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0017] Example 1: Please see Figure 1 This application provides a lightning nowcasting method based on spatiotemporal Transformer and uncertainty perception, including: Acquire multi-source meteorological observation data, which includes at least radar data, lightning location data, and automatic weather station data; The multi-source meteorological observation data are preprocessed and spatiotemporally aligned to generate a multimodal input sequence under a unified spatiotemporal grid; The multimodal input sequence is fed into the trained uncertainty-aware spatiotemporal Transformer model UQ-STNet. The model extracts and fuses multimodal features to obtain multimodal fusion features, and performs spatiotemporal dependency modeling based on the multimodal fusion features using a geographic distance-aware biased Transformer attention mechanism to obtain spatial augmentation features. The spatial augmentation features are then aggregated in the temporal dimension and their spatial resolution restored to generate nominal lightning occurrence probability fields under multiple forecast lead times. Uncertainty quantification is performed on the nominal lightning occurrence probability fields to obtain the prediction mean and variance for each spatial location. Based on the variance, a confidence discount is applied to the nominal probability of high-uncertainty areas to obtain the discounted lightning occurrence probability field. Based on the lightning occurrence probability field, lightning forecast information is determined and output.

[0018] To address the problems of high false alarm rates, insufficient characterization of complex spatiotemporal evolution of convection, inadequate fusion of multi-source data, and lack of probabilistic and uncertainty representation in existing lightning nowcasting methods for local warning scenarios, this invention proposes a lightning nowcasting method based on spatiotemporal Transformer and uncertainty awareness. An uncertainty-aware spatiotemporal Transformer model, UQ-STNet (Uncertainty-aware Spatiotemporal Transformer Network), is constructed. This method takes multi-source observations from radar, lightning location, and automatic weather station as inputs, aligning and fusing them on a unified latitude and longitude grid. The distance-constrained spatiotemporal Transformer module ST-Block captures the long-range spatial dependence of convective systems and suppresses false alarms triggered by isolated pixels. Simultaneously, the uncertainty awareness module UQCD explicitly learns and predicts the mean and variance, discounting high-confidence triggers in high-uncertainty regions, thereby significantly reducing the false alarm rate while maintaining high detection capability.

[0019] The specific technical solution is as follows: Data input and processing: To ensure complementary information from multiple sources and reduce systematic biases caused by scale inconsistencies, this invention spatially registers all data sources to the same latitude and longitude grid and temporally aligns them to a uniform sampling rate before inputting them into the model. The multi-source data used in this invention includes: Radar data: Q-region dual-polarization Doppler radar volume scan data, updated approximately every 6 minutes, and mapped to approximately... The grid extracts products such as comprehensive reflectivity (CREF), echo top height (ET), and vertical liquid water content (VIL) to characterize convection intensity, vertical development, and condensate content, and forms a three-dimensional tensor input network that varies with time.

[0020] Automatic Weather Station Data (AWS): Synchronous observations from high-density automatic weather stations in Province S were used. Key elements included air temperature (T), maximum wind speed (WIN_S_MAX), and relative humidity (RHU). This set of variables provides boundary layer thermodynamic-dynamic-moisture constraints, providing near-surface diagnostic information for modeling the variability of lightning occurrence environments and noise sensitivity.

[0021] Lightning observation data: Data from the lightning monitoring and positioning network of Province S was used, covering the period from approximately August 2022 to October 2024. After spatiotemporal binning, the data was mapped into a regular grid and formed an hourly-scale lightning density / intensity grid field. The sampling interval was aligned with the hourly scale of the automatic stations for fusion.

[0022] Model data processing: The uncertainty-aware spatiotemporal Transformer model trained in this embodiment includes at least a multimodal encoder, a feature fusion module, a distance-constrained spatiotemporal Transformer module ST-Block, a decoder, and an uncertainty-aware module UQCD.

[0023] Multimodal coding and feature fusion: On a unified latitude and longitude grid, rasterized lightning prediction is formulated as a multimodal spatiotemporal learning problem. Let time... Lower mode The preprocessed raster input is: (1); in, and These represent the spatial height and width of the grid, respectively. Representing modes The number of feature channels per. The relevant spatiotemporal information of the corresponding observation sources is encoded. The data of each mode are spatially registered to the same grid and temporally aligned to a uniform sampling rate to avoid systematic prediction bias caused by aliasing and to ensure cross-modal consistency.

[0024] For each mode, a mode-specific encoder is used to map it to the latent feature space, resulting in the corresponding feature map: (2); in, and Indicates the spatial dimensions of the feature map. Indicates the number of channels. Indicates time Lower mode Latent feature map, Indicates time Lower mode Preprocessed raster input, Indicates modality A dedicated encoder. These latent features capture key spatiotemporal patterns of each modality, providing a foundation for subsequent multimodal fusion and Transformer-based processing.

[0025] Subsequently, the three modal features are concatenated in the channel dimension, through... Projective convolution completes dimension alignment and fusion, resulting in multimodal fused features. , represented as: (3); in, and The spatial dimension of the fused feature map. For the number of channels, This represents the encoded radar modal feature map. This represents the encoded lightning localization modal feature map. This represents the encoded modal feature map of the automatic weather station. The resulting feature map integrates key spatiotemporal information from multiple modes, providing a unified input for the subsequent Transformer module. This fusion method preserves complementary information from the three types of data, which helps distinguish between real convective signals and background noise, structurally reducing false detections.

[0026] The Spatiotemporal Transformer module ST-Block is designed to effectively capture long-range spatial dependencies and enhance the representation of convection and lightning-related features in a rasterized spatiotemporal context.

[0027] The spatiotemporal Transformer module operates on the fused multimodal feature maps, primarily comprising the following three stages: First, the feature maps are tokenized and injected with 2D location encoding to preserve spatial structure information. Second, a Transformer layer incorporating multi-head self-attention and distance-aware bias models nonlocal interactions while introducing physically reasonable geographic constraints, enabling the network to correlate convection signals over long distances—crucial for lightning prediction influenced by large-scale atmospheric circulation. Finally, the temporal decoder aggregates the processed features and upsamples them back to the original grid resolution, outputting pixel-by-pixel lightning probability forecasts at multiple lead times.

[0028] By integrating spatial attention, location information, and temporal decoding, the spatiotemporal Transformer module ST-Block becomes a core component for characterizing the complex spatiotemporal relationships of convective systems.

[0029] Tokenization and location encoding: for each time step We first fuse multimodal features Convert to a token sequence and inject explicit two-dimensional location information: (4); in, Represents grid cells ( The quantity of ) This represents a reshaping operation that flattens the spatial dimensions of the feature map into a sequence. Indicates time The characteristic token sequence, Each line corresponds to a time Next latitude and longitude grid point Dimensionality representation. Matrix It is a learnable two-dimensional position encoding used to encode the spatial coordinates of each grid cell, enabling the Transformer to perceive absolute position and relative geometric relationships, rather than relying on permutation-invariant coordinates. This is particularly important for convective systems, because the same echo pattern appearing in different regions may correspond to drastically different physical mechanisms.

[0030] Transformer layer: The token sequence is processed through a standard Transformer encoder block, which includes a multi-head self-attention (MSA) network and a position-wise feedforward network (FFN), and employs pre-normalization and residual connections. (5a); (5b); in, For layer normalization, In all Multi-head self-attention is calculated at each grid point. For each Two independently operating feedforward subnetworks. (Intermediate representation) By aggregating non-local information through attention, The final token sequence incorporates global context and local nonlinear transformations. This enables the model to capture long-range spatial dependencies that are difficult to model with pure convolutional structures, such as long-range triggering relationships between upstream and downstream convection units.

[0031] Distance-aware attention: To further enhance nonlocal interactions that conform to physical laws, we introduce a distance-aware bias matrix into the attention scoring. : (6); in Each attention head is separate The query, key, and value matrix obtained by linear projection The dimension of the key. Bias matrix. Adjust before softmax The similarity score between features reduces the weight of geographically distant or physically unreasonable associations, while enhancing the salience of connections in neighboring regions or meteorologically relevant areas. In this way, the attention mechanism can simultaneously consider the similarity of learned features and the underlying geophysical geometry.

[0032] set up express The latitude and longitude, and The Haversine distance is: (7); in, For the Earth's radius, This represents the great circle distance between two grid cells on a sphere. Compared to Euclidean distance, the Haversine distance provides a more accurate separation at the regional scale and naturally reflects the propagation path of large-scale weather systems.

[0033] We map the distance to a bias term: (8a); (8b); in, The average pairwise distance used for normalization, For normalized distance, It is a scalar hyperparameter. When the distance exceeds the sparse radius Strict penalties are imposed to discourage attention from leaping over unrealistically long distances; This represents a geophysical mask used to further encourage or inhibit certain connections. Equations (8a) and (8b) make it more likely that nearby convective structures will exchange information, while very distant or physically disconnected regions contribute less to attention.

[0034] After distance-aware attention and Transformer processing, the token sequence Reconstructed back into a grid format, serving as a spatial enhancement feature for subsequent temporal decoding: (9); in, This represents a spatially enhanced feature map that has been reshaped back into a mesh format.

[0035] Time decoding: Decoder aggregates along the time axis And upsampled to through two stages of deconvolution. Output One forecast lead time: (10); in, This represents the spatial feature sequence generated by ST-Block. It is a time decoder used to aggregate information along the time axis and output the lead amount. Feature map; and It is a two-dimensional deconvolution, used to progressively convert features from... Upsample back to the original mesh The final one Convolution generates one for each grid cell , Output lead probability forecast This decoding path ensures that the long-range spatial context captured by ST-Block can be reliably projected back into a high-resolution, pixel-by-pixel probability field, meeting the needs of business applications.

[0036] Uncertainty Awareness Module: To address the issue of deep models being prone to "overconfidence" in scenarios with weak signals, strong interference, sparse observations, or clutter contamination, we propose an uncertainty quantification module with confidence discounting. This module models prediction dispersion, suppresses unreliable high-probability predictions, and improves overall detection performance. In this module, the mean and variance of the logit distribution are predicted simultaneously using decoder features, resulting in a probability field with uncertainty awareness within a logistic-normal prediction framework. Subsequently, a confidence discounting mechanism adaptively reduces the confidence level of high-dispersion predictions, thereby achieving a controllable trade-off between detection rate and false alarm rate in ambiguous regions.

[0037] Parameterized uncertainty head: for lead time Decoding features, using The header outputs the logit mean and log-variance. The nominal probability is a sigmoid function of the mean, and the standard deviation is recovered from the log-variance. (11); in, Lead time Decoding feature map. It is the deterministic logit output of the traditional model, which is then processed by the sigmoid function to obtain the nominal probability. In ambiguous regions, such as convective boundaries, weak echoes, sparse data, or clutter-contaminated areas, these models tend to be overconfident; conversely, The standard deviation is obtained from the log-variance of the encoded predictions. And introduce To ensure numerical stability, this dispersion parameter characterizes the model's stability. The degree of uncertainty provides the additional degrees of freedom needed to break the classic sensitivity-accuracy dilemma, in which a higher probability of detection usually comes at the cost of an increased false alarm rate.

[0038] Logistic-normal prediction perspective: We treat the prediction probability as a random variable induced by Gaussian perturbations in the logit space, thus obtaining the logistic-normal model: Treating the predicted probability as a random variable induced by Gaussian perturbations in the logit space, which follows a logistic-normal distribution, we can express it as follows: (12); in, Lead time The probability of random lightning occurrence, It is a standard normal random variable. Characterizing the central trend of the forecast, Control the degree of dispersion induced by the logistic-normal distribution. Larger... It often appears in regions where convection is more complex or observation is less sufficient. These are precisely the regions where deterministic single logit outputs are prone to overconfidence. Therefore, this mechanism explicitly marks high uncertainty regions that need to be discounted in the decision-making stage. Here, the logistic-normal model refers to the mathematical model expressed by formula (12).

[0039] Confidence discounting for alarm control: To suppress overconfident alarms under high uncertainty, based on the confidence level... The nominal probability is reduced in weight proportionally and then clipped to a smaller value. Interval: (13); in, This is the probability after uncertainty discounting that is ultimately used for alarms. Limit the results to the range of valid probabilities. It is a positive hyperparameter used to control uncertainty. The intensity of the penalty for high nominal probabilities; the experiment adopted As a robust default value.

[0040] In practice, the element-wise operation formula (14) is used: (14); in, It is the first in the batch One sample, Indicates the lead time for forecasting. The UQCD module significantly reduces false high-confidence forecasts at locations with high model estimation uncertainty, while having almost no impact on high-confidence forecasts with sufficient observations. This reduction in false alarm rate is achieved at the cost of a small amount of sensitivity, directly addressing the sensitivity-accuracy dilemma observed in previous lightning nowcasting studies.

[0041] The UQ-STNet model architecture of this invention organically combines the long-range dependency modeling capability of the spatiotemporal Transformer with uncertainty quantification and confidence discounting mechanisms, achieving efficient spatiotemporal representation learning and low-false-probability lightning nowcasting from multi-source meteorological data. Specifically, the model employs multi-source inputs such as radar, lightning location, and automatic weather stations, aligning and fusing them on a unified grid. ST-Block is used to tokenize and positionally encode the rasterized feature sequences, and distance-constrained attention is introduced to aggregate key spatial context globally. This enables explicit characterization of the spatial organization and impact range of complex convective systems, avoiding over-response to isolated noise or local anomalies due to reliance solely on local convolution, thus structurally improving the spatial consistency and stability of forecasts.

[0042] Meanwhile, this invention introduces a UQCD module at the decoding output end, which simultaneously predicts the mean and variance of the probability distribution for each pixel, expanding the prediction from a deterministic output to a probabilistic expression with uncertainty characterization. Furthermore, it adaptively lowers the trigger probability in high uncertainty areas through a confidence discount strategy, enabling the model to actively "cool down" in scenarios prone to false alarms, such as weak lightning signals, strong background interference, or high observation noise. This significantly suppresses false alarms and reduces the false alarm rate while maintaining effective detection capabilities, thereby enhancing the credibility of the early warning results and their interpretability for risk decision-making.

[0043] Example 2: Please see Figure 2 and Figure 3 Based on the same inventive concept, this application also provides a method for training a lightning nowcasting model based on spatiotemporal Transformer and uncertainty perception, including: Acquire and preprocess multi-source meteorological observation data to obtain a multimodal training sample set under a unified spatiotemporal grid. The multi-source meteorological observation data includes at least radar data, lightning location data, and automatic weather station data. The multimodal training sample set is input into the uncertainty-aware spatiotemporal Transformer model to be trained. Multimodal fusion features are obtained through multimodal feature extraction and fusion. Based on the multimodal fusion features, spatiotemporal dependency modeling is performed through a geographic distance-aware biased Transformer attention mechanism to obtain spatial enhancement features. The spatial enhancement features are aggregated in the temporal dimension and spatial resolution is restored to generate nominal lightning occurrence probability fields under multiple forecast lead times. Uncertainty quantification is performed on the nominal lightning occurrence probability fields to obtain the prediction mean and variance for each spatial location. Based on the variance, the nominal probability of high uncertainty areas is discounted with confidence to obtain the discounted lightning occurrence probability field. The discounted lightning occurrence probability field is compared with the corresponding real lightning observation field to calculate the model prediction loss. With the goal of minimizing the model prediction loss, the parameters of the uncertainty-aware spatiotemporal Transformer model are optimized until the model converges to obtain the lightning nowcasting model.

[0044] Please see Figure 2 , Figure 2 The training architecture of the UQ-STNet model is shown. The training architecture of the model of this invention can be divided into four parts: input layer, model computation, training process, and output layer.

[0045] Multi-source data input (bottom of input layer): In the input layer, multi-source observation data is acquired and input, including automatic weather station data (AWS), lightning location data (LIG), and radar data (RADAR), to construct a multimodal spatiotemporal input sequence.

[0046] Data Preprocessing and Alignment (Middle of Input Layer): The AWS, LIG, and RADAR data are preprocessed. Preprocessing operations include outlier removal, grid interpolation, radar grid matching, and time alignment. The preprocessing of AWS, LIG, and RADAR data mainly includes the following steps: outlier removal, grid interpolation, radar grid matching, and time alignment. Outlier removal eliminates missing values, anomalous observations, or noise points; grid interpolation converts discrete station observation data into a regular grid field; radar grid matching spatially registers different data sources to a unified latitude and longitude grid; and time alignment unifies the data sources to the same sampling time series to avoid time aliasing and systematic bias caused by inconsistent sampling rates.

[0047] Unified format organization and dataset construction (upper part of input layer): The preprocessed and aligned multi-source raster data is converted into a unified binary raster file, and the dataset is divided based on the binary raster file to form a training set, a validation set and a test set for subsequent model training and inference.

[0048] Temporal slice input and network forward computation: Constructing multimodal raster slices at each time step as model input. , Figure 2 China and Israel This is an example, and the data is entered into UQ-STNet in chronological order.

[0049] The UQ-STNet performs spatiotemporal feature extraction and dependency modeling on the input sequence. Figure 2 UQ-STNet1…UQ-STNet t-1 This illustrates the continuous modeling process for multiple historical time steps.

[0050] Decoding and Autoregressive Feedback: The model internally sets up a decoder to aggregate the spatiotemporal enhancement features in the time dimension and output the prediction results of multiple forecast lead times. When performing rolling nowcasts, the decoder can adopt a cyclic autoregressive feedback method, using the prediction results of the previous moment or the previous lead time as one of the input information for subsequent predictions, so as to achieve continuously updated nowcasts.

[0051] Uncertainty Estimation and Confidence Discounting: An uncertainty perception module is set up at the output end to quantify the uncertainty of the pixel-level prediction results and discount high-confidence triggers in high-uncertainty regions, thereby reducing false alarms. Specifically, the uncertainty module is used to obtain the prediction center trend and dispersion, and performs adaptive weighting on the nominal probability based on the dispersion to obtain the discounted probability used for alarm determination.

[0052] Output results: In the output layer, a two-dimensional grid prediction probability is output, which is a pixel-level lightning occurrence probability field; in some implementations, information related to uncertainty can also be output simultaneously to form a forecast result with both probability and confidence, which can be used by the business platform for threshold determination, early warning area generation or risk classification display.

[0053] Please see Figure 3 , Figure 3 The training flowchart for the UQ-STNet model is shown below: S1. Acquisition and preprocessing of multi-source observation data, unified time series organization: This embodiment acquires raw data from three observation sources: radar data, lightning location data, and automatic weather station data, and organizes them into a sample sequence according to a unified time series. Furthermore, the raw data undergoes quality control preprocessing, including missing data identification, outlier removal, and necessary data consistency checks, to provide a reliable data foundation for subsequent spatial rasterization and temporal alignment.

[0054] 1) Radar data: Key derived quantities such as comprehensive reflectivity, echo top height and vertical liquid water content are extracted from volume scan data to characterize convection intensity, vertical development degree and water content, respectively; the derived quantities are preprocessed and combined into observation tensors that vary with time, and used as radar mode input.

[0055] 2) Lightning location data: Lightning events are spatiotemporally binned, mapped to regular grids, and the spatial location, temporal location, and discharge intensity of the events are tensorized and organized to form a lightning density field or intensity field arranged according to time steps, which serves as the input for lightning modes.

[0056] 3) Automatic weather station data: Acquire station elements such as temperature, maximum wind speed, and relative humidity to provide boundary layer thermal, dynamic, and water vapor constraint information, which serves as the modal input for automatic weather stations.

[0057] S2, Spatial rasterization and temporal alignment: To eliminate differences in resolution and sampling intervals among multi-source data, this embodiment performs unified latitude and longitude grid mapping on the three types of data, forming multimodal input slices at the same time, thereby avoiding systematic biases introduced by time aliases or spatial mismatches. Radar products generate corresponding time slices according to their scanning update rhythm, and then map them to the unified grid. Automatic station elements are mapped to the same grid through spatial interpolation to ensure pixel-by-pixel alignment with the radar grid. Lightning location events are first spatiotemporally binned, and then mapped to a lightning activity grid field of the same grid.

[0058] S3. Design the UQ-STNet model architecture (multimodal fusion, distance-constrained space Transformer, uncertain output): 3.1 Input layer and multimodal parallel access: At each moment, the radar grid field, lightning positioning grid field and automatic station grid field are used as parallel inputs. The three share a unified latitude and longitude grid in space and are aligned to a unified sampling rhythm in time.

[0059] 3.2 Multimodal encoder and feature fusion module: Set up corresponding encoding branches for each type of input, and map each modal raster data into a latent feature map; then concatenate the multi-path features in the channel dimension, and complete channel alignment and fusion through 1×1 convolution to obtain a fused feature map as the input of the subsequent Transformer.

[0060] 3.3 Distance-aware spatiotemporal Transformer module ST-Block: The fused features are unfolded into a spatial sequence by grid and injected with two-dimensional position encoding; global grid cell association is established through multi-head self-attention to achieve non-local information aggregation; and a distance bias obtained by the great circle distance of latitude and longitude is superimposed on the attention score to suppress physically unreasonable long-distance association, thereby reducing the spread of false alarms caused by spurious correlations; the Transformer output is then restored to raster form as the subsequent decoding input.

[0061] 3.4 Temporal Decoding and Resolution Restoration: The decoder is set to aggregate spatial enhancement features along the time dimension, and the spatial resolution is gradually restored through an upsampling structure, so that the output returns to the original grid size, thereby obtaining a pixel-level, multi-lead lightning occurrence probability field.

[0062] 3.5 Uncertainty Quantification and Confidence Discount Output: An Uncertainty Awareness Module (UQCD) is set at the decoding output end. For each pixel, parameters are output to characterize the predicted central trend and parameters to characterize the predicted dispersion. Adaptive confidence discount and pruning are performed on the nominal probability based on the uncertainty magnitude to suppress overconfident alarms in ambiguous areas and reduce false alarms.

[0063] S4. Model Training and Optimization: This invention constructs a supervisory signal (sample label) using a lightning activity grid field and employs the Adam optimizer for parameter updates. The loss function is based on binary cross-entropy loss and can be combined with uncertainty-related terms or regularization terms to achieve more robust training. By adjusting key hyperparameters such as the learning rate and discount strength, the model achieves a balance between detection capability and false alarm control. Simultaneously, it can monitor false alarm-related indicators based on the validation set to avoid overfitting or a resurgence of false alarms in the later stages of training.

[0064] S5, Model Output and Threshold Alarm Application: After training, during the inference phase, a pixel-level lightning occurrence probability field with multiple lead times is output, along with uncertainty-related information. Further, the nominal probability is discounted based on the uncertainty to obtain discounted probability raster data. In operational applications, this discounted probability raster data can be used for threshold alarm determination, generating gridded warning areas or risk level maps. Simultaneously, the uncertainty map can be used to identify high-risk but unreliable areas, providing auxiliary data for on-duty analysis, alarm interpretation, and threshold adjustment.

[0065] Please see Figure 4 and Figure 5 The UQ-STNet of the present invention includes a multimodal encoder, a feature fusion module, a distance-aware spatiotemporal Transformer module ST-Block, a decoder, and an uncertainty-aware module UQCD.

[0066] The multi-source raster inputs are processed by their respective encoders to extract features, which are then aligned and fused in the feature fusion module to obtain a unified feature tensor. Then, the ST-Block is input for spatiotemporal representation learning; the output of UQ_STNet is upsampled by the decoder to restore the spatial resolution and obtain the multi-advance prediction sequence. And in the uncertainty perception module UQCD, the probability center trend and the degree of dispersion are obtained simultaneously. This leads to a confidence discount, which is then used to form the final discounted probability for alarms. Finally, the rasterization probability result is output.

[0067] like Figure 4As shown, the multimodal encoder includes a radar encoder, an automatic weather station encoder, and a lightning location encoder. For any modal input raster tensor, local texture and intensity variation features are first extracted through at least two layers of 2D convolution (Conv2D). Then, the long-range spatial dependency within the modality is modeled using a spatiotemporal Transformer, outputting the corresponding modal feature representation, denoted as: radar modal features. Modal characteristics of automatic weather stations Lightning location mode characteristics .like Figure 5 As shown, the feature fusion module performs convolutional projection alignment on the three modal features respectively: The aligned features are then fused along the channel dimension to obtain a unified fused feature tensor. and will As input to subsequent UQ_STNet.

[0068] like Figure 4 As shown, at the current time step, excluding fused features In addition, spatial information at the current time step can be introduced as auxiliary input. This spatial information is preprocessed by convolutional layers Conv_1 and Conv_2 and then... Both are input into UQ_STNet to achieve joint modeling of the current time step and the historical sequence. In some implementations, an autoregressive loop mechanism is used: the prediction output from the previous time step is used... Feedback information is used in the next step of inference, namely the "autoregressive loop" (dashed line in the diagram), thus achieving rolling nowcasting. The output of UQ_STNet undergoes resolution restoration and pixel-by-pixel mapping via the decoder. The decoder includes two levels of deconvolutional / transposed convolutional layers (Conv_2DTranspose) and two-dimensional convolutional layers (Conv_2D), and uses Cropping3D to crop or align the output to the target grid range, ultimately obtaining the predicted sequence output. In multi-leadership forecasting scenarios, the decoder output can be represented as the prediction results of multiple lead times. And in subsequent modules, they are combined to form tensors. .

[0069] like Figure 5 As shown, the uncertainty perception module receives the predicted feature tensor output by the decoder. A parameterized uncertainty head is constructed using Dropout and Conv2D, outputting the central tendency parameter and the dispersion parameter, denoted as follows: and Based on the above and The predicted probability is adjusted for uncertainty confidence discounting, and the final discounted probability used for business alarm determination is obtained using formula (13). Furthermore, the module can perform a log-odds transformation on the probability: (15); in, This is a numerically stable term; and the result is adjusted to a raster tensor consistent with the output interface through Reshape to form the final output.

[0070] In terms of system integration, the output results of this invention can be written into a structured file or interface service in chronological order for the early warning platform to call and display; it can also be connected with the existing early warning workflow for real-time rolling updates of lightning near-term forecasts.

[0071] Example 3: Please see Figure 6 Based on the same inventive concept, this application also provides a lightning nowcasting system based on spatiotemporal Transformer and uncertainty perception, comprising: The acquisition module is used to acquire multi-source meteorological observation data, which includes at least radar data, lightning location data, and automatic weather station data. The preprocessing module is used to preprocess and spatiotemporally align the multi-source meteorological observation data to generate a multimodal input sequence under a unified spatiotemporal grid. The data processing module loads a spatiotemporal Transformer model with uncertainty awareness. The model includes at least a multimodal encoder, a feature fusion module, a distance-constrained spatiotemporal Transformer module ST-Block, a decoder, and an uncertainty awareness module UQCD. The multimodal input sequence is input into a trained uncertainty-aware spatiotemporal Transformer model. The multimodal encoder and feature fusion module are used to perform multimodal feature extraction and fusion on the multimodal input sequence. The spatiotemporal Transformer module ST-Block is used to perform spatiotemporal dependency modeling on the fused features based on a distance-constrained attention mechanism, and outputs a spatially enhanced feature representation. The decoder is used to perform temporal dimension aggregation and spatial resolution restoration on the spatially enhanced feature representation to generate a nominal lightning occurrence probability field under at least one forecast lead. The uncertainty-aware module UQCD is used to quantize the uncertainty of the nominal lightning occurrence probability field, obtain the prediction mean and variance corresponding to each spatial location, and perform confidence discounting on the nominal probability of high uncertainty areas based on the variance, and output the final lightning occurrence probability field used for early warning. The output module is used to determine and output lightning forecast information based on the lightning occurrence probability field.

[0072] Example 4: Based on the same inventive concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the lightning proximity forecasting method based on spatiotemporal Transformer and uncertainty perception as described above.

[0073] Example 5: To validate the effectiveness of the UQ-STNet model, we evaluated and compared the performance of several baseline models using the same learning rate. Evaluation metrics included Threat Score (TS), Equal Weight Threat Score (ETS), Detection Rate (POD), and False Alarm Rate (FAR). Figure 7 As shown, UQ-STNet performs better overall in terms of comprehensive skills and detection capabilities, while maintaining a more robust level in false alarm control.

[0074] In terms of comprehensive evaluation metrics, UQ-STNet achieved a threat score of 0.3768, higher than ConvLSTM's 0.3235 and StepDeep's 0.3148, and significantly better than S-Mamba's 0.1971. The equally weighted threat score also showed a consistent trend, with UQ-STNet reaching 0.3401, exceeding ConvLSTM's 0.2986, StepDeep's 0.3060, and significantly higher than S-Mamba's 0.1321. These results indicate that UQ-STNet has a more stable advantage in overall forecasting skill, balancing hit rate and false alarm rate.

[0075] In terms of detection capability, UQ-STNet achieves a detection rate of 0.6619, superior to ConvLSTM's 0.5601, and also higher than S-Mamba's 0.6235 and StepDeep's 0.6516, indicating that it captures real lightning events more comprehensively and misses fewer false alarms. Meanwhile, in terms of false alarm rate, UQ-STNet's value is 0.3764, lower than ConvLSTM's 0.3947; in contrast, S-Mamba's false alarm rate reaches 0.7947 and StepDeep's reaches 0.6256, both significantly higher. This comparison shows that UQ-STNet improves the detection rate without introducing false alarm inflation, achieving more effective alarm suppression, thus offering a greater advantage in balancing high detection and low false alarm rates, which is crucial for business operations.

[0076] In summary, the above comparison shows that UQ-STNet, by introducing the spatial modeling capabilities of Transformer, more fully learns the spatial organization and long-range dependencies of convective systems. At the same time, by combining uncertainty quantification and confidence discounting mechanisms, it more cautiously adjusts the trigger probability in high uncertainty regions, enabling the model to effectively control the risk of false alarms while pursuing high detection rates.

[0077] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Although this application has disclosed preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make some modifications or alterations to the above-mentioned technical content to create equivalent embodiments without departing from the scope of the technical solution of this application. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of this application.

Claims

1. A lightning nowcasting method based on spatiotemporal Transformer and uncertainty perception, characterized in that, include: Acquire multi-source meteorological observation data, which includes at least radar data, lightning location data, and automatic weather station data; The multi-source meteorological observation data are preprocessed and spatiotemporally aligned to generate a multimodal input sequence under a unified spatiotemporal grid; The multimodal input sequence is input into the trained uncertainty-aware spatiotemporal Transformer model UQ-STNet. The model obtains multimodal fusion features through multimodal feature extraction and fusion, and performs spatiotemporal dependency modeling based on the multimodal fusion features through a geographic distance-aware biased Transformer attention mechanism to obtain spatially enhanced features. The spatial enhancement features are aggregated in the time dimension and the spatial resolution is restored to generate nominal lightning occurrence probability fields under multiple forecast lead amounts; the uncertainty of the nominal lightning occurrence probability fields is quantified to obtain the prediction mean and variance of each spatial location, and the nominal probability of high uncertainty areas is discounted based on the variance to obtain the discounted lightning occurrence probability field. The spatially enhanced features are obtained by performing spatiotemporal dependency modeling based on the multimodal fusion features through a geographic distance-aware biased Transformer attention mechanism, including: The multimodal fusion features are converted into a token sequence and injected with two-dimensional positional encoding. The token sequence is then processed by a Transformer layer containing multi-head self-attention and distance-aware bias to output the final token sequence, which is the spatial augmented feature. Among them, a distance-aware bias matrix is ​​introduced into the multi-head self-attention. To improve attention score calculation, it is expressed as: (6)； in, These represent the query, key, and value matrices obtained by linearly projecting the token sequence onto each attention head; Represents the bias matrix. For the dimension of the key, represents the attention scoring function, and Softmax represents the Softmax function; The introduction of distance-aware bias matrix ,include: definition express The latitude and longitude, and The Haversine distance is: （7）； in, For the Earth's radius, On a sphere and The great circle distance between them; Mapping the distance to a bias term is represented as follows: (8a)； (8b)； in, The average pairwise distance used for normalization, For normalized distance, It is a scalar hyperparameter; This indicates that the distance exceeds the sparse radius. Imposing harsh penalties at times; Indicates a geophysical mask; Based on the lightning occurrence probability field, lightning forecast information is determined and output.

2. The method according to claim 1, characterized in that, The multimodal fusion features obtained through multimodal feature extraction and fusion include: At the point of time The system receives data from multiple modalities. The data for each modality is preprocessed into a raster format. For each modality, a modality-specific encoder is used for feature extraction. The input data is mapped to a shared latent feature space, and a feature map is output. The feature maps extracted from multiple modalities are concatenated along the channel dimension, and then projected through a convolutional layer to finally output a unified multimodal fusion feature. .

3. The method according to claim 1, characterized in that, The process of aggregating the spatial enhancement features in the temporal dimension and restoring the spatial resolution generates multiple nominal lightning occurrence probability fields under forecast lead, including: Time decoder aggregates along the time axis And upsampled to through two stages of deconvolution. Output One forecast lead time: (10)； in, This represents the spatial feature sequence generated by ST-Block. It is a time decoder used to aggregate information along the time axis and output the lead amount. Feature map; and It is a two-dimensional deconvolution, used to progressively convert features from... Upsample back to the original mesh The final Convolution generates one for each grid cell , Output lead probability forecast .

4. The method according to claim 1, characterized in that, The uncertainty quantification of the nominal lightning occurrence probability field, obtaining the predicted mean and variance for each spatial location, and the confidence discounting of the nominal probability in high-uncertainty regions based on the variance, to obtain the discounted lightning occurrence probability field, includes: For lead time Decoding features ,use The convolutional head outputs the logit mean and log-variance, the nominal probability is the sigmoid of the mean, and the standard deviation is recovered from the log-variance. （11a）； （11b）； （11c）； in, It is the mean of the deterministic predictions output by the traditional model. Represents the logarithmic variance. Indicates the standard deviation of the forecast. This represents the sigmoid function. This represents the nominal probability of lightning occurring. Represents the numerically stable term; Treating the predicted probability as a random variable induced by Gaussian perturbations in the logit space, we obtain the logistic-normal model: Treating the predicted probability as a random variable induced by Gaussian perturbations in the logit space, following a logistic-normal distribution, it can be expressed as: （12）； in, Lead time The probability of random lightning occurrence, For a standard normally distributed random variable, Characterizing the central trend of the forecast, Control the degree of dispersion induced by the logistic-normal distribution; To suppress overconfidence alarms under high uncertainty, based on and The nominal probability is reduced in weight proportionally and then clipped to a smaller value. Interval: （13）； in, The probability, discounted for uncertainty, is ultimately used for alerting. It is a positive hyperparameter used to control uncertainty. The intensity of the penalty for high nominal probabilities; Limit the results to the range of valid probabilities, that is, let Then we have: ； In practice, the element-wise operation formula (14) is used: (14)； in, It is the first in the batch One sample, Indicates the lead time for forecasting. For spatial grid positions, Indicates the first Individual samples in predicting lead time Below, located in the spatial grid The nominal probability of lightning occurring at a given location. Indicates the first Individual samples in predicting lead time Below, located in the spatial grid The standard deviation of the prediction at that location.

5. A lightning nowcasting system based on spatiotemporal Transformer and uncertainty perception, characterized in that, To implement the method as described in claim 1, comprising: The acquisition module is used to acquire multi-source meteorological observation data, which includes at least radar data, lightning location data, and automatic weather station data. The preprocessing module is used to preprocess and spatiotemporally align the multi-source meteorological observation data to generate a multimodal input sequence under a unified spatiotemporal grid. The data processing module loads a spatiotemporal Transformer model with uncertainty awareness. The model includes at least a multimodal encoder, a feature fusion module, a distance-constrained spatiotemporal Transformer module ST-Block, a decoder, and an uncertainty awareness module UQCD. The multimodal input sequence is fed into the trained uncertainty-aware spatiotemporal Transformer model UQ-STNet. The multimodal encoder and feature fusion module are used to perform multimodal feature extraction and fusion on the multimodal input sequence. The spatiotemporal Transformer module ST-Block is used to perform spatiotemporal dependency modeling on the fused features based on a distance-constrained attention mechanism, and outputs a spatially enhanced feature representation. The decoder is used to perform temporal dimension aggregation and spatial resolution restoration on the spatially enhanced feature representation to generate a nominal lightning occurrence probability field under at least one forecast lead. The uncertainty-aware module UQCD is used to quantize the uncertainty of the nominal lightning occurrence probability field, obtain the prediction mean and variance corresponding to each spatial location, and perform confidence discounting on the nominal probability of high uncertainty regions based on the variance, and output the final lightning occurrence probability field used for early warning. The output module is used to determine and output lightning forecast information based on the lightning occurrence probability field.

Images