A thunderstorm path prediction method based on multi-source data and deep learning

Abstract

The application discloses a thunderstorm path prediction method based on multi-source data and deep learning, first, multi-source data is collected and preprocessed to ensure data quality, then, through a data fusion method based on a graph neural network, the multi-source data is integrated to extract more rich features, next, a convolutional neural network model is constructed to identify and track thunderstorm areas from the fused data, on the basis of thunderstorm identification and tracking, a recurrent neural network model is used to predict the state of future thunderstorms, and the movement path of the thunderstorm is found by optimizing an objective function, so that the purpose of predicting the thunderstorm path is achieved. The application can realize high-precision thunderstorm activity prediction, and has important practical application value for preventing and responding to thunderstorm disasters.

Description

Technical Field

[0001] This invention belongs to the field of meteorological observation. Background Technology

[0002] Thunderstorms are a common weather phenomenon occurring in small- to medium-scale convective systems. Often accompanied by lightning, thunderstorms can cause forest and oil depot fires, power supply and communication system failures or damage, and pose a significant threat to aerospace, mining, and some important and sensitive high-tech equipment. Since the 1980s, the hazards caused by lightning have increased significantly, particularly in fields closely related to high technology, such as aerospace, defense, communications, power, computers, and electronics. The widespread use of large-scale and very large-scale integrated circuits, which are extremely sensitive to lightning electromagnetic interference, has greatly increased the probability of lightning strikes. Therefore, the prediction and early warning of thunderstorm paths is of paramount importance.

[0003] Currently, there are several methods for predicting thunderstorm activity:

[0004] The first method, radar extrapolation, predicts the location and path of thunderstorms based on radar echo data from radar scans through extrapolation and interpolation. This method offers high real-time performance and spatial resolution; however, radar extrapolation is limited by radar coverage and echo analysis, potentially restricting its prediction range. A commonly used algorithm is TITAN (Thunderstorm Identification, Tracking, Analysis, and Nowcasting) (see "Research on Thunderstorm Path Tracking Based on Improved TITAN Algorithm" (Mechatronics Information, 2018)). TITAN is a method for storm tracking using a smooth linear trend model. It calculates the search radius based on the product of the maximum storm speed and the time interval, considering all possible matches between storms within a continuous time period. A cost function is constructed, transforming the tracking problem into a combinatorial optimization problem. However, in practical applications, the algorithm has high complexity and long computation time, potentially leading to forecast delays. Furthermore, TITAN forecasts rely on prior assumptions and parameter settings, and its forecast efficiency may be less than ideal when dealing with variable meteorological environments and complex thunderstorms.

[0005] The second method, numerical simulation, uses numerical weather models to simulate the atmosphere and determine the generation, evolution, and movement trajectory of thunderstorms ("Numerical simulation of the charge structure evolution characteristics of a thunderstorm cell in the mature stage during a squall line in Guangdong" (Meteorological Science (2020)). This method can provide predictions over a longer time range and takes into account factors such as atmospheric circulation and physical parameters. However, it is highly dependent on observational data, and model errors can affect the accuracy of the prediction results.

[0006] The third type of statistical machine learning method analyzes historical thunderstorm event data to construct statistical models or apply machine learning algorithms to predict future thunderstorm paths ("Deep Learning Lightning Forecasting Model Based on Radar Reflectivity Factor and Lightning Location Data" (Meteorological Science and Technology, 2022)). These methods can comprehensively consider multiple influencing factors and have a certain degree of predictive accuracy and flexibility, but they have certain requirements for the timeliness and quality of the measured data. Summary of the Invention

[0007] Purpose of the invention: In order to solve the problems existing in the prior art, this application provides a thunderstorm path prediction method based on multi-source data and deep learning.

[0008] Technical solution: This invention provides a thunderstorm path prediction method based on multi-source data and deep learning, specifically including the following steps:

[0009] Step 1: Collect thunderstorm-related data at time t, including radar data M. radar t Ground flash positioning data M lightning t Satellite data M sat t Meteorological sounding data M sonde t Surface meteorological observation data M surface t ;

[0010] Step 2: Preprocess the collected data;

[0011] Step 3: Treat each type of thunderstorm-related data collected in Step 1 as a node, and fuse the data of one node with other nodes to obtain the fused value of each node;

[0012] Step 4: Input the fused data into the CNN network to obtain the candidate thunderstorm region identification result Y at time t. t CNN ;

[0013] Step 5: From Y through thresholding and connectivity analysis t CNN Extract thunderstorm regions and calculate the state of the i-th thunderstorm region. The state of the thunderstorm region includes the center point position, area, and probability value output by the CNN; i = 1, 2, ..., I, where I represents the number of thunderstorm regions at time t.

[0014] Step 6: Use the state sequence of the thunderstorm region in the time series [tn, t-n+1, ​​..., t] as input to the RNN model to obtain the center point location, area, and probability value of the thunderstorm region at time t+1, thus obtaining the state of the thunderstorm region at time t+1; n represents the length of the time series;

[0015] Step 7: Using the thunderstorm area at time t as the starting point and the thunderstorm area at time t+1 as the ending point, calculate the cost of the path from the starting point to the ending point, and select the path with the minimum cost as the thunderstorm path.

[0016] Furthermore, in step 3, the following formula is used for fusion:

[0017]

[0018] in, Let σ represent the value of the i'-th node after the (k+1)th iteration fusion, σ represent the activation function, N represent the total number of nodes, and W represent the value of the node after the (k+1)th iteration fusion. i，j， Let A(.) represent the weight of the edge between node i' and node j', and let A(.) represent the information integration function. This represents the value of the i'-th node after k iterations of fusion. Let i' represent the value of the j'-th node after k iterations of fusion. Let i' = 1, 2, 3, 4, 5. When i' = 1, it represents the node corresponding to radar data; when i' = 2, it represents the node corresponding to ground flash positioning data; when i' = 3, it represents the node corresponding to satellite data; when i' = 4, it represents the node corresponding to meteorological sounding data; and when i' = 5, it represents the node corresponding to ground meteorological observation data. Let N = 5.

[0019] Furthermore, in step 5, threshold processing and connectivity analysis are performed on Y... t CNN Extracting the thunderstorm region specifically involves: [Y] t CNN Candidate thunderstorm regions with intensity greater than a preset threshold are assigned a value of 1, while candidate thunderstorm regions with intensity less than or equal to the preset threshold are assigned a value of 0, thus obtaining a binary map of the candidate thunderstorm regions. In the binary map, adjacent pixels with values ​​all equal to 1 are divided into a connected region, and each connected region is taken as a thunderstorm region.

[0020] Furthermore, the squared difference loss is used as the loss function L for the RNN model, and the specific expression is as follows:

[0021]

[0022] Where T represents the number of training samples, with one training sample corresponding to one time step. The cost function predicted by the RNN model. For the actual cost function, and The expression is as follows:

[0023]

[0024]

[0025] Where w1, w2, and w3 are all weights, This represents the change in the coordinates of the center point of the i-th thunderstorm region in the t'-th training sample to the coordinates of the j-th thunderstorm region in the t'+1-th training sample predicted by the model. This represents the change in area from the i-th thunderstorm region of the t'-th training sample to the area of ​​the j-th thunderstorm region of the t'+1-th training sample predicted by the model; This represents the change in the intensity of the i-th thunderstorm region in the t'-th training sample compared to the probability value of the j-th thunderstorm region predicted by the model in the t'+1-th training sample. This represents the change in the coordinates of the center point of the i-th thunderstorm region in the t'-th training sample to the coordinates of the center point of the j-th thunderstorm region in the actual t'+1-th training sample; This represents the change in area from the i-th thunderstorm region of the t'-th training sample to the j-th thunderstorm region of the actual t'+1-th training sample; This represents the change in the intensity of the i-th thunderstorm region in the t'-th training sample to the probability value of the j-th thunderstorm region in the actual t'+1-th training sample.

[0026] Furthermore, in step 7, the cost of the path from the starting point to the ending point is calculated using the following formula:

[0027]

[0028] Where R represents the total number of thunderstorm areas at time t, and M represents the total number of thunderstorm areas at time t+1. The expression is:

[0029]

[0030] Where w1, w2, and w3 are all weights. The change in the coordinates of the center point of the i-th thunderstorm region at time t to the coordinates of the center point of the j-th thunderstorm region predicted by the RNN model at time t+1; This represents the change in the area of ​​the i-th thunderstorm region at time t to the area of ​​the j-th thunderstorm region predicted by the RNN model at time t+1. Let be the change in intensity of the i-th thunderstorm region at time t to the probability value of the j-th thunderstorm region predicted by the RNN model at time t+1.

[0031] Beneficial effects:

[0032] 1. This invention employs a multi-source data fusion technology based on graph neural networks, which can effectively integrate and extract useful information from multiple sources of data, such as radar data, ground flash location data, satellite data, meteorological sounding data, and ground meteorological observation data.

[0033] 2. This invention introduces convolutional neural networks and recurrent neural networks to construct a model that can automatically learn complex features in data and achieve high-precision thunderstorm identification and prediction.

[0034] 3. This invention defines a thunderstorm state including thunderstorm location, area and intensity, and constructs a cost function to measure the change of thunderstorm state. By optimizing the weights in the cost function, the model can more accurately predict the change of thunderstorm state.

[0035] 4. This invention proposes an optimal thunderstorm activity prediction algorithm based on an optimization objective function, which can find the path that minimizes the total cost as the movement path of the thunderstorm and effectively predict the movement path of the thunderstorm.

[0036] 5. This invention proposes a data-driven weight selection and optimization technique. By defining and minimizing the loss function, the model parameters can be automatically selected and optimized, giving the model good universality and robustness. Attached Figure Description

[0037] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0039] like Figure 1 The embodiment shown provides a method for predicting thunderstorm activity, and the specific steps are as follows:

[0040] Step 1: Preparation and preprocessing of thunderstorm-related data.

[0041] Collect multi-source data, including radar data, ground flash location data, satellite data, meteorological sounding data, and ground meteorological observation data.

[0042] Radar data: This data mainly consists of radar reflectivity (R). tRadial Velocity (V) t Spectrum Width W t Composition, these data can be represented as matrix M radar t .

[0043] Lightning location data: mainly lightning data, including the location of the lightning (x). lightning t Time t lightning t Intensity I lightning t polarity P lightning t We can represent it as matrix M lightning t .

[0044] Satellite data: including cloud image C t Temperature T sat t Humidity H sat t Wind direction D sat t Wind speed S sat t It can be represented as matrix M sat t .

[0045] Meteorological sounding data: This data includes temperature T sonde t Humidity H sonde t air pressure P sonde t Wind direction D sonde t Wind speed S sonde t It can be represented as matrix M sonde t .

[0046] Surface meteorological observation data: This data includes temperature T surface t Humidity H surface t air pressure P surface t Wind direction D surface t Wind speed S surface t Precipitation P rain t It can be represented as matrix M surface t .

[0047] In the above, t represents a point in time, and all matrices are defined at time t.

[0048] Perform data preprocessing:

[0049] ① Data cleaning: Remove invalid or erroneous data. Check the logical consistency of the data. For example, if the area of ​​a thunderstorm suddenly changes from 100 square kilometers to 1 square kilometer, it could be due to measurement error or data entry error.

[0050] ② Fill missing values: Use interpolation or statistical methods to fill missing data.

[0051] Step 2: Multi-source data fusion.

[0052] This invention proposes a data fusion method based on Graph Neural Networks (GNNs) to enhance the accuracy and stability of thunderstorm identification and prediction by integrating multi-source data, including radar data, ground flash location data, satellite data, meteorological sounding data, and surface meteorological observation data. This method can more effectively extract useful information from different data sources and integrate this information efficiently. In this method, each data source is first treated as a node, and the data in each node is considered as the features of that node. Then, the weights of the edges between nodes are defined to represent the correlation between data sources. This results in a graph structure that describes the relationships between the data sources.

[0053] First, each data source is represented as a node. For each node, the data within it is used as its feature. Therefore, for a radar data node, its feature is M. radar t For ground flash positioning data nodes, the characteristic is M lightning t For satellite data nodes, the characteristic is M. sat t For meteorological sounding data nodes, the characteristic is M sonde t For ground meteorological observation data nodes, the characteristic is M surface t Therefore, there are five nodes at time t.

[0054] Next, the weights of the edges between nodes are defined to represent the relationships between data sources. Assuming there are N data sources, the weight of the edge between node i' and node j' is represented as W. i，j， The features of node i', after k rounds of updates, are represented as follows: Therefore, the information propagation and integration process in a graph neural network can be represented as:

[0055]

[0056]

[0057] in, X represents the value of the i'-th node after the (k+1)th iteration fusion. i’ The feature value of the i'th node, σ represents the activation function, N represents the total number of nodes, and W i，j， The weight of the edge between node i' and node j', where A(.) represents the information integration function. This represents the value of the i'-th node after k iterations of fusion. Let i' represent the value of the j'-th node after k iterations of fusion, where i' = 1, 2, 3, 4, 5. When i' = 1, it represents the node corresponding to radar data; when i' = 2, it represents the node corresponding to ground flash positioning data; when i' = 3, it represents the node corresponding to satellite data; when i' = 4, it represents the node corresponding to meteorological sounding data; and when i' = 5, it represents the node corresponding to ground meteorological observation data. N = 5.

[0058] After K rounds of information dissemination and integration, the final fused data is obtained. K represents the maximum number of iterations. This fused data includes information from both the original data and the relationships between different data sources.

[0059] Step 3: Initial identification and tracking of thunderstorms.

[0060] Using the fused data, a convolutional neural network (CNN) model is constructed. This model can automatically learn complex features in the data, thereby more accurately identifying thunderstorm regions. In this context, the CNN acts as an image segmentation network, and its output is a matrix of the same size as the input image, where each value represents the probability that the corresponding pixel belongs to a thunderstorm region. The fused data is input to obtain the identification results of candidate thunderstorm regions, denoted as Y. t CNN .

[0061] Then, through thresholding and connectivity analysis, from Y t CNN Multiple thunderstorm regions were identified. For each identified possible thunderstorm region e, its center position (x) was calculated. t e ,y t e ), area s t e And the probability value I based on the CNN output. t eThis gives us the state S of the thunderstorm region at time t. t e =(x t e ,y t e ,s t e ,I t e ).

[0062] For thresholding, a threshold value, such as 0.5, is typically set. Candidate regions with an output probability greater than 0.5 are identified as potential thunderstorm regions and assigned a value of 1, while candidate regions with an output probability less than 0.5 are identified as non-thunderstorm regions and assigned a value of 0. This transforms the probability map into a binary map. Connectivity analysis is performed on the binary map. All adjacent pixels with a value of 1 are grouped into connected regions, and each connected region is identified as a thunderstorm region. For each connected region, its center position, area, and average intensity value can be calculated. This yields the state of the thunderstorm region at time t.

[0063] Next, the identified thunderstorm states are used as input to the RNN model. The RNN model takes the thunderstorm state sequence as input and outputs a prediction of the thunderstorm state Y for future time intervals. t RNN ,

[0064] Step 4: State definition and cost function construction.

[0065] Define the state of each thunderstorm region i at time t as S. t i =(x t i ,y t i ,s t i ,I t i ), where x t i and y t i Indicates the location of the thunderstorm, s t i I represents the area of ​​a thunderstorm. t i This represents the average probability value of a thunderstorm area.

[0066] The output Y of the RNN model t RNNIn essence, it's a prediction of the state of each thunderstorm region over a future period. This output can be further transformed into predictions of the future location, area, and intensity of the thunderstorm region. This yields the thunderstorm state at time t+1.

[0067] Then, construct a cost function C. ij t To measure thunderstorms from state S t i Change to state S t+1 j The cost. This cost function can be expressed as:

[0068]

[0069] Where w1, w2, and w3 are all weights. The change in the coordinates of the center point of the i-th thunderstorm region at time t to the coordinates of the center point of the j-th thunderstorm region predicted by the RNN model at time t+1; This represents the change in the area of ​​the i-th thunderstorm region at time t to the area of ​​the j-th thunderstorm region predicted by the RNN model at time t+1. Let be the change in intensity of the i-th thunderstorm region at time t to the probability value of the j-th thunderstorm region predicted by the RNN model at time t+1.

[0070] Step 5: Weight selection and optimization.

[0071] In the model of this invention, weights w1, w2, and w3 determine the relative importance of positional, area, and intensity changes in the cost function of thunderstorm state changes. The selection of these weights typically depends on the characteristics of the actual problem; this invention proposes an automatic selection method driven by data.

[0072] This embodiment defines a loss function to measure the difference between the cost predicted by the model and the actual cost. For example, a common loss function is the squared error loss:

[0073]

[0074] Where R represents the number of training samples, with one training sample corresponding to one time step. The cost function predicted by the RNN model. For the actual cost function, and The expression is as follows:

[0075]

[0076]

[0077] Where w1, w2, and w3 are all weights, This represents the change in the coordinates of the center point of the i-th thunderstorm region in the t'-th training sample to the coordinates of the j-th thunderstorm region in the t'+1-th training sample predicted by the model. This represents the change in area from the i-th thunderstorm region of the t'-th training sample to the area of ​​the j-th thunderstorm region of the t'+1-th training sample predicted by the model; This represents the change in the intensity of the i-th thunderstorm region in the t'-th training sample compared to the probability value of the j-th thunderstorm region predicted by the model in the t'+1-th training sample. This represents the change in the coordinates of the center point of the i-th thunderstorm region in the t'-th training sample to the coordinates of the center point of the j-th thunderstorm region in the actual t'+1-th training sample; This represents the change in area from the i-th thunderstorm region of the t'-th training sample to the j-th thunderstorm region of the actual t'+1-th training sample; This represents the change in the intensity of the i-th thunderstorm region in the t'-th training sample to the probability value of the j-th thunderstorm region in the actual t'+1-th training sample.

[0078] i = 1, 2, ..., R, j = 1, 2, ..., M, where R represents the total number of thunderstorm regions in the t'th training sample, and M represents the total number of thunderstorm regions in the t'+1th training sample.

[0079] Then, the weights can be selected by optimizing this loss function. Specifically, the goal is to find a set of weights (w1, w2, w3) that minimizes the loss function:

[0080]

[0081] This optimization problem can be solved using the gradient descent algorithm.

[0082] Step 6: Minimize the objective function.

[0083] In the task of predicting thunderstorm activity, a path is found that minimizes the sum of the cost functions of the thunderstorm region at time t and the thunderstorm region at time t+1. This embodiment uses dynamic programming or other optimization algorithms to find such a path.

[0084] The expression for the objective function is:

[0085]

[0086] Where R is the number of starting points and M is the number of ending points. It is the cost of moving from point i to point j, including changes in the location, area, and intensity of the thunderstorm.

[0087] Step 7: Thunderstorm path prediction.

[0088] After obtaining the minimized objective function Q, the path corresponding to Q is selected as the predicted thunderstorm movement path. Assume there is a set of paths {P1, P2, ..., P...} K The corresponding total cost is {Q1, Q2, ..., Q}. K The path with the smallest Q-value is selected as the predicted thunderstorm path, i.e.:

[0089]

[0090] Among them, P pred It is the predicted optimal path, Q k It is path P k The corresponding total cost, This means selecting the path that minimizes the total cost Q among all paths.

[0091] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A thunderstorm path prediction method based on multi-source data and deep learning, characterized in that, Specifically, the steps include the following: Step 1: Collect thunderstorm-related data at time t, including radar data M. radar t Ground flash positioning data M lightning t Satellite data M sat t Meteorological sounding data M sonde t Surface meteorological observation data M surface t ; Step 2: Preprocess the collected data; Step 3: Treat each type of thunderstorm-related data collected in Step 1 as a node, and fuse the data of one node with other nodes to obtain the fused value of each node; Step 4: Input the fused data into the CNN network to obtain the candidate thunderstorm region identification result Y at time t. t CNN ; Step 5: From Y through thresholding and connectivity analysis t CNN Extract thunderstorm regions and calculate the state of the i-th thunderstorm region. The state of the thunderstorm region includes the center point position, area, and probability value output by the CNN; i = 1, 2, ..., I, where I represents the number of thunderstorm regions at time t. Step 6: Use the state sequence of the thunderstorm region in the time series [tn, t-n+1, ​​..., t] as input to the RNN model to obtain the center point location, area, and probability value of the thunderstorm region at time t+1, thus obtaining the state of the thunderstorm region at time t+1; n represents the length of the time series; Step 7: Using the thunderstorm area at time t as the starting point and the thunderstorm area at time t+1 as the ending point, calculate the cost of the path from the starting point to the ending point, and select the path with the minimum cost as the thunderstorm path.

2. The thunderstorm path prediction method based on multi-source data and deep learning according to claim 1, characterized in that, The following formula is used for fusion in step 3: in, Let σ represent the value of the i'-th node after the (k+1)th iteration fusion, σ represent the activation function, N represent the total number of nodes, and W represent the value of the node after the (k+1)th iteration fusion. i，j， Let A(.) represent the weight of the edge between node i' and node j', and let A(.) represent the information integration function. This represents the value of the i'-th node after k iterations of fusion. Let i' represent the value of the j'-th node after k iterations of fusion. Let i' = 1, 2, 3, 4, 5. When i' = 1, it represents the node corresponding to radar data; when i' = 2, it represents the node corresponding to ground flash positioning data; when i' = 3, it represents the node corresponding to satellite data; when i' = 4, it represents the node corresponding to meteorological sounding data; and when i' = 5, it represents the node corresponding to ground meteorological observation data. Let N = 5.

3. The thunderstorm path prediction method based on multi-source data and deep learning according to claim 1, characterized in that, In step 5, threshold processing and connectivity analysis are used to analyze Y. t CNN Extracting the thunderstorm region specifically involves: [Y] t CNN Candidate thunderstorm regions with intensity greater than a preset threshold are assigned a value of 1, while candidate thunderstorm regions with intensity less than or equal to the preset threshold are assigned a value of 0, thus obtaining a binary map of the candidate thunderstorm regions. In the binary map, adjacent pixels with values ​​all equal to 1 are divided into a connected region, and each connected region is taken as a thunderstorm region.

4. The thunderstorm path prediction method based on multi-source data and deep learning according to claim 1, characterized in that, The squared difference loss is used as the loss function L for the RNN model, and the specific expression is as follows: Where T represents the number of training samples, with one training sample corresponding to one time step. The cost function predicted by the RNN model. For the actual cost function, and The expression is as follows: Where w1, w2, and w3 are all weights, This represents the change in the coordinates of the center point of the i-th thunderstorm region in the t'-th training sample to the coordinates of the j-th thunderstorm region in the t'+1-th training sample predicted by the model. This represents the change in area from the i-th thunderstorm region of the t'-th training sample to the area of ​​the j-th thunderstorm region of the t'+1-th training sample predicted by the model; This represents the change in the intensity of the i-th thunderstorm region in the t'-th training sample compared to the probability value of the j-th thunderstorm region predicted by the model in the t'+1-th training sample. This represents the change in the coordinates of the center point of the i-th thunderstorm region in the t'-th training sample to the coordinates of the center point of the j-th thunderstorm region in the actual t'+1-th training sample; This represents the change in area from the i-th thunderstorm region of the t'-th training sample to the j-th thunderstorm region of the actual t'+1-th training sample; This represents the change in the intensity of the i-th thunderstorm region in the t'-th training sample to the probability value of the j-th thunderstorm region in the actual t'+1-th training sample.

5. The thunderstorm path prediction method based on multi-source data and deep learning according to claim 1, characterized in that, In step 7, the cost of the path from the starting point to the ending point is calculated using the following formula: Where R represents the total number of thunderstorm areas at time t, and M represents the total number of thunderstorm areas at time t+1. The expression is: Where w1, w2, and w3 are all weights. The change in the coordinates of the center point of the i-th thunderstorm region at time t to the coordinates of the center point of the j-th thunderstorm region predicted by the RNN model at time t+1; This represents the change in the area of ​​the i-th thunderstorm region at time t to the area of ​​the j-th thunderstorm region predicted by the RNN model at time t+1. Let be the change in intensity of the i-th thunderstorm region at time t to the probability value of the j-th thunderstorm region predicted by the RNN model at time t+1.

Images