Lightning early warning model optimization method of deep learning algorithm

Abstract

The invention discloses a thunder early warning model optimization method based on a deep learning algorithm. The thunder early warning model optimization method comprises the steps that five risk early warning levels are divided according to the distance between a thunder occurrence position and a target point, namely six corresponding early warning levels are obtained; selecting lightning positioning, radar extrapolation results and atmospheric electric field instrument data around the target point from the sample data set; and risk grading preprocessing is carried out on the thunder early warning probabilities and critical thresholds of the three early warning factors. The method has the following advantages that the precision is high, analysis and modeling are carried out through comprehensive multi-source data, information related to thunder and lightning occurrence and development can be captured more comprehensively, and therefore the precision of early warning is improved. Data can be obtained and processed in real time, early warning information can be updated in time, and the requirement of a user for the real-time performance is met. Personalized customization: personalized risk grading and early warning setting can be carried out according to characteristics of different target point locations and user requirements.

Description

Technical Field

[0001] The present invention relates to the field of lightning warning using deep learning algorithms, and more specifically to a method for optimizing a lightning warning model using a deep learning algorithm. Background Art

[0002] Lightning is a very destructive natural phenomenon that can cause serious damage to power systems, communication facilities, aviation, building structures, and human life and property. Globally, there are countless fires, equipment damage, casualties, and other accidents caused by lightning every year, which have caused huge losses to the society and economy. Traditional lightning warning methods are mainly based on simple threshold judgments or empirical models of single or a few monitoring methods such as atmospheric electric field meters, lightning location systems, and radars. For example, atmospheric electric field meters warn of lightning by monitoring changes in the intensity of the atmospheric electric field, and issue a warning signal when the electric field intensity exceeds the preset threshold. However, this method has obvious limitations. It can only respond when lightning activity has already had a significant impact on the local electric field, and the advance warning is limited. Although the lightning location system can determine the location of lightning, it is difficult to accurately predict the movement path and intensity changes of lightning, and it is impossible to effectively warn the areas that may be affected for a long time in advance. Although radar can monitor the dynamics of meteorological clouds, it is insufficient to capture the microscopic characteristics of lightning activity, and under complex meteorological conditions, the accuracy and timeliness of lightning warnings are greatly reduced.

[0003] The current lightning warning model has the following problems:

[0004] (1) Time series information processing problem: Lightning warning involves data that changes over time, such as lightning location, radar data, and atmospheric electric field. Traditional methods may find it difficult to fully utilize the long-term dependencies in these time series information.

[0005] (2) Lightning warning feature extraction and heterogeneous data fusion problem: Different warning factors (such as lightning location, radar data and atmospheric electric field) have their own characteristics. How to effectively extract and fuse these features is crucial to improving the accuracy of warnings.

[0006] (3) Overfitting problem of deep learning models: When training complex deep learning models, overfitting is prone to occur, resulting in a decrease in the generalization ability of the model in practical applications.

[0007] (4) The problem of warning level classification affecting the accuracy of warning: The existing lightning warning level classification may not be detailed or accurate enough, which affects the actual effect of the warning. Summary of the invention

[0008] Therefore, in order to solve the above-mentioned shortcomings, the present invention provides a lightning warning model optimization method based on a deep learning algorithm.

[0009] The present invention is implemented by constructing a lightning warning model optimization method of a deep learning algorithm, which is specifically as follows;

[0010] According to the distance between the lightning strike location and the target point, it is divided into five risk warning levels (i.e. 50km is type 1, 40km is type 2, 30km is type 3, 20km is type 4, 10km is risk 5; the warning is lifted or there is no lightning safety risk is type 0), which corresponds to six warning levels;

[0011] The lightning location around the target point, radar extrapolation results, and atmospheric electric field instrument data were selected from the sample data set; first, the lightning warning probability and critical threshold of the three warning factors were pre-processed for risk classification;

[0012] Construct the TCN-GRU-MATT model structure, including:

[0013] (1) Input 1: Input forecast factors, which are lightning location, radar, and atmospheric electric field risk classification data every 6 minutes when there is a thunderstorm weather process;

[0014] (2) Temporal convolutional network layer: set up 2 layers of residual units, each layer of residual units contains 2 convolution units and 1 nonlinear mapping; the convolution unit uses the ReLU function as the activation function to normalize the weight of the convolution kernel; the convolution kernel size is 2; and the Dropout coefficient is set to 0.4. The Dropout setting can randomly select some neurons to inactivate to prevent overfitting in training and speed up the convergence of the model; set the expansion coefficient to (1, 2, 4, 8, 16, 32); the filter is 128; the input and output of the residual unit have different dimensions and cannot be directly added. A 1×1 convolution layer is added to the residual mapping for dimensionality reduction;

[0015] (3) Input 2: Input classification labels, which are the six warning level results corresponding to every 6 minutes;

[0016] (4) Gated recurrent unit layer: Set up a 3-layer GRU unit. First, connect the output g of the TCN model with the output x2 of input 2. It is necessary to ensure that the dimension and number of samples of the data are the same, and obtain the data c. t ;

[0017] c t =f(x (2，t) , g (t) )

[0018] Where: f represents the connection operation, which is implemented using the concatenate function of the keras model to connect the temporal convolutional network and input 2 as the input of the gated recurrent unit;

[0019] (5) Self-attention mechanism layer: Use the self-attention mechanism to weight the output of the GRU model and select the features that have the greatest impact on the prediction results;

[0020] (6) Integrate the weighted prediction results of all subsequences to obtain the final prediction result.

[0021] According to the method for optimizing a lightning warning model using a deep learning algorithm of the present invention, the method is characterized in that the TCN neural network layer in step (4) is a simple and universal convolutional neural network architecture for solving time series problems; the TCN model is composed of a group of residual units, each of which is a small neural network with residual connections. The residual connections can accelerate the feedback and convergence of the deep network and solve the "degeneration phenomenon" caused by the increase in network levels; the residual connection is to add the input x of the residual unit to the output f(x), and the network is forced to be modeled as follows:

[0022] f(x)=h(x)-x

[0023] The residual unit contains two convolution units and nonlinear mapping. First, a one-dimensional dilated causal convolution is performed in the convolution unit. The sampling interval is adjusted by the dilation coefficient to achieve a larger receptive field (RF), that is, the area that the features on the convolution layer can see, so that the network can remember long enough historical information and only convolve the input before time t to obtain the output at time t to ensure that future information will not be leaked. Then the weights are normalized and the ReLU function is used as the activation function. Finally, the Dropout operation is used to randomly discard neurons with a certain probability to prevent overfitting and accelerate the model training speed. Nonlinear mapping is to reduce the dimension of high-dimensional data when the input and output of the residual unit have different dimensions. The operation formula of the one-dimensional dilated causal convolution is:

[0024]

[0025] Where: x is the input sequence; f is the filter; d is the dilation coefficient; k is the convolution kernel size; s-di ensures that convolution operations can only be performed on past inputs.

[0026] According to the method for optimizing a lightning warning model of a deep learning algorithm of the present invention, it is characterized in that in step (4), the GRU is a simplified version of the LSTM, which integrates the input gate and the forget gate of the LSTM into an update gate, has no output gate, but adds a reset gate; the update gate controls the degree of retention of the state information at the previous moment, and the larger the value of the update gate, the greater the influence of the state information at the previous moment on the current state; the reset gate controls the degree of integration of the current state with the previous information, and the smaller the value, the more information is ignored; the calculation formula of h(t) in the GRU unit is:

[0027]

[0028] Where: x t is the input at time t; h t is the output or state at time t; h t-1 is the state at time t-1; w is the weight; σ is the activation function sigmoid(·); tanh(·) is the activation function.

[0029] According to the method for optimizing a lightning warning model using a deep learning algorithm of the present invention, the method further includes the following experimental evaluation criteria;

[0030] The experiment selected mean absolute percentage error, mean square error, test time and determination coefficient R 2 As a model measurement indicator; MAPE can measure the accuracy of the prediction model, RMSE can reflect the size of the average prediction error, R 2 It can characterize the fitting degree of the prediction model; in the classification prediction of lightning risk level, the smaller the values ​​of MAPE and RMSE, the better the R 2 The closer the value is to 1, the better the model fit is.

[0031]

[0032] The lightning warning model optimization of the present invention has the following advantages: (1) High accuracy: By analyzing and modeling comprehensive multi-source data (lightning location, radar, atmospheric electric field, terrain, meteorology, etc.), it is possible to capture information related to the occurrence and development of lightning more comprehensively, thereby improving the accuracy of the warning. Deep learning models, etc. can automatically learn complex features and patterns in the data, further improving the accuracy of the prediction. (2) Strong real-time performance: Data can be acquired and processed in real time, and warning information can be updated in a timely manner to meet the user's demand for real-time performance. This is of great significance for some scenarios that require high lightning warning time, such as outdoor activities, aerospace, power facilities, petrochemicals, etc. (3) Personalized customization: It can perform personalized risk classification and warning settings according to the characteristics of different target points and user needs. For example, for different industry users (such as electricity, communications, construction, etc.), targeted warning services can be provided to meet their specific safety needs. (4) Multi-dimensional evaluation: Not only the characteristics of lightning itself are considered, but also factors such as terrain and meteorology are combined to perform multi-dimensional risk assessment. In this way, the impact of lightning on the target point can be more comprehensively evaluated, and more accurate warning information can be provided. (5) Strong adaptability: It can adapt to the characteristics of lightning activities in different regions, seasons, and climate conditions, and has strong adaptability and generalization capabilities. By continuously updating and optimizing the model, its performance in various complex environments can be improved. (6) Early warning: It can issue an early warning within a certain period of time before the occurrence of lightning, providing users with sufficient time to take corresponding protective measures and reduce the losses caused by lightning disasters. (7) Comprehensive risk assessment: The lightning safety risks are graded so that users can more intuitively understand the degree of risk and make reasonable decisions and arrangements. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the TCN residual unit involved in this application;

[0034] Figure 2 It is a schematic diagram of the GRU unit structure of this application;

[0035] Figure 3 It is a schematic diagram of the MATT structure of this application;

[0036] Figure 4 This is a schematic diagram of risk classification preprocessing of lightning warning factors in this application;

[0037] Figure 5 It is the model calculation flow chart;

[0038] Figure 6 It is the structure diagram of the model of this application;

[0039] Figure 7 The 2024 thunderstorm weather process data set in Qingdao area is used to divide the model into training set and prediction set;

[0040] Figure 8-Figure 11 This is a schematic diagram comparing the prediction effects of the training sets and prediction sets of the four models. DETAILED DESCRIPTION

[0041] The following will be combined with the attached Figure 1-Figure 11 The present invention is described in detail, and the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The present invention provides a lightning warning model optimization method of a deep learning algorithm by improvement, and the specific optimization is as follows;

[0043] 1. Deep Learning Model

[0044] (1) TCN neural network layer: The TCN model is a simple and universal convolutional neural network architecture for solving time series problems. The TCN model consists of a group of residual units, each of which is a residual connection. [] The residual connection of a small neural network can speed up the feedback and convergence of the deep network and solve the "degeneration phenomenon" caused by the increase of network layers. Figure 1 As shown. The residual connection adds the input x of the residual unit to the output f(x), and the network is forced to be modeled as:

[0045] f(x)=h(x)-x

[0046] The residual unit contains two convolution units and nonlinear mapping. In the convolution unit, a one-dimensional dilated causal convolution is first performed. The sampling interval is adjusted by the dilation coefficient to achieve a larger receptive field (RF), that is, the area that the features on the convolution layer can see, so that the network can remember historical information long enough, and only convolve the input before time t to obtain the output at time t to ensure that future information will not be leaked; then the weights are normalized and the ReLU function is used as the activation function; finally, the Dropout operation is used to randomly discard neurons with a certain probability to prevent overfitting and accelerate the model training speed. Nonlinear mapping is to reduce the dimension of high-dimensional data when the input and output of the residual unit have different dimensions. The operation formula of one-dimensional dilated causal convolution is:

[0047]

[0048] Where: x is the input sequence; f is the filter; d is the dilation coefficient; k is the convolution kernel size; s-di ensures that convolution operations can only be performed on past inputs.

[0049] (2) GRU neural network layer: GRU is a simplified version of LSTM. It integrates the input gate and forget gate of LSTM into an update gate. It has no output gate, but adds a reset gate. The update gate controls the degree of retention of the state information at the previous moment. The larger the value of the update gate, the greater the influence of the state information at the previous moment on the current state. The reset gate controls the degree of integration of the current state with the previous information. The smaller the value, the more information is ignored. The GRU unit structure is as follows: Figure 2 shown.

[0050] The calculation formula of h(t) in the GRU unit is:

[0051]

[0052] Where: x t is the input at time t; h t is the output or state at time t; h t-1 is the state at time t-1; w is the weight; σ is the activation function sigmoid(·); tanh(·) is the activation function.

[0053] (3) MATT multi-head self-attention mechanism layer: Multi-head attention is an extended form of attention mechanism widely used in the Transformer model. It obtains the attention distribution of different subspaces of the input sequence by running multiple independent attention mechanisms in parallel, thereby more comprehensively capturing the potential multiple semantic associations in the sequence.

[0054] In multi-head attention, the input sequence first passes through three different linear transformation layers to obtain the query, key, and value respectively. Then, these transformed vectors are divided into several "heads", each with its own independent query, key, and value matrices. For each head, a scaled dot-product attention operation is performed, that is:

[0055]

[0056] Finally, the outputs of all heads are concatenated together and then fused through a linear layer to obtain the final attention output vector. In this way, multi-head attention can process the input sequence from different angles in parallel, improving the model's ability to understand and capture complex dependencies. In practice, multi-head attention can significantly improve the performance of the Transformer model in natural language processing and other sequence data processing tasks.

[0057] Lightning safety risk early warning methods include:

[0058] According to the distance between the lightning strike location and the target point, it is divided into five risk warning levels (i.e. 50km is type 1, 40km is type 2, 30km is type 3, 20km is type 4, and 10km is risk 5. The warning is lifted or there is no lightning safety risk, which is type 0), which corresponds to six warning levels.

[0059] This paper uses the historical data of thunderstorm weather in Qingdao in 2024 as a sample data set, and finally screens 893 data. The sample data set selects lightning location around the target point, radar extrapolation results, and atmospheric electric field instrument data. First, the three warning factors of lightning warning probability and critical threshold are preprocessed by risk classification. The method is as follows:

[0060] The TCN-GRU-MATT model structure includes:

[0061] (1) Input 1: Input forecast factors, which are lightning location, radar, and atmospheric electric field risk classification data every 6 minutes when there is a thunderstorm.

[0062] (2) Temporal convolutional network layer. Set up 2 layers of residual units. Each layer of residual units contains 2 convolution units and 1 nonlinear mapping. The convolution unit uses the ReLU function as the activation function to normalize the weight of the convolution kernel. The convolution kernel size is 2; and the Dropout coefficient is set to 0.4. The Dropout setting can randomly select some neurons to inactivate to prevent overfitting in training and speed up the convergence of the model; set the expansion coefficient to (1, 2, 4, 8, 16, 32); the filter is 128. The input and output of the residual unit have different dimensions and cannot be directly added. A 1×1 convolution layer is added to the residual mapping for dimensionality reduction.

[0063] (3) Input 2: Input classification labels, which are the six warning level results corresponding to every 6 minutes.

[0064] (4) Gated recurrent unit layer. Set up a 3-layer GRU unit. First, connect the output g of the TCN model with the output x2 of input 2. It is necessary to ensure that the dimension and number of samples of the data are the same, and obtain the data c.t .

[0065] c t =f(x (2，t) , g (t) )

[0066] Where: f represents the connection operation, which is implemented using the concatenate function of the keras model to connect the temporal convolutional network and input 2 as the input of the gated recurrent unit.

[0067] (5) Self-attention mechanism layer: The self-attention mechanism is used to weight the output of the GRU model and select the features that have the greatest impact on the prediction results.

[0068] (6) Integrate the weighted prediction results of all subsequences to obtain the final prediction result.

[0069] Experimental evaluation criteria:

[0070] The experiment selected mean absolute percentage error (MAPE), root mean square error (RMSE), test time and determination coefficient R 2 As a model measurement indicator, MAPE can measure the accuracy of the prediction model, RMSE can reflect the size of the average prediction error, and R 2 It can characterize the fitting degree of the prediction model. In the classification prediction of lightning risk level, the smaller the values ​​of MAPE and RMSE, the better the R 2 The closer the value is to 1, the better the model fit is.

[0071]

[0072] Calculation results

[0073] Experimental environment: The computer used in the calculations in this article is configured with Window 11 system, 32G running memory, and RTX4060Ti graphics card; the programming environment used is Matlab 2023a.

[0074] Comparison of model prediction results: In order to verify the performance of the model, the experiment compared the TCN-GRU-MATT model with the TCN and GRU models. The number of network layers of the TCN, GRU and TCN-GRU models was uniformly set to 3. The model was divided into training set and prediction set for the 2024 thunderstorm weather process data set in Qingdao. The results are as follows: Figure 7 shown.

[0075] The following table compares the prediction results of the four model training sets and prediction sets.

[0076]

[0077] The innovation of this application lies in:

[0078] First, the existing technology addresses the problem of (1) processing time series information: Lightning warning involves data that changes over time, such as lightning location, radar data, and atmospheric electric field. Traditional methods may find it difficult to fully utilize the long-term dependencies in these time series information.

[0079] However, this application uses the TCN neural network layer and its one-dimensional dilated causal convolution to achieve a larger receptive field, which can remember historical information long enough and only convolve past inputs to ensure that future information will not be leaked, thereby better processing time series data.

[0080] Second, the existing technology addresses the problem of (2) lightning warning feature extraction and heterogeneous data fusion: different warning factors (such as lightning location, radar data, and atmospheric electric fields) have their own characteristics. How to effectively extract and fuse these features is crucial to improving the accuracy of warnings.

[0081] This application uses the GRU neural network layer to simplify the LSTM structure and process sequence data more efficiently; it introduces the MATT multi-head self-attention mechanism layer to capture the attention distribution of the input sequence from different subspaces and obtain potential semantic associations more comprehensively, so as to better extract and fuse the characteristics of multi-source warning factors.

[0082] Third, the existing technology addresses the problem of (3) overfitting of deep learning models: when training complex deep learning models, overfitting is prone to occur, resulting in a decrease in the generalization ability of the model in practical applications.

[0083] The present application adopts the Dropout operation in the convolution unit of the TCN neural network layer to randomly discard neurons, thereby preventing overfitting and speeding up the model training.

[0084] Fourth, the existing technology addresses the problem that (4) the classification of warning levels affects the accuracy of warnings: the existing classification of lightning warning levels may not be detailed or accurate enough, which affects the actual effect of the warning.

[0085] This application accurately divides the distance between the lightning strike location and the target point into five risk warning levels (50km for type 1, 40km for type 2, 30km for type 3, 20km for type 4, and 10km for type 5), and combines multi-source data such as lightning location, radar data, and atmospheric electric field to perform risk grading preprocessing to improve the accuracy of warning level division.

[0086] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightning warning model optimization method based on a deep learning algorithm, characterized in that: The details are as follows: According to the distance between the lightning strike location and the target point, it is divided into five risk warning levels (i.e. 50km is type 1, 40km is type 2, 30km is type 3, 20km is type 4, and 10km is risk 5; Warning cancellation or no lightning safety risk is type 0), which corresponds to six warning levels; The lightning location around the target point, radar extrapolation results, and atmospheric electric field instrument data were selected from the sample data set; first, the lightning warning probability and critical threshold of the three warning factors were pre-processed for risk classification; Construct the TCN-GRU-MATT model structure, including: (1) Input 1: Input forecast factors, which are lightning location, radar, and atmospheric electric field risk classification data every 6 minutes when there is a thunderstorm weather process; (2) Temporal convolutional network layer: set up 2 layers of residual units, each layer of residual units contains 2 convolution units and 1 nonlinear mapping; the convolution unit uses the ReLU function as the activation function to normalize the weight of the convolution kernel; the convolution kernel size is 2; and the Dropout coefficient is set to 0.

4. The Dropout setting can randomly select some neurons to inactivate to prevent overfitting in training and speed up the convergence of the model; set the expansion coefficient to (1, 2, 4, 8, 16, 32); the filter is 128; the input and output of the residual unit have different dimensions and cannot be directly added. A 1×1 convolution layer is added to the residual mapping for dimensionality reduction; (3) Input 2: Input classification labels, which are the six warning level results corresponding to every 6 minutes; (4) Gated recurrent unit layer: Set up a 3-layer GRU unit. First, connect the output g of the TCN model with the output x2 of input 2. It is necessary to ensure that the dimension and number of samples of the data are the same, and obtain the data c. t ; ct=f(x(2,t),g(t)) Where: f represents the connection operation, which is implemented using the concatenate function of the keras model to connect the temporal convolutional network and input 2 as the input of the gated recurrent unit; (5) Self-attention mechanism layer: Use the self-attention mechanism to weight the output of the GRU model and select the features that have the greatest impact on the prediction results; (6) Integrate the weighted prediction results of all subsequences to obtain the final prediction result.

2. The method for optimizing a lightning warning model using a deep learning algorithm according to claim 1, characterized in that: The TCN neural network layer in step (4) is a simple and universal convolutional neural network architecture for solving time series problems. The TCN model consists of a group of residual units, each of which is a small neural network with residual connections. The residual connection can accelerate the feedback and convergence of the deep network and solve the "degeneration phenomenon" caused by the increase of network layers. The residual connection is to add the input x of the residual unit to the output f(x), and the network is forced to be modeled as follows: f(x)=h(x)-x The residual unit contains two convolution units and nonlinear mapping. First, a one-dimensional dilated causal convolution is performed in the convolution unit. The sampling interval is adjusted by the dilation coefficient to achieve a larger receptive field (RF), that is, the area that the features on the convolution layer can see, so that the network can remember long enough historical information and only convolve the input before time t to obtain the output at time t to ensure that future information will not be leaked. Then the weights are normalized and the ReLU function is used as the activation function. Finally, the Dropout operation is used to randomly discard neurons with a certain probability to prevent overfitting and accelerate the model training speed. Nonlinear mapping is to reduce the dimension of high-dimensional data when the input and output of the residual unit have different dimensions. The operation formula of the one-dimensional dilated causal convolution is: Where: x is the input sequence; f is the filter; d is the dilation coefficient; k is the convolution kernel size; s-di ensures that convolution operations can only be performed on past inputs.

3. The method for optimizing a lightning warning model using a deep learning algorithm according to claim 1, characterized in that: In step (4), GRU is a simplified version of LSTM, which integrates the input gate and forget gate of LSTM into an update gate. There is no output gate, but a reset gate is added. The update gate controls the degree of retention of the state information at the previous moment. The larger the value of the update gate, the greater the influence of the state information at the previous moment on the current state. The reset gate controls the degree of integration of the current state with the previous information. The smaller the value, the more information is ignored. The calculation formula of h(t) in the GRU unit is: Where: x t is the input at time t; h t is the output or state at time t; h t-1 is the state at time t-1; w is the weight; σ is the activation function sigmoid(·); tanh(·) is the activation function.

4. The method for optimizing a lightning warning model using a deep learning algorithm according to claim 1, characterized in that: It also includes the following experimental evaluation criteria; The experiment selected mean absolute percentage error, mean square error, test time and determination coefficient R 2 As a model measurement indicator; MAPE can measure the accuracy of the prediction model, RMSE can reflect the size of the average prediction error, R 2 It can characterize the fitting degree of the prediction model; in the lightning risk classification prediction, the smaller the values ​​of MAPE and RMSE, the better the R 2 The closer the value is to 1, the better the model fit is.

Images