A method for identifying and predicting severe convection based on deep learning

Abstract

The present invention relates to the field of meteorological technology, and discloses a method for identifying and forecasting severe convection based on deep learning. The method first constructs a severe convection identification model and a forecasting model, which are used for feature classification and prediction of weather occurrence time, location and intensity, respectively. The two models are trained by a deep learning algorithm, and their residuals are used to construct a residual fitting model to correct the forecast results. Combined with real-time and historical meteorological data, the method can generate accurate severe convective weather identification and forecasting, effectively make up for the limitations of traditional methods, improve forecast accuracy and timeliness, and is of great value to disaster prevention and mitigation and ensuring stable social and economic development.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorology, and in particular to a method for identifying and predicting severe convection based on deep learning. Background Art

[0002] Severe convective weather, such as thunderstorms, hail, tornadoes, short-term heavy rainfall, etc., is a very destructive natural phenomenon in the meteorological field, posing a serious threat to people's lives and property safety, and also has a significant impact on agricultural production, transportation, urban planning and other fields. Therefore, accurate and timely identification and forecasting of severe convective weather is of great significance for disaster prevention and mitigation and ensuring the stable development of social economy.

[0003] At present, the identification and forecasting of severe convective weather mainly rely on meteorological observation data, weather map analysis, numerical weather forecast models and other means. However, these methods have many limitations. Meteorological observation data are limited by the distribution density of observation sites and the accuracy of observation instruments, and it is often difficult to fully and accurately capture the subtle characteristics of severe convective weather. Weather map analysis relies on the experience and subjective judgment of meteorologists, and the forecast results are greatly affected by human factors. Although numerical weather forecast models can simulate atmospheric movement processes, their forecast accuracy is restricted by many factors such as model resolution, initialization conditions, and physical process parameterization schemes. For severe convective weather with strong suddenness and locality, the forecast effect is often unsatisfactory. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying and predicting severe convection based on deep learning to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for identifying and predicting severe convection based on deep learning, the method comprising:

[0006] Construct a severe convection identification model and a severe convection forecast model. The severe convection identification model is used to identify the characteristics of severe convective weather and classify it, and the severe convection forecast model is used to predict the time, location and intensity of severe convective weather based on current and historical meteorological data.

[0007] A deep learning algorithm is used to train a severe convection identification model and a severe convection prediction model respectively, and the residuals of the two models are used to construct a residual fitting model, which is used to correct and refine the results of severe convection identification and prediction;

[0008] The trained model is used in combination with real-time and historical meteorological data to generate identification and forecast results of severe convective weather.

[0009] Preferably, a convolutional neural network (CNN) algorithm is used to construct a strong convection recognition model, and spatial features in meteorological data are extracted through multi-layer convolution and pooling operations, and finally classified through a fully connected layer.

[0010] Preferably, the training steps of the strong convection identification model include:

[0011] Data preprocessing: collect and clean meteorological data, including wind speed, wind direction, air pressure, temperature, humidity, precipitation rate and its rate of change, and standardize the collected data;

[0012] Constructing a deep neural network (CNN) structure: Designing a CNN model, including an input layer, multiple convolutional layers, a pooling layer, a fully connected layer, and an output layer. The convolutional layer is used to extract local spatial features of meteorological data, the pooling layer is used to reduce feature dimensions and retain important information, and the fully connected layer is used to map features to classification results;

[0013] Model training: The preprocessed meteorological data is input into the CNN model, the output is calculated through forward propagation, the cross entropy loss function is used to measure the difference between the model prediction and the true label, and the model parameters are updated through the back propagation algorithm. The training is iterated until the loss function converges or the preset training rounds are reached;

[0014] Model evaluation and optimization: Evaluate the performance of the model on the validation set and adjust the model structure or hyperparameters based on the evaluation results, including the convolution kernel size, number of convolution layers, and learning rate.

[0015] Preferably, a long short-term memory network (LSTM) algorithm is used to construct a severe convection forecast model, through which the time dependency of meteorological data is learned and the forecast results of severe convective weather in the future period are output.

[0016] Preferably, the training steps of the severe convection forecast model include:

[0017] Time series data preparation: Organize historical meteorological data to form a time series data set, including meteorological observations at consecutive time points and corresponding severe convective weather labels;

[0018] Construct the LSTM structure of the long short-term memory network: Design the LSTM model, including the input layer, LSTM layer, fully connected layer and output layer. The LSTM layer is used to capture the long-term temporal dependencies in the meteorological data, and the fully connected layer is used to map the output of the LSTM layer to the future severe convective weather forecast value;

[0019] Sequence prediction training: Input the time series data set into the LSTM model, calculate the output sequence through iterative calculation of time steps, use the mean square error MSE as the loss function, and optimize the model parameters through the gradient descent algorithm until the loss function reaches the minimum value or the training process is stable;

[0020] Model validation and tuning: Evaluate the prediction performance of the model on the validation set and adjust the model structure or training strategy based on the evaluation results, including increasing the number of LSTM layers, adjusting the number of hidden units, and changing the learning rate.

[0021] Preferably, the step of constructing the residual fitting model includes:

[0022] The residuals between the actual output and the expected output of the severe convection identification model and the severe convection prediction model are calculated, which are recorded as identification residuals and prediction residuals respectively;

[0023] The identification residuals and forecast residuals are used as new features and combined with the original meteorological data to construct a residual feature set;

[0024] Select the deep neural network DNN algorithm, train the residual feature set, and obtain the residual fitting model;

[0025] The structure of the residual fitting model includes an input layer, multiple hidden layers and an output layer. The input layer receives the residual feature set, and the output layer outputs the corrected severe convection identification and forecast results.

[0026] Preferably, when the deep neural network DNN algorithm trains the residual fitting model, the loss function used is the mean square error MSE, and its calculation formula is:

[0027]

[0028] Among them, y i is the actual residual, is the residual predicted by the DNN model, and N is the number of samples; the parameters of the DNN model are optimized by minimizing MSE.

[0029] Preferably, during the training process of the residual fitting model, a back propagation algorithm is used to update parameters, and the specific steps include:

[0030] Forward propagation: Input the residual feature set into the DNN model and obtain the predicted value of the output layer by layer calculation;

[0031] Calculate the error: Use the MSE loss function to calculate the error between the predicted value and the actual residual;

[0032] Back propagation: Calculate the gradient of each layer's parameters based on the error, update the parameters using the gradient descent algorithm, and iterate until the loss function converges or the preset number of training rounds is reached.

[0033] Preferably, during the training of the residual fitting model, an early stopping method is used to prevent the model from overfitting, and the performance of the model is monitored on the validation set. When the model performance does not improve after multiple consecutive training rounds, the training is stopped and the optimal model parameters are saved.

[0034] Preferably, when constructing the residual feature set, it also includes performing feature engineering processing on the identification residuals and the prediction residuals, including calculating the first-order difference, second-order difference and moving average of the residuals.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] By building a severe convection identification model and a severe convection forecast model, the present invention can fully utilize the powerful feature extraction capability of deep learning to automatically learn and identify the complex features of severe convective weather, such as radar echo structure, wind direction and speed changes, temperature and humidity distribution, etc., so as to achieve more accurate classification and forecasting. Compared with traditional methods, deep learning models can capture more subtle weather changes and improve the accuracy of forecasts.

[0037] The deep learning model can quickly process a large amount of real-time and historical meteorological data, and combined with an efficient training algorithm, the forecast results can be quickly generated and updated. This is crucial for natural phenomena such as severe convective weather, which are highly sudden and change rapidly. The present invention can shorten the forecast period and provide more timely information support for disaster prevention and mitigation. By constructing a residual fitting model, the present invention can use the residual information of the severe convection identification model and the forecast model to correct and refine the preliminary forecast results. This multi-level, multi-model fusion strategy can further improve the refinement of the forecast and provide more accurate data support for refined disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A step diagram of a method for identifying and predicting severe convection based on deep learning according to the present invention;

[0039] Figure 2 This is the training flow chart of the severe convection identification model. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0041] See also Figure 1-2 The present invention provides a technical solution: a method for identifying and predicting severe convection based on deep learning, the method comprising:

[0042] Construct severe convection identification model and severe convection forecast model:

[0043] According to the characteristics of severe convective weather, a severe convection identification model is constructed. The model uses deep learning algorithms, such as convolutional neural networks (CNN) or recurrent neural networks (RNN), to identify the characteristics of severe convective weather and classify it. The model input includes meteorological observation data, such as radar echo maps, satellite cloud maps, ground observation station data, etc., and the output is the type of severe convective weather, such as thunderstorms, hail, tornadoes, or short-term heavy precipitation.

[0044] At the same time, a severe convection forecast model is constructed. This model also uses deep learning algorithms, but focuses more on the prediction ability of time series, such as long short-term memory networks (LSTM) or gated recurrent units (GRU). The model input includes current and historical meteorological data, such as temperature, humidity, wind speed, wind direction, air pressure, etc., as well as weather type information output by the severe convection identification model. The model output is the predicted time, location and intensity of severe convective weather in the future.

[0045] Train the model and build a residual fit model:

[0046] A large amount of historical meteorological data and severe convective weather events are used as training samples to train the severe convective identification model and severe convective forecast model respectively. During the training process, appropriate loss functions and optimization algorithms, such as the cross entropy loss function and the Adam optimization algorithm, are used to minimize the prediction error of the model.

[0047] In order to improve the accuracy and refinement of the forecast, the residuals of the two models are used to construct a residual fitting model. The residual refers to the difference between the model prediction value and the actual observation value. By analyzing the characteristics and rules of the residuals, another deep learning model is constructed to fit these residuals, thereby correcting and refining the results of severe convection identification and forecasting.

[0048] Generate identification and forecast results of severe convective weather:

[0049] In practical applications, real-time meteorological data is first obtained and input into the trained severe convective recognition model to obtain the type information of severe convective weather. Then, the real-time meteorological data and the recognition results are input into the severe convective forecast model to obtain the forecast results of severe convective weather in the future. Finally, the residual fitting model is used to correct and refine the forecast results to obtain the final severe convective weather recognition and forecast results.

[0050] The present invention will be further described below in conjunction with Examples 1 to 3:

[0051] Embodiment 1:

[0052] This embodiment is used to describe a method for constructing a strong convection identification model. The specific implementation steps are as follows:

[0053] Collect a wide range of meteorological data, including but not limited to wind speed, wind direction, air pressure, temperature, humidity, precipitation rate and its rate of change. The data comes from multiple channels such as ground observation stations, radar echoes, satellite remote sensing, etc. to ensure the comprehensiveness and representativeness of the data.

[0054] Clean the data, remove outliers, missing values, etc., to ensure data quality. After cleaning, standardize the data to unify the data of different dimensions to the same scale for the convenience of subsequent model processing. Standardization can use the zero mean unit variance standardization method, that is, subtract the mean of each feature and divide it by its standard deviation.

[0055] Construct a convolutional neural network CNN structure:

[0056] Design a CNN model whose structure includes input layer, multiple convolutional layers, pooling layer, fully connected layer and output layer.

[0057] Input layer: receives preprocessed meteorological data in the form of multidimensional tensors, where the dimensions include time, space (such as longitude and latitude) and meteorological characteristics.

[0058] Convolution layer: Multiple layers of convolution operations are used, and each layer uses convolution kernels of different sizes and numbers to extract local spatial features in meteorological data. The convolution kernel slides on the data and generates feature maps through dot product operations. The activation function uses ReLU (Rectified Linear Unit) to increase the nonlinear expression ability of the model.

[0059] Pooling layer: Following each convolutional layer, a maximum pooling or average pooling operation is used to reduce the dimension of the feature map and the amount of computation while retaining important information.

[0060] Fully connected layer: Flatten the feature map output by the pooling layer into a one-dimensional vector and map it to the classification result through a multi-layer fully connected network.

[0061] Output layer: The Softmax function is used to convert the output of the fully connected layer into a probability distribution, which represents the predicted probability of each type of severe convective weather.

[0062] The preprocessed meteorological data is divided into a training set and a validation set. The training set is used for model training, and the validation set is used for model evaluation.

[0063] The training set data is input into the CNN model and the output is calculated through forward propagation. The cross entropy loss function is used to measure the difference between the model prediction and the true label. The cross entropy loss function is defined as the negative log-likelihood between the predicted probability distribution and the true label distribution.

[0064] The gradient of the loss function to the model parameters is calculated through the back-propagation algorithm, and the model parameters are updated using an optimization algorithm (such as Adam or SGD). The training process is iterated until the loss function converges or the preset training rounds are reached.

[0065] The performance of the model was evaluated on the validation set, and indicators such as accuracy, precision, recall, and F1 score were used to measure the model's ability to identify severe convective weather.

[0066] According to the evaluation results, adjust the model structure or hyperparameters to optimize the model performance. The adjusted parameters include but are not limited to the size and number of convolution kernels, the number of convolution layers, the type and size of pooling layers, the number of layers and neurons in the fully connected layer, the learning rate, the dropout ratio, etc.

[0067] Through repeated iterative training, evaluation and optimization, we finally obtained a strong convection recognition model with excellent performance. This model can accurately identify strong convective weather types such as thunderstorms, hail, tornadoes, short-term heavy rainfall, etc.

[0068] Embodiment 2:

[0069] This embodiment is used to describe the construction method of the severe convective forecast model, which aims to output accurate forecast results of severe convective weather in the future by learning the time dependence of meteorological data. The specific implementation steps of the method are as follows:

[0070] Historical meteorological data are obtained from data sources such as meteorological observation stations or meteorological satellites, including but not limited to key meteorological elements such as temperature, humidity, wind speed, wind direction, air pressure, and precipitation. These data need to be organized in chronological order to form a time series data set. The data at each time point contains a complete set of meteorological observations and corresponds to a severe convective weather label, which is annotated according to the severe convective events in the historical records, such as "yes" or "no" to indicate whether severe convective weather occurred at that time point. In order to enhance the generalization ability of the model, the data set needs to be preprocessed, including missing value filling, outlier detection and processing, data standardization or normalization, etc., to ensure the quality and consistency of the data.

[0071] When designing an LSTM model, you need to consider the model's input layer, LSTM layer, fully connected layer, and output layer. The input layer is responsible for receiving preprocessed time series data, and the LSTM layer is used to capture long-term temporal dependencies in meteorological data. The number of hidden units in the LSTM layer can be adjusted according to the complexity of the dataset and the requirements of the model, and the optimal value is usually determined through experiments.

[0072] The fully connected layer is located after the LSTM layer and is used to map the output of the LSTM layer to the future severe convective weather forecast value. The output layer is designed as a classification output (such as a binary classification problem) or a regression output (such as predicting the specific intensity of severe convective weather) according to the task requirements.

[0073] In addition, in order to improve the performance of the model, you can also consider increasing the number of LSTM layers to form a deep LSTM network, or introduce other auxiliary structures, such as attention mechanisms, gating mechanisms, etc.

[0074] The preprocessed time series dataset is divided into a training set and a validation set. During the training process, the training set is input into the LSTM model, and the output sequence is calculated by iterative time steps. In order to evaluate the prediction performance of the model, the mean square error (MSE) is used as the loss function, which measures the difference between the model prediction value and the true value.

[0075] The gradient descent algorithm (such as the Adam optimizer) is used to optimize the model parameters. By continuously iteratively updating the parameters, the loss function is gradually reduced until it reaches the minimum value or the training process becomes stable. During the training process, attention should also be paid to the convergence speed and stability of the model to avoid overfitting or underfitting.

[0076] Evaluate the prediction performance of the model on the validation set, and use evaluation indicators such as accuracy, recall, and F1 score to quantify the prediction effect of the model. Adjust and optimize the model structure or training strategy based on the evaluation results.

[0077] Specific tuning strategies include: increasing the number of LSTM layers to improve the expressiveness of the model; adjusting the number of hidden units to balance the complexity and generalization ability of the model; changing the learning rate to control the convergence speed and stability of the model; introducing regularization techniques (such as L1 or L2 regularization) to reduce overfitting; and adopting strategies such as early stopping and learning rate decay to further improve the performance of the model.

[0078] The above steps are iterated repeatedly until a severe convection forecast model that meets the forecast performance requirements is obtained.

[0079] Embodiment 3:

[0080] In order to more effectively improve the recognition and forecast accuracy of severe convective weather, this embodiment provides an improved method based on the residual fitting model. The following are the detailed implementation steps of the method:

[0081] The existing severe convection identification model and severe convection forecast model are used to predict historical meteorological data, and the difference between the actual output and the expected output of these models is calculated, which are defined as identification residual and forecast residual, respectively. Assume that the original meteorological data set is X, the actual output of the severe convection identification model is Y1_actual, and the expected output is Y1_expected, then the identification residual R1 is calculated as R1 = Y1_actual-Y1_expected; similarly, for the severe convection forecast model, its actual output is Y2_actual, and the expected output is Y2_expected, then the forecast residual R2 is calculated as R2 = Y2_actual-Y2_expected.

[0082] The calculated identification residual R1 and forecast residual R2 are used as new features and combined with the original meteorological data X to form a residual feature set. In order to further enrich the feature set, feature engineering processing is performed on the identification residual and forecast residual, including calculating the first-order difference of the residual (i.e., the difference between two adjacent residual values), the second-order difference (i.e., the difference between two adjacent first-order differences), and the moving average (using a fixed window size to average the residual sequence). These processing helps to capture the dynamic changes and smooth trends in the residual sequence and improve the model's ability to recognize residual patterns.

[0083] Deep neural network (DNN) is used as the core algorithm of the residual fitting model. The structure of the DNN model is designed to include an input layer, multiple hidden layers and an output layer. The input layer receives the residual feature set processed by feature engineering; the hidden layer extracts high-level features through nonlinear transformation; and the output layer outputs the corrected severe convection identification and forecast results. During the training process, the mean square error (MSE) is selected as the loss function, and its calculation formula is: MSE = (1 / N) * Σ ((R_actual-R_predicted)^2), where R_actual is the actual residual, R_predicted is the residual predicted by the DNN model, and N is the number of samples. The weights and bias parameters of the DNN model are optimized by minimizing the MSE.

[0084] The back propagation algorithm is used to update the parameters of the DNN model. The specific steps include:

[0085] Forward propagation: Input the residual feature set into the DNN model and calculate layer by layer until the predicted value of the output layer is obtained.

[0086] Calculate the error: Use the MSE loss function to calculate the error between the predicted value and the actual residual.

[0087] Back propagation: Calculate the gradient of each layer's parameters based on the error, use the gradient descent algorithm (such as the Adam optimizer) to update the parameters, and iterate this process until the loss function converges or reaches the preset training rounds.

[0088] In the training process of the residual fitting model, the early stopping method is used to prevent the model from overfitting. Specifically, while training the model on the training set, the performance of the model (such as MSE) is monitored on the validation set. When the model performance on the validation set does not improve after multiple training rounds (such as 10 rounds), the model is considered to have begun to overfit. At this time, the training is stopped and the model parameters with the best performance are saved.

[0089] After the training is completed, the residual fitting model is applied to the new meteorological data. First, the severe convection identification model and forecast model are used to make preliminary predictions on the new data, then the identification residuals and forecast residuals are calculated, and these residuals and the original meteorological data are input into the residual fitting model to obtain the revised severe convection identification and forecast results.

[0090] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0091] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and predicting severe convection based on deep learning, characterized in that: The method comprises: Construct a severe convection identification model and a severe convection forecast model. The severe convection identification model is used to identify the characteristics of severe convective weather and classify it, and the severe convection forecast model is used to predict the time, location and intensity of severe convective weather based on current and historical meteorological data. A deep learning algorithm is used to train a severe convection identification model and a severe convection prediction model respectively, and the residuals of the two models are used to construct a residual fitting model, which is used to correct and refine the results of severe convection identification and prediction; The trained model is used in combination with real-time and historical meteorological data to generate identification and forecast results of severe convective weather.

2. The method for identifying and predicting severe convection based on deep learning according to claim 1, characterized in that: The convolutional neural network (CNN) algorithm is used to build a severe convection recognition model. The spatial features in the meteorological data are extracted through multi-layer convolution and pooling operations, and finally classified through a fully connected layer.

3. The method for identifying and predicting severe convection based on deep learning according to claim 2, characterized in that: The training steps of the severe convection identification model include: Data preprocessing: collect and clean meteorological data, including wind speed, wind direction, air pressure, temperature, humidity, precipitation rate and its rate of change, and standardize the collected data; Construct the convolutional neural network (CNN) structure: Design a CNN model, including an input layer, multiple convolutional layers, a pooling layer, a fully connected layer, and an output layer. The convolutional layer is used to extract the local spatial features of the meteorological data, the pooling layer is used to reduce the feature dimension and retain important information, and the fully connected layer is used to map the features to the classification results; Model training: The preprocessed meteorological data is input into the CNN model, the output is calculated through forward propagation, the cross entropy loss function is used to measure the difference between the model prediction and the true label, and the model parameters are updated through the back propagation algorithm. The training is iterated until the loss function converges or the preset training rounds are reached; Model evaluation and optimization: Evaluate the performance of the model on the validation set and adjust the model structure or hyperparameters based on the evaluation results, including the convolution kernel size, number of convolution layers, and learning rate.

4. The method for identifying and predicting severe convection based on deep learning according to claim 1, characterized in that: The long short-term memory network (LSTM) algorithm is used to construct a severe convection forecast model. The algorithm learns the time dependence of meteorological data and outputs the forecast results of severe convective weather in the future.

5. The method for identifying and predicting severe convection based on deep learning according to claim 4, characterized in that: The training steps of the severe convection forecast model include: Time series data preparation: Organize historical meteorological data to form a time series data set, including meteorological observations at consecutive time points and corresponding severe convective weather labels; Construct the LSTM structure of the long short-term memory network: Design the LSTM model, including the input layer, LSTM layer, fully connected layer and output layer. The LSTM layer is used to capture the long-term temporal dependencies in the meteorological data, and the fully connected layer is used to map the output of the LSTM layer to the future severe convective weather forecast value; Sequence prediction training: Input the time series data set into the LSTM model, calculate the output sequence through iterative calculation of time steps, use the mean square error MSE as the loss function, and optimize the model parameters through the gradient descent algorithm until the loss function reaches the minimum value or the training process is stable; Model validation and tuning: Evaluate the prediction performance of the model on the validation set and adjust the model structure or training strategy based on the evaluation results, including increasing the number of LSTM layers, adjusting the number of hidden units, and changing the learning rate.

6. The method for identifying and predicting severe convection based on deep learning according to claim 1, characterized in that: The steps to construct a residual fitting model include: The residuals between the actual output and the expected output of the severe convection identification model and the severe convection prediction model are calculated, which are recorded as identification residuals and prediction residuals respectively; The identification residuals and forecast residuals are used as new features and combined with the original meteorological data to construct a residual feature set; Select the deep neural network DNN algorithm, train the residual feature set, and obtain the residual fitting model; The structure of the residual fitting model includes an input layer, multiple hidden layers and an output layer. The input layer receives the residual feature set, and the output layer outputs the corrected severe convection identification and forecast results.

7. The method for identifying and predicting severe convection based on deep learning according to claim 6, characterized in that: When the deep neural network DNN algorithm trains the residual fitting model, the loss function used is the mean square error MSE, and its calculation formula is: Among them, y i is the actual residual, is the residual predicted by the DNN model, and N is the number of samples; the parameters of the DNN model are optimized by minimizing MSE.

8. The method for identifying and predicting severe convection based on deep learning according to claim 7, characterized in that: During the training process of the residual fitting model, the back propagation algorithm is used to update the parameters. The specific steps include: Forward propagation: Input the residual feature set into the DNN model and obtain the predicted value of the output layer by layer calculation; Calculate the error: Use the MSE loss function to calculate the error between the predicted value and the actual residual; Back propagation: Calculate the gradient of each layer's parameters based on the error, update the parameters using the gradient descent algorithm, and iterate until the loss function converges or the preset number of training rounds is reached.

9. The method for identifying and predicting severe convection based on deep learning according to claim 7, characterized in that: During the training process of the residual fitting model, the early stopping method is used to prevent the model from overfitting. The performance of the model is monitored on the validation set. When the model performance does not improve after multiple consecutive training rounds, the training is stopped and the optimal model parameters are saved.

10. The method for identifying and predicting severe convection based on deep learning according to claim 6, characterized in that: When constructing the residual feature set, feature engineering processing of the identification residuals and prediction residuals is also included, including calculating the first-order difference, second-order difference and moving average of the residuals.

Images