Optimization method for simulating the frequency of compound extreme wind and rain events in climate models based on deep learning

Abstract

The present invention relates to the technical field of climate event simulation, and in particular to an optimization method for simulating the frequency of compound extreme wind and rain events using a climate model based on deep learning, comprising the following steps: obtaining observation data of precipitation and wind speed in a target area and simulation data of multiple climate models; using a percentile threshold method to calculate precipitation and wind speed thresholds, and determining the frequency of occurrence of compound extreme wind and rain events; using observation data to evaluate climate model simulation results, selecting their frequencies as feature values ​​of a deep learning model, and using the observation frequencies as true values; constructing a multi-layer perceptron neural network, and optimizing model training by integrating a weighted loss function that combines mean square error, standard deviation ratio, and minimum value constraints; inputting the simulation results of the climate model into a trained model, and generating a corrected result of the frequency of compound extreme wind and rain events, thereby improving the simulation accuracy of the frequency of compound extreme wind and rain events, being able to effectively reduce deviations, and providing support for disaster prediction and decision-making under the background of climate change.

Description

Technical Field

[0001] The present invention relates to the technical field of climate event simulation, and in particular to an optimization method for simulating the frequency of compound extreme wind and rain events using a climate model based on deep learning. Background Art

[0002] Against the backdrop of intensified global climate change, extreme weather and climate events such as extreme high temperatures, heat waves and heavy rainfall occur frequently. Extreme events of a single meteorological variable have a profound impact on social economy and people's lives. Compound extreme events caused by the co-occurrence of extreme events of multiple meteorological variables cause far more damage than single extreme events. Coastal areas are often affected by tropical cyclones, and the frequency of compound extreme wind and rain events caused by the co-occurrence of disastrous weather such as floods and storm surges is higher than that inland. Therefore, it is particularly important to use climate models to accurately estimate the frequency of compound extreme wind and rain events in coastal areas.

[0003] The global climate model is a numerical model that simulates the Earth's climate system through a series of mathematical and physical formulas. These models play an important role in gaining a deeper understanding of the changes in extreme weather and climate events. However, due to the relatively low resolution of global climate model simulations, there are large deviations in the simulation of regional climate characteristics, resulting in a certain degree of uncertainty in the climate model's prediction of extreme weather and climate events. Therefore, it is of great significance to reduce the deviation of extreme weather and climate events simulated by global climate models at a regional scale.

[0004] In order to correct the bias of climate model simulation, dynamical downscaling and statistical downscaling methods are often used in the past. The dynamical downscaling method relies on the simulation of physical processes to provide more comprehensive climate information. However, it has a relatively large demand for computing resources and is therefore difficult to implement. The statistical downscaling method achieves bias correction by establishing a statistical relationship between climate models and observational data, and deep learning technology is one of them. In recent years, deep learning technology has been widely used in the meteorological field because it can better handle nonlinear relationships between data. Certain results have also been achieved in using deep learning technology to handle the bias correction problem of climate models.

[0005] However, since the existing deep learning models are affected by their own structures, the models trained by different combinations of neural networks and loss functions have limitations and output averaging problems. That is, the model output results tend to be average values ​​and cannot accurately reflect the actual distribution characteristics of compound extreme events. At the same time, it is difficult to fully utilize the information output by multiple climate models, resulting in insufficient extraction of data features from different models, affecting the accuracy and reliability of the simulation. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides an optimization method for simulating the frequency of compound extreme wind and rain events based on deep learning climate models, which is characterized by comprising the following steps:

[0007] S1. Obtain observation data and multiple climate model output data in the target area, and construct a grid network in the target area;

[0008] S2. Based on the percentile threshold method, the observed rainfall threshold and the observed wind speed threshold are calculated according to the observed data, the climate model precipitation threshold and the climate model wind speed threshold are calculated according to the output data of multiple climate models, and the frequency of occurrence of the first composite extreme wind and rain event is calculated according to the observed rainfall threshold and the observed wind speed threshold, and the frequency of occurrence of the second composite extreme wind and rain event is calculated according to the climate model precipitation threshold and the climate model wind speed threshold;

[0009] S3. Based on the observed data, the frequency of the second composite extreme wind and rain event is evaluated, and the frequency of the second composite extreme wind and rain event is selected as the characteristic value of the deep learning model according to the evaluation result, and the frequency of the first composite extreme wind and rain event at the same time and grid point location is selected as the true value of the deep learning model;

[0010] S4. Based on a multi-layer perceptron neural network, a deep learning model is constructed to reduce the error value of the frequency of the second composite extreme wind and rain event, and the deep learning model is trained by a weighted loss function according to the characteristic value and the true value obtained in step S3;

[0011] S5. Input the frequency of occurrence of the second composite extreme wind and rain event into the deep learning model trained in step S4 for training, and generate the observation constrained optimization result of the frequency of composite extreme wind and rain events in the target area;

[0012] The observation data includes gridded data of precipitation and gridded data of maximum wind speed actually observed in the target area;

[0013] The plurality of climate model output data include gridded precipitation data and gridded maximum wind speed data obtained by simulating a plurality of climate models in the target area;

[0014] The step S2 comprises the following steps:

[0015] S201. Based on the percentile threshold method, the observed rainfall threshold and the observed wind speed threshold are calculated according to the observed data, and the climate model precipitation threshold and the climate model wind speed threshold are calculated according to the output data of multiple climate models. The percentile threshold method can be expressed as:

[0016]

[0017] Where k is the percentile, Nsorted (i, t) is the sequence of the i-th grid position at the t-th time and is sorted from small to large. is the length of the sequence, P(i,t) represents the N sorted Middle The threshold value of the position;

[0018] S202. According to the obtained observed rainfall threshold, observed wind speed threshold, climate model precipitation threshold and climate model wind speed threshold, it is determined whether an extreme wind and rain event occurs in each grid of the target area grid network, which can be expressed as:

[0019]

[0020] Among them, P (i,t) and W (i,t) Respectively represent the precipitation and maximum wind speed at the tth time in the i-th grid, and Respectively represent the precipitation threshold and wind speed threshold at the tth time in the i-th grid; when the variable Z (i,t) When it is 1, it means that a composite extreme wind and rain event occurs at the corresponding time point, otherwise it does not occur;

[0021] S203. According to the judgment result, the occurrence frequency of the first composite extreme wind and rain event and the occurrence frequency of the second composite extreme wind and rain event are calculated respectively, and the calculation formula can be expressed as:

[0022]

[0023] Among them, F(i,year) is the cumulative frequency of compound extreme wind and rain events in one year, year is the year sequence, and year=[y1,y2,y3,…,y n ], i represents the grid position, d is the day sequence, m is the month sequence, D m is the day of the month, Z （i,d,m，year） is the composite extreme wind and rain event occurring in the year th, month m, and day d in the ith grid.

[0024] Furthermore, the deep learning model in step S4 includes:

[0025] The input layer is used to receive the occurrence of the second-frequency composite extreme wind and rain event, and randomly extract the frequency samples of the second-frequency composite extreme wind and rain event to establish a feature set, which can be expressed as:

[0026]

[0027] Where a represents the number of climate models, j represents the grid position of the randomly selected model simulation results, and y ’Indicates the year of the random selection model simulation results, n ’ Indicates the number of randomly selected samples;

[0028] The hidden layer includes a fully connected layer and a discard layer connected in sequence. There are multiple fully connected layers and discard layers. The fully connected layer is used to transmit feature information extracted from the feature set through an activation function. The activation function can be expressed as:

[0029]

[0030] Among them, W is the weight matrix, x is the feature set vector, b is the bias vector, and y is the LR is the feature information vector, is the activation function;

[0031] The discard layer is used to filter the feature information output by the fully connected layer, and randomly receive the feature information vector output results of some neurons in the fully connected layer;

[0032] The output layer is used to integrate the feature information extraction results and output the optimization results.

[0033] Furthermore, the weighted loss function in step S4 includes the mean square error, the ratio of standard deviation and the minimum value constraint function, and its calculation formula is:

[0034]

[0035] Among them, Loss represents the degree of deviation between the model prediction value and the true value, MSE is the mean square error, W1 is the weight parameter of MSE, RSD is the ratio of standard deviation, W2 is the weight parameter of RSD, Min_Loss is the minimum value constraint function, and W3 is the weight parameter of Min_Loss. The W1, W2, and W3 are used to adjust the influence of different constraints on the model. By setting multiple groups of parameter tuning experiments, the best weight combination is determined.

[0036] Furthermore, the MSE can be expressed as:

[0037]

[0038] Among them, y n The frequency of compound extreme wind and rain events simulated by the deep learning model, is the frequency of the first composite extreme wind and rain event, n is the nth randomly selected sample, and N represents the total number of randomly selected samples.

[0039] Further, the RSD can be expressed as:

[0040]

[0041] in, is the standard deviation of the frequency of composite extreme wind and rain events simulated by the deep learning model, is the standard deviation of the frequency of the first composite extreme wind and rain event.

[0042] Furthermore, the Min_Loss can be expressed as:

[0043]

[0044] Among them, y n The frequency of compound extreme wind and rain events simulated by the deep learning model, is the frequency of the first composite extreme wind and rain event, n is the nth randomly selected sample, and N represents the total number of randomly selected samples.

[0045] Furthermore, the deep learning model used in step 5 updates parameters through back propagation calculation, and continuously iterates and updates parameters until the value of the weighted loss function converges to a preset range, thereby obtaining the observation-constrained optimization result of the frequency of compound extreme wind and rain events in the target area. The back propagation calculation can be expressed as:

[0046]

[0047] Among them, w is the weight parameter, b is the bias parameter, θ is the parameter set of w and b, θold(w, b) is the parameter set before the model parameters are updated, θnew(w, b) is the parameter set after the model parameters are updated, η is the learning rate, and L is the loss function.

[0048] The beneficial effects of the present invention are:

[0049] (1) The neural network deep learning scheme based on the multi-layer perceptron of the present invention utilizes the characteristics of the fully connected layer to convert the frequency of compound extreme wind and rain events output by multiple climate models into a higher-dimensional perspective to process complex data feature information, thereby better capturing the nonlinear relationship characteristics between the output results of multiple climate models and the actual observation results. Compared with the traditional method of taking the ensemble average of the simulation results of multiple climate models with equal weights and the ensemble median statistical method, it not only considers the simulation information of multiple climate models, but also imposes observation constraints on the information of multiple climate simulations, so that the information extraction in the climate model is more sufficient, and the output result has lower deviation and is closer to the actual situation.

[0050] (2) The present invention proposes an improved weighted loss function that integrates the mean square error, the standard deviation ratio and the minimum value constraint function, while taking into account the difference between the output results and the actual results and the data distribution characteristics, effectively alleviating the tendency of deep learning model simulation results to tend to mean, making the model simulation results closer to the actual observation results, thereby effectively establishing the relationship between multiple climate model simulation results and the actual observation results.

[0051] (3) In the process of constructing a model for fusing the simulation results of multiple climate models, the present invention adds a dropout layer after each fully connected layer on the basis of the traditional multi-sensor neural network model, which helps to reduce overfitting, improve training efficiency and control the complexity of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 , a flow chart of the optimization method for simulating the frequency of complex extreme wind and rain events based on deep learning climate models of the present invention.

[0053] Figure 2 , schematic diagram of the weighted loss function optimization deep learning model in an embodiment of the present invention.

[0054] Figure 3 , a schematic diagram of a deep learning model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the content of the present invention and make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the embodiments and drawings. The illustrative embodiments of the present invention and their description are only used to explain the present invention and are not intended to further limit the present invention.

[0056] like Figure 1 As shown, an embodiment of the present invention provides an optimization method for simulating the frequency of compound extreme wind and rain events using a climate model based on deep learning, comprising the following steps:

[0057] Step 1: Obtain the daily cumulative precipitation grid data and daily 10-meter maximum wind speed grid data output by multiple climate models and the actual observed daily cumulative precipitation and 10-meter maximum wind speed grid data of the target area of ​​interest, and use the interpolation method to unify the model and observation data at the same resolution.

[0058] Step 2: Calculate the precipitation threshold and wind speed threshold corresponding to each grid point for the interpolated multiple climate models and actual observation data, and record the occurrence of composite extreme wind and rain events when the precipitation and the maximum wind speed at 10 meters on the same grid at the same time exceed their corresponding thresholds, and count the annual cumulative frequency of composite extreme wind and rain events. The specific steps are as follows:

[0059] Step 2.1: Use the percentile threshold method to determine the 95th percentile extreme precipitation threshold and extreme wind speed threshold. The percentile threshold P is calculated as:

[0060]

[0061] Where k is the percentile, Nsorted (i, t) is the sequence of the i-th grid position at the t-th time and is sorted from small to large. is the length of the sequence, P(i,t) represents the N sorted Middle The threshold value of the position;

[0062] Step 2.2: Calculate the frequency of composite extreme wind and rain events, using a binary variable formula to indicate whether an event occurs at each grid point. The variable Z (i,t) When it is 1, it means that a composite extreme wind and rain event occurs in the grid at the corresponding time, otherwise it does not occur, which can be expressed as:

[0063]

[0064] Among them, P (i,t) and W (i,t) Respectively represent the precipitation and maximum wind speed at the tth time in the i-th grid, and They represent the precipitation threshold (greater than 1 mm) and wind speed threshold (greater than 0.5 m / s) at the tth time in the i-th grid respectively;

[0065] Step 2.3, based on the judgment results, the occurrence frequency of the first composite extreme wind and rain event and the occurrence frequency of the second composite extreme wind and rain event are calculated respectively, and the calculation formula can be expressed as:

[0066]

[0067] Where F(i,year) is the cumulative frequency of compound extreme wind and rain events in one year (unit: times), year is the year sequence, and year=[y1,y2,y3,…,y n ], i represents the grid position, d is the day sequence, m is the month sequence, D m is the day of the month, Z （i,d,m，year） is the composite extreme wind and rain event occurring in the year th, month m, and day d in the ith grid.

[0068] Step 3: The frequency of the second composite extreme wind and rain event in the target area is evaluated based on the frequency of the first composite extreme wind and rain event, using the correlation coefficient, root mean square error and Taylor skill score of the climate model evaluation indicators commonly used in the meteorological field. The specific form is as follows:

[0069] Correlation coefficient: It is often used to quantify the correlation between different modes. The larger the correlation coefficient value, the higher the correlation between the frequency of the second composite extreme wind and rain event and the frequency of the first composite extreme wind and rain event. It can be expressed as:

[0070]

[0071] Where n represents the total number of grids, X i represents the frequency of the second composite extreme wind and rain event on the i-th grid after time averaging, Y i represents the frequency of the first composite extreme wind and rain event on the i-th grid after time averaging, and They respectively represent the average values ​​of the occurrence frequency of the second composite extreme wind and rain event and the occurrence frequency of the first composite extreme wind and rain event.

[0072] Root mean square error: It is used to measure the difference between the frequency of the second composite extreme wind and rain event and the frequency of the first composite extreme wind and rain event. The lower the root mean square error, the smaller the difference between the frequency of the second composite extreme wind and rain event and the frequency of the first composite extreme wind and rain event. It can be expressed as:

[0073]

[0074] Where n represents the total number of grids, X i represents the frequency of the second composite extreme wind and rain event on the i-th grid after time averaging, Y i Represents the frequency of the first composite extreme wind and rain event on the i-th grid after time averaging.

[0075] Taylor skill score: It is used to quantitatively evaluate the frequency of the second composite extreme wind and rain event. The higher the Taylor skill score, the more accurate the frequency of the second composite extreme wind and rain event. It can be expressed as:

[0076]

[0077] Among them, σ m and σ o They respectively represent the spatial standard deviation of the frequency of the second composite extreme wind and rain event and the frequency of the first composite extreme wind and rain event, COR and COR0 respectively represent the correlation coefficient and the maximum correlation coefficient of the selected frequency of the second composite extreme wind and rain event with the frequency of the first composite extreme wind and rain event;

[0078] According to the evaluation results, the frequency of the second composite extreme wind and rain events was selected as the characteristic value of the deep learning model, and the frequency of the first composite extreme wind and rain events at the same time and grid position was selected as the true value of the deep learning model.

[0079] Step 4, such as Figure 2As shown, a composite weighted loss function that integrates the mean squared error (MSE), the ratio of the standard deviation (RSD) indicators and the construction of the minimum value constraint function (Min_Loss), as well as a multi-layer perceptron neural network structure, is used to construct a climate model simulation deviation correction optimization model that integrates deep learning and is trained; wherein, the neural network structure of the model mainly includes an input layer, a hidden layer and an output layer; the input layer is the annual cumulative frequency of the composite extreme wind and rain events output by multiple models obtained in step S2; the hidden layer includes multiple fully connected layers and a discard layer, wherein the fully connected layer transmits the proposed feature information, and the introduction of the discard layer makes the network easier to train and better extract features; the output layer collects the feature information extracted by the hidden layer to generate the final simulation result; the specific steps are as follows:

[0080] Step 4.1 introduces an innovative loss function, namely the weighted loss function, which is used to measure the difference between the output of the deep learning model and the true value. This function combines the three functions of MSE, RSD and Min_Loss, and its specific form is as follows:

[0081] Mean square error: MSE in the weighted loss function is used in deep learning models to measure the difference between the model output and the true value. The specific formula is:

[0082]

[0083] Among them, y n The frequency of compound extreme wind and rain events simulated by the deep learning model, is the frequency of the first composite extreme wind and rain event, n is the nth randomly selected sample, and N represents the total number of randomly selected samples.

[0084] Ratio of standard deviation: used in climate model assessment to measure the difference in data distribution between the model and observations. By introducing RSD into the loss function, the model is designed to better capture the data distribution between the eigenvalues ​​and the observed values, thereby improving the accuracy of model simulation. Since the model parameter update adopts the gradient descent method, the lower the loss value, the better the model simulation effect. The closer the RSD value is to 1, the closer the simulation result is to the observed data distribution. The RSD function is adjusted. The specific formula is:

[0085]

[0086] Among them, σ y is the standard deviation of the frequency of composite extreme wind and rain events simulated by the deep learning model, is the standard deviation of the frequency of the first composite extreme wind and rain event.

[0087] Underestimation constraint function: It aims to consider the adjustment of simulated underestimation during model training to reduce simulation bias, which can be expressed as:

[0088]

[0089] Among them, y n The frequency of compound extreme wind and rain events simulated by the deep learning model, is the frequency of the first composite extreme wind and rain event, n is the nth randomly selected sample, and N represents the total number of randomly selected samples.

[0090] The weighted loss function combines the three functions of MSE, RSD and Min_Loss, and assigns weights to different functions in a weighted manner, aiming to adjust the performance of the model in different aspects. The specific formula is:

[0091]

[0092] Among them, Loss represents the degree of deviation between the model prediction value and the true value, MSE is the mean square error, W1 is the weight parameter of MSE, RSD is the ratio of standard deviation, W2 is the weight parameter of RSD, Min_Loss is the minimum value constraint function, W3 is the weight parameter of Min_Loss, W1, W2 and W3 represent the weight parameters of each function, which respectively adjust the relative importance of MSE, RSD and Min_Loss in the overall loss calculation, so that the model can better adapt to the task of simulating compound extreme wind and rain events.

[0093] Step 4.2, a deep learning optimization model is constructed using a multi-perceptron neural network structure. The model mainly includes an input layer, a hidden layer, and an output layer. Its specific form is:

[0094] Input layer: The model receives the annual cumulative frequency of compound extreme wind and rain events simulated by multiple climate models obtained in step S3, and integrates the output results of multiple climate models through the input layer to establish a one-dimensional feature set, which can be specifically expressed as:

[0095]

[0096] Where a represents the number of climate models, j represents the grid position of the randomly selected model simulation results, and y ’ Indicates the year of the random selection model simulation results, n ’ Indicates the number of randomly selected samples;

[0097] Hidden layer: This layer consists of multiple fully connected layers, each of which is followed by a dropout layer. The input layer integrates information from multiple model simulation results. The fully connected layer helps to effectively extract important feature information from high-dimensional data and capture complex nonlinear relationships in the data. The output of each fully connected layer is calculated as follows:

[0098]

[0099] Among them, W is the weight matrix, x is the feature set vector, b is the bias vector, and y is the LR is the feature information vector, is the activation function (Leaky Rectified Linear Unit); the dropout layer randomly "discards" the output results of neurons in the fully connected layer according to a certain probability, that is, only the output results of some neurons in the fully connected layer are considered during the training process. The introduction of the dropout layer helps to make the network easier to train and better extract features.

[0100] Output layer: The final output layer generates a simulation of compound extreme wind and rain events. This layer integrates the feature information proposed in the hidden layer and outputs the final result.

[0101] Step 4.3, train the model constructed in step 4.2. Use the feature values ​​and true values ​​obtained in step S3 to train the model, update the parameters in the model and finally get a trained model. The gradient descent method is mainly used to update the parameters between models. By calculating the loss value between the model simulation result and the true value, the back propagation calculation is further performed to update the parameters. The parameters are continuously updated iteratively until the loss result is relatively good. The specific parameter update calculation form is as follows:

[0102]

[0103] Among them, w is the weight parameter, b is the bias parameter, θ is the parameter set of w and b, θold(w, b) is the parameter set before the model parameters are updated, θnew(w, b) is the parameter set after the model parameters are updated, η is the learning rate, and L is the loss function.

[0104] Step 5: According to the model trained in step 4, Figure 3 As shown, the composite extreme wind and rain events simulated by multiple climate models for the target area are input into the model to obtain the improved results of the frequency of composite extreme wind and rain events simulated by the climate models in the target area.

[0105] The present invention is explained from the perspectives of purpose of use, efficiency, progress and novelty. It is practical and progressive and meets the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings of this application are only preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc. that are similar or identical to those of this application, that is, all equivalent replacements or modifications made within the scope of the patent application of this application, should fall within the scope of protection of the patent application of this application.

Claims

1. An optimization method for simulating the frequency of compound extreme wind and rain events in a climate model based on deep learning, characterized in that: The following steps are involved: S1. Obtain observation data and multiple climate model output data in the target area, and construct a grid network in the target area; S2. Based on the percentile threshold method, the observed rainfall threshold and the observed wind speed threshold are calculated according to the observed data, the climate model precipitation threshold and the climate model wind speed threshold are calculated according to the output data of multiple climate models, and the frequency of occurrence of the first composite extreme wind and rain event is calculated according to the observed rainfall threshold and the observed wind speed threshold, and the frequency of occurrence of the second composite extreme wind and rain event is calculated according to the climate model precipitation threshold and the climate model wind speed threshold; S3. Based on the observed data, the frequency of the second composite extreme wind and rain event is evaluated, and the frequency of the second composite extreme wind and rain event is selected as the characteristic value of the deep learning model according to the evaluation result, and the frequency of the first composite extreme wind and rain event at the same time and grid point location is selected as the true value of the deep learning model; S4. Based on a multi-layer perceptron neural network, a deep learning model is constructed to reduce the error value of the frequency of the second composite extreme wind and rain event, and the deep learning model is trained by a weighted loss function according to the characteristic value and the true value obtained in step S3; S5. Input the frequency of occurrence of the second composite extreme wind and rain event into the deep learning model trained in step S4 for training, and generate the observation constrained optimization result of the frequency of composite extreme wind and rain events in the target area; The observation data includes gridded data of precipitation and gridded data of maximum wind speed actually observed in the target area; The plurality of climate model output data include gridded precipitation data and gridded maximum wind speed data obtained by simulating a plurality of climate models in the target area; The step S2 comprises the following steps: S201. Based on the percentile threshold method, the observed rainfall threshold and the observed wind speed threshold are calculated according to the observed data, and the climate model precipitation threshold and the climate model wind speed threshold are calculated according to the output data of multiple climate models. The percentile threshold method can be expressed as: Where k is the percentile, N sorted (i, t) is the sequence of the i-th grid position at the t-th time and is sorted from small to large. is the length of the sequence, P(i,t) represents the N sorted Middle The threshold value of the position; S202. According to the obtained observed rainfall threshold, observed wind speed threshold, climate model precipitation threshold and climate model wind speed threshold, it is determined whether an extreme wind and rain event occurs in each grid of the target area grid network, which can be expressed as: Among them, P (i,t) and W (i,t) Respectively represent the precipitation and maximum wind speed at the tth time in the i-th grid, and Respectively represent the precipitation threshold and wind speed threshold at the tth time in the i-th grid; when the variable Z (i,t) When it is 1, it means that a composite extreme wind and rain event occurs at the corresponding time grid point, otherwise it does not occur; S203. According to the judgment result, the occurrence frequency of the first composite extreme wind and rain event and the occurrence frequency of the second composite extreme wind and rain event are calculated respectively, and the calculation formula can be expressed as: Among them, F(i,year) is the cumulative frequency of compound extreme wind and rain events in one year, year is the year sequence, and year=[y1,y2,y3,…,y n ], i represents the grid position, d is the day sequence, m is the month sequence, D m is the day of the month, Z （i,d,m，year） is the composite extreme wind and rain event occurring in the year th, month m, and day d in the ith grid.

2. The method for optimizing the frequency of composite extreme wind and rain events simulated by a climate model based on deep learning according to claim 1, characterized in that: The deep learning model in step S4 includes: The input layer is used to receive the occurrence of the second-frequency composite extreme wind and rain event, and randomly extract the frequency samples of the second-frequency composite extreme wind and rain event to establish a feature set, which can be expressed as: Where a represents the number of climate models, j represents the grid position of the randomly selected model simulation results, and y ’ Indicates the year of the random selection model simulation results, n ’ Indicates the number of randomly selected samples; The hidden layer includes a fully connected layer and a discard layer connected in sequence. There are multiple fully connected layers and discard layers. The fully connected layer is used to transmit feature information extracted from the feature set through an activation function. The activation function can be expressed as: Among them, W is the weight matrix, x is the feature set vector, b is the bias vector, and y is the LR is the feature information vector, is the activation function; The discard layer is used to filter the feature information output by the fully connected layer, and randomly receive the feature information vector output results of some neurons in the fully connected layer; The output layer is used to integrate the feature information extraction results and output the optimization results.

3. The method for optimizing the frequency of composite extreme wind and rain events simulated by a climate model based on deep learning according to claim 1, characterized in that: The weighted loss function in step S4 includes the mean square error, the ratio of standard deviation and the minimum value constraint function, and its calculation formula is: Among them, Loss represents the degree of deviation between the model prediction value and the true value, MSE is the mean square error, W1 is the weight parameter of MSE, RSD is the ratio of standard deviations, W2 is the weight parameter of RSD, Min_Loss is the minimum value constraint function, and W3 is the weight parameter of Min_Loss.

4. The method for optimizing the frequency of composite extreme wind and rain events simulated by a climate model based on deep learning according to claim 3, characterized in that: The MSE can be expressed as: Among them, y n The frequency of compound extreme wind and rain events simulated by the deep learning model, is the frequency of the first composite extreme wind and rain event, n is the nth randomly selected sample, and N represents the total number of randomly selected samples.

5. The method for optimizing the frequency of composite extreme wind and rain events simulated by a climate model based on deep learning according to claim 3, characterized in that: The RSD can be expressed as: in, is the standard deviation of the frequency of composite extreme wind and rain events simulated by the deep learning model, is the standard deviation of the frequency of the first composite extreme wind and rain event.

6. The method for optimizing the frequency of composite extreme wind and rain events simulated by a climate model based on deep learning according to claim 3, characterized in that: The Min_Loss can be expressed as: Among them, y n The frequency of compound extreme wind and rain events simulated by the deep learning model, is the frequency of the first composite extreme wind and rain event, n is the nth randomly selected sample, and N represents the total number of randomly selected samples.

7. The method for optimizing the frequency of composite extreme wind and rain events simulated by a climate model based on deep learning according to claim 1, characterized in that: The deep learning model used in step 5 updates parameters through back propagation calculation, and continuously iterates and updates parameters until the value of the weighted loss function converges to a preset range, thereby obtaining the observation-constrained optimization result of the frequency of composite extreme wind and rain events in the target area. The back propagation calculation can be expressed as: Among them, w is the weight parameter, b is the bias parameter, θ is the parameter set of w and b, θold(w, b) is the parameter set before the model parameters are updated, θnew(w, b) is the parameter set after the model parameters are updated, η is the learning rate, and L is the loss function.

Images