A method for interpreting numerical model forecast products and related devices

Abstract

The present invention discloses a method for interpreting a numerical model forecast product. First, a space-time neighbor relationship cuboid corresponding to the space-time position to be interpreted is generated through a space-time neighbor relationship neural network. Then, a space-time weight vector corresponding to the space-time position to be interpreted is generated based on the space-time neighbor relationship cuboid and the space-time weighted neural network. Finally, the interpretation result is obtained based on the corresponding space-time weight vector, the regression coefficient vector representing the overall average regression relationship, and the vector synthesized by the product to be interpreted and the auxiliary variable. The space-time neighbor relationship neural network is used to more accurately model the neighbor relationship between space-time positions, and the space-time weighted neural network is used to more accurately model the relationship between the numerical model forecast product and the auxiliary variables that change with time and space and the actual situation, thereby further improving the interpretation level of the numerical model forecast product. The present invention also provides a device, equipment and storage medium, which also have the above-mentioned beneficial effects.

Description

Technical Field

[0001] The present invention relates to the technical field of interpretation and utilization of earth science numerical model forecast products, and in particular to a numerical model forecast product interpretation and utilization method, a numerical model forecast product interpretation and utilization device, a numerical model forecast product interpretation and utilization equipment, and a computer-readable storage medium. Background Art

[0002] With the development of earth detection technology, meteorology, climatology, dynamic meteorology, information and communication technology, supercomputer technology, numerical calculation methods, data assimilation technology, etc., numerical weather forecasting has made great progress, and numerical model forecasting has become the basis of weather forecasting and climate prediction. However, due to the dynamic framework of numerical models, the physical process is only an approximation of the real atmosphere, the initial values ​​of the model obtained by data assimilation also inevitably have errors, and the chaotic characteristics of atmospheric motion, there are always uncertainties in the forecast products of numerical models. In addition, due to the differences in model resolution, model initialization, data assimilation technology and types of assimilated data, dynamic and physical parameterization schemes, and the description of terrain, different models have different forecast capabilities for different regions, different times, and different elements. Therefore, it is an important means to improve the forecast level of numerical models to revise the forecast products of numerical models or to integrate the forecast products of different numerical models. From the existing schemes, when correcting or integrating numerical model forecast products, some only use samples from a single grid point or station, some use all samples of a certain area at the same time, and some assume that all samples have the same importance. These problems may make it impossible for existing schemes to accurately interpret numerical model forecast products. The above problems exist not only in the interpretation of weather and climate numerical model forecast products, but also in the interpretation of ocean and other numerical model forecast products.

[0003] Therefore, how to propose a method for interpreting numerical model forecast products with wider applicability and higher accuracy is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0004] The purpose of the present invention is to provide a method for interpreting a numerical model forecast product, which has a better interpretation effect; another purpose of the present invention is to provide a numerical model forecast product interpretation device, a numerical model forecast product interpretation equipment and a computer-readable storage medium, which have a better interpretation effect.

[0005] In order to solve the above technical problems, the present invention provides a method for interpreting numerical model forecast products, comprising:

[0006] Acquire the product to be released and the auxiliary variables required for releasing the product to be released; the product to be released corresponds to the area to be released;

[0007] According to the space-time position to be interpreted and the pre-established regular space-time control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established;

[0008] Calling a spatiotemporal neighbor relationship neural network to obtain a spatiotemporal neighbor relationship cuboid according to the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid, and the time distance cuboid;

[0009] Calling a spatiotemporal weighted neural network to calculate a spatiotemporal weight vector according to the spatiotemporal proximity relationship cuboid;

[0010] The spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variable are multiplied to obtain an intermediate vector, and the intermediate vectors are summed to obtain the interpretation result; the regression coefficient vector is a regression coefficient vector determined according to the sample set during training by a preset regression model.

[0011] Optionally, the spatiotemporal proximity neural network includes multiple sub-networks, and the weights of the multiple sub-networks are shared.

[0012] Optionally, the spatiotemporal weighted neural network is any one of the following:

[0013] Convolutional neural network, fully connected neural network.

[0014] Optionally, the auxiliary variables include altitude and / or reporting time error.

[0015] Optionally, the obtaining of the product to be released and the auxiliary variables required for releasing the product to be released include:

[0016] The product to be released and the auxiliary variables required for releasing the product to be released are obtained and interpolated to a preset discrete grid.

[0017] Optionally, also include:

[0018] Construct a sample set; the samples in the sample set include observation data, historical numerical model forecast products corresponding to the observation data, and auxiliary variables corresponding to the observation data;

[0019] Determine the types of products to be released and auxiliary variables that will eventually be involved in the release, as well as the regression coefficient vector, through a preset regression model;

[0020] The spatiotemporal proximity neural network and the spatiotemporal weighted neural network are trained simultaneously through the sample set and the regression coefficient vector.

[0021] Optionally, the regression model includes:

[0022] lasso regression model, lasso-logistic regression model.

[0023] The observation data include:

[0024] Site observation data, buoy observation data, multi-source data fusion real-time data, and reanalysis data.

[0025] The present invention also provides a numerical model forecast product interpretation device, comprising:

[0026] A product acquisition module is used to acquire the product to be released and the auxiliary variables required for releasing the product to be released; the product to be released corresponds to the area to be released;

[0027] A building module is used to build a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid according to the space-time position to be released and a pre-established regular space-time control point array;

[0028] A spatiotemporal proximity module, used to call a spatiotemporal proximity neural network to obtain a spatiotemporal proximity cuboid according to the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid and the time distance cuboid;

[0029] A weight module, used for calling a spatiotemporal weighted neural network to calculate a spatiotemporal weight vector according to the spatiotemporal proximity cuboid;

[0030] The interpretation result module is used to multiply the spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variable to obtain an intermediate vector, and sum the intermediate vectors to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model.

[0031] The present invention also provides a numerical model forecast product interpretation device, the device comprising:

[0032] Memory: used to store computer programs;

[0033] Processor: used to implement the steps of the method for interpreting numerical model forecast products as described in any one of the above items when executing the computer program.

[0034] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method for interpreting numerical model forecast products as described in any one of the above items are implemented.

[0035] A method for interpreting a numerical model forecast product provided by the present invention includes obtaining a product to be interpreted and auxiliary variables required for interpreting the product to be interpreted; the product to be interpreted corresponds to an area to be interpreted; according to the space-time position to be interpreted and a pre-established regular space-time control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established; a space-time neighbor relationship neural network is called to obtain a space-time neighbor relationship cuboid according to the distance cuboid in the u-coordinate direction, the distance cuboid in the v-coordinate direction, and the time distance cuboid; a space-time weighted neural network is called to calculate a space-time weight vector according to the space-time neighbor relationship cuboid; an intermediate vector is obtained by multiplying the space-time weight vector, the regression coefficient vector, and a vector synthesized by the product to be interpreted and the auxiliary variables, and the intermediate vectors are summed to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model.

[0036] First, the spatiotemporal neighbor relationship cuboid corresponding to the spatiotemporal position to be interpreted is generated through the spatiotemporal neighbor relationship neural network, and then the spatiotemporal weight vector corresponding to the spatiotemporal position to be interpreted is generated based on the spatiotemporal neighbor relationship cuboid and the spatiotemporal weighted neural network. Finally, the interpretation result is obtained based on the spatiotemporal weight vector, the regression coefficient vector representing the overall average regression relationship, and the vector synthesized by the product to be interpreted and the auxiliary variable. By using the spatiotemporal neighbor relationship neural network to more accurately model the neighbor relationship between spatiotemporal positions, and by using the spatiotemporal weighted neural network to more accurately model the relationship between the numerical model forecast products and auxiliary variables that change with time and space and the actual situation, the interpretation level of the numerical model forecast products can be further improved.

[0037] The present invention also provides a numerical model forecast product interpretation device, a numerical model forecast product interpretation equipment and a computer-readable storage medium, which also have the above-mentioned beneficial effects and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 A flowchart of a method for interpreting a numerical model forecast product provided by an embodiment of the present invention;

[0040] Figure 2 for Figure 1 Model diagram for interpretation of numerical model forecast products;

[0041] Figure 3A flowchart of a specific method for interpreting numerical model forecast products provided by an embodiment of the present invention;

[0042] Figure 4 A structural block diagram of a numerical model forecast product interpretation device provided by an embodiment of the present invention;

[0043] Figure 5 A structural block diagram of a numerical model forecast product interpretation device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The core of the present invention is to provide a method for interpreting numerical model forecast products. From the existing solutions, when correcting or integrating numerical model forecast products, some only use samples from a single grid point or station, some use all samples of a certain area at the same time, and some assume that all samples have the same importance. These problems may cause the existing solutions to be unable to accurately describe the complex spatiotemporal relationship between numerical model forecast products and actual conditions, thereby making it impossible to accurately interpret numerical model forecast products.

[0045] The present invention provides a method for interpreting a numerical model forecast product, including obtaining a product to be interpreted and auxiliary variables required for interpreting the product to be interpreted; the product to be interpreted corresponds to an area to be interpreted; according to the spatiotemporal position to be interpreted and a pre-established regular spatiotemporal control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established; a spatiotemporal neighbor relationship neural network is called to obtain a spatiotemporal neighbor relationship cuboid according to the distance cuboid in the u-coordinate direction, the distance cuboid in the v-coordinate direction, and the time distance cuboid; a spatiotemporal weighted neural network is called to calculate a spatiotemporal weight vector according to the spatiotemporal neighbor relationship cuboid; an intermediate vector is obtained by multiplying the spatiotemporal weight vector, the regression coefficient vector, and a vector synthesized by the product to be interpreted and the auxiliary variables, and the intermediate vectors are summed to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model.

[0046] First, the spatiotemporal neighbor relationship cuboid corresponding to the spatiotemporal position to be interpreted is generated through the spatiotemporal neighbor relationship neural network, and then the spatiotemporal weight vector corresponding to the spatiotemporal position to be interpreted is generated based on the spatiotemporal neighbor relationship cuboid and the spatiotemporal weighted neural network. Finally, the interpretation result is obtained based on the spatiotemporal weight vector, the regression coefficient vector representing the overall average regression relationship, and the vector synthesized by the product to be interpreted and the auxiliary variable. By using the spatiotemporal neighbor relationship neural network to more accurately model the neighbor relationship between spatiotemporal positions, and by using the spatiotemporal weighted neural network to more accurately model the relationship between the numerical model forecast products and auxiliary variables that change with time and space and the actual situation, the interpretation level of the numerical model forecast products can be further improved.

[0047] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] Please refer to Figure 1 as well as Figure 2 , Figure 1 A flowchart of a method for interpreting a numerical model forecast product provided by an embodiment of the present invention; Figure 2 for Figure 1 Model diagram for interpreting numerical model forecast products.

[0049] See also Figure 1 as well as Figure 2 In an embodiment of the present invention, the method for interpreting a numerical model forecast product includes:

[0050] S101: Acquire the product to be released and the auxiliary variables required for releasing the product to be released.

[0051] In the embodiment of the present invention, the product to be released corresponds to the area to be released. The area to be released is a defined geographical range, such as Hainan Island or 10° to 20° north latitude, 100° to 130° east longitude, etc. The specific location of the area to be released needs to be set according to the actual situation and is not specifically limited here.

[0052] In this step, it is necessary to obtain the product to be interpreted, that is, the numerical model forecast product to be interpreted and the auxiliary variables required for the interpretation of the product to be interpreted. The auxiliary variables usually include but are not limited to altitude, error in the starting time, etc., that is, in the embodiment of the present invention, the auxiliary variables may optionally include altitude and / or error in the starting time. Taking altitude as an example, altitude will have a significant impact on forecast products such as temperature. In the following steps, it is necessary to refer to the auxiliary variables to generate the final interpretation result. The specific type of the auxiliary variables or whether they are needed can be set according to the actual situation, and no specific limitation is made here. It should be noted that the numerical model forecast product of the present invention can be an integrated interpretation of numerical model forecast products from multiple countries such as Europe, the United States, and China. For example, the final interpretation result = b0 + b1 * European model forecast product + b2 * American model forecast product + b3 * Chinese model forecast product, where b0, b1, b2, and b3 are only coefficients in the illustrative example.

[0053] It should also be noted that in the embodiments of the present invention, the above-mentioned numerical models include but are not limited to weather and climate numerical models, ocean numerical models, earth system models, etc.; the above-mentioned numerical model forecast products include but are not limited to temperature, precipitation, relative humidity, air pressure, wind speed, visibility, ground temperature, soil moisture predicted by weather and climate numerical models and earth system models, including but not limited to sea water temperature, salinity, density, and flow rate predicted by ocean numerical models; the above-mentioned products to be released can be deterministic forecasts, ensemble averages, or combinations thereof from one or more models.

[0054] S102: According to the space-time position to be interpreted and the pre-established regular space-time control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established.

[0055] The above-mentioned time-space position to be released is a grid point in a discrete grid pre-established in the area to be released during the release period as needed. In addition, in the embodiment of the present invention, it is necessary to first set a regular time-space control point array. In space, the above-mentioned regular time-space control point array covers the entire area to be released. In time, the above-mentioned regular time-space control point array divides the sample set into several time periods. The regular time-space control point array needs to be determined according to the size of the area to be released, the time span of the sample set, the computing power and the modeling accuracy requirements. The regular time-space control point array also constitutes a time-space cuboid. In this step, it is necessary to obtain the u-coordinate direction distance, v-coordinate direction distance, and time distance of all time-space control points in the time-space cuboid and the time-space position point i to be released, respectively forming a u-coordinate direction distance cuboid, a v-coordinate direction distance cuboid, and a time distance cuboid, that is, the input of the time-space neighbor relationship neural network. The above-mentioned sample set is the sample set required for training each network model, and the specific content of the sample set will be described in detail in the following invention embodiment.

[0056] It should be noted that in the embodiment of the present invention, corresponding distance cuboids can also be established according to other parameters related to the neighbor relationship, so as to set more types of distance cuboids as inputs of the neighbor relationship neural network. The types of other parameters related to the neighbor relationship can be set according to actual conditions and are not specifically limited here.

[0057] Typically, the spatiotemporal neighbor relationship neural network includes multiple subnetworks, and the number of subnetworks needs to be the same as the number of spatiotemporal control points in the above-mentioned spatiotemporal cuboid. In the following steps, any subnetwork needs to calculate the corresponding elements in the final spatiotemporal neighbor relationship cuboid according to the corresponding elements in the u-coordinate distance cuboid, the v-coordinate distance cuboid, and the time distance cuboid of the spatiotemporal position point i to be interpreted. In the embodiment of the present invention, the specific structure of each subnetwork, such as the number of hidden layers, is not specifically limited and depends on the specific situation.

[0058] S103: calling the spatiotemporal neighbor relationship neural network, and obtaining the spatiotemporal neighbor relationship cuboid according to the distance cuboid in the u-coordinate direction, the distance cuboid in the v-coordinate direction, and the time distance cuboid.

[0059] In this step, a pre-trained spatiotemporal proximity neural network is called, which includes multiple sub-networks. Each sub-network outputs the corresponding elements in the spatiotemporal proximity cuboid according to the corresponding elements in the u-coordinate distance cuboid, the v-coordinate distance cuboid, and the time distance cuboid of the spatiotemporal position point i to be interpreted. The resulting spatiotemporal proximity cuboid serves as the input of the spatiotemporal weighted neural network in the following steps.

[0060] Preferably, in the embodiment of the present invention, in order to reduce the amount of parameters, each of the above sub-networks will adopt a weight sharing strategy.

[0061] S104: Calling the spatiotemporal weighted neural network to calculate the spatiotemporal weight vector according to the spatiotemporal proximity relationship cuboid.

[0062] In the embodiment of the present invention, the above-mentioned spatiotemporal weighted neural network is usually a convolutional neural network. Of course, the spatiotemporal weighted neural network can also be specifically a fully connected neural network or other types of network models, depending on the specific circumstances, and is not specifically limited here. Usually, in the embodiment of the present invention, when the spatiotemporal weighted neural network is a convolutional neural network and the time dimension of the spatiotemporal neighbor relationship cuboid is greater than 1, the pooling layer in the spatiotemporal weighted neural network is a three-dimensional pooling layer, and the convolution layer is a three-dimensional convolution layer, so as to generate an accurate spatiotemporal weight vector.

[0063] In this step, the spatiotemporal weighted neural network will calculate the spatiotemporal weight vectors corresponding to the products to be interpreted and the auxiliary variables based on the above-mentioned spatiotemporal proximity relationship cuboid, so as to finally calculate the interpretation results in the following steps.

[0064] S105: Multiply the spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variables to obtain an intermediate vector, and sum the intermediate vectors to obtain the interpretation result.

[0065] In an embodiment of the present invention, the regression coefficient vector is a regression coefficient vector determined by a preset regression model according to a sample set during training. The regression coefficient vector represents the overall average regression relationship between the product to be released and the auxiliary variables and the actual situation, and the change in time and space between the product to be released and the auxiliary variables and the actual situation can be regarded as the fluctuation of the "average regression relationship". The specific content of the regression model will be described in detail in the following invention embodiments, and will not be repeated here. The specific content of the training and sample set will be described in detail in the following invention embodiments, and will not be repeated here.

[0066] In this step, it is necessary to multiply the corresponding elements of the above-mentioned spatiotemporal weight vector, regression coefficient vector, and the vector synthesized by the product to be released and the auxiliary variable to obtain an intermediate vector, and sum the intermediate vector to obtain the final interpretation result. It should be noted that in order to correspond to the "intercept" of the regression model, it is necessary to adaptively modify the vector synthesized by the above-mentioned product to be released and the auxiliary variable. It is usually necessary to add a constant before the vector synthesized by the product to be released and the auxiliary variable, usually 1, to form "a constant, a vector synthesized by the product to be released and the auxiliary variable".

[0067] A method for interpreting a numerical model forecast product provided by an embodiment of the present invention includes obtaining a product to be interpreted and auxiliary variables required for interpreting the product to be interpreted; the product to be interpreted corresponds to an area to be interpreted; according to the spatiotemporal position to be interpreted and a pre-established regular spatiotemporal control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established; a spatiotemporal neighbor relationship neural network is called to obtain a spatiotemporal neighbor relationship cuboid according to the distance cuboid in the u-coordinate direction, the distance cuboid in the v-coordinate direction, and the time distance cuboid; a spatiotemporal weighted neural network is called to calculate a spatiotemporal weight vector according to the spatiotemporal neighbor relationship cuboid; an intermediate vector is obtained by multiplying the spatiotemporal weight vector, the regression coefficient vector, and a vector synthesized by the product to be interpreted and the auxiliary variables, and the intermediate vectors are summed to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model.

[0068] First, the spatiotemporal neighbor relationship cuboid corresponding to the spatiotemporal position to be interpreted is generated through the spatiotemporal neighbor relationship neural network, and then the spatiotemporal weight vector corresponding to the spatiotemporal position to be interpreted is generated based on the spatiotemporal neighbor relationship cuboid and the spatiotemporal weighted neural network. Finally, the interpretation result is obtained based on the spatiotemporal weight vector, the regression coefficient vector representing the overall average regression relationship, and the vector synthesized by the product to be interpreted and the auxiliary variable. By using the spatiotemporal neighbor relationship neural network to more accurately model the neighbor relationship between spatiotemporal positions, and by using the spatiotemporal weighted neural network to more accurately model the relationship between the numerical model forecast products and auxiliary variables that change with time and space and the actual situation, the interpretation level of the numerical model forecast products can be further improved.

[0069] The specific contents of the method for interpreting and using a numerical model forecast product provided by the present invention will be described in detail in the following embodiments of the invention.

[0070] Please refer to Figure 3 , Figure 3 The present invention provides a flowchart of a specific method for interpreting numerical model forecast products.

[0071] See also Figure 3 In an embodiment of the present invention, the method for interpreting a numerical model forecast product includes:

[0072] S201: Construct a sample set.

[0073] In an embodiment of the present invention, the samples in the sample set include observation data, historical numerical model forecast products corresponding to the observation data, and auxiliary variables corresponding to the observation data. The observation data generally include: site observation data, buoy observation data, multi-source data fusion real-time data, and reanalysis data. Taking temperature as an example, the above-mentioned observation data can be selected from the historical measured temperature, the historical numerical model forecast product is the forecast temperature obtained by forecasting the above-mentioned measured temperature in history, and the auxiliary variable can be selected from the altitude of the above-mentioned corresponding observation data.

[0074] In this step, it is usually necessary to eliminate redundant samples and abnormal samples in the above sample set. The specific evaluation criteria for redundant samples and abnormal samples can be set according to actual conditions and are not specifically limited here.

[0075] S202: Determine the types of products to be released and auxiliary variables that will eventually be involved in the release, as well as the regression coefficient vector, through a preset regression model.

[0076] In this step, the types of products to be released and auxiliary variables that will eventually be used for release and the regression coefficient vector need to be determined through a preset regression model to establish an overall average regression relationship between the products to be released and auxiliary variables and the actual situation.

[0077] In an embodiment of the present invention, the regression model includes: lasso regression model, lasso-logistic regression model. Specifically, when there are multiple numerical model forecast products and auxiliary variables, the types of products to be interpreted and auxiliary variables that are ultimately involved in the interpretation and their regression coefficient vectors can be determined by the lasso regression model; when there is only one numerical model forecast product or auxiliary variable, the ordinary linear regression model is used to calculate the regression coefficient vector of the product to be interpreted or the auxiliary variable.

[0078] When logistic regression is needed to interpret the products to be interpreted, such as the weather forecast, when there are multiple numerical model forecast products and auxiliary variables, the lasso-logistic regression model can be used to determine the types of products to be interpreted and auxiliary variables and their regression coefficient vectors that will eventually participate in the interpretation. The lasso-logistic regression model is a combination of the lasso regression model and the logistic regression model; when there is only one numerical model forecast product or auxiliary variable, the logistic regression model is used to calculate the regression coefficient vector of the product to be interpreted or the auxiliary variable. Accordingly, the sigmoid activation function can be used after summing the intermediate vectors. Accordingly, during the training process, the loss function and evaluation index can be replaced with the cross entropy loss function and TS score.

[0079] S203: The spatiotemporal proximity neural network and the spatiotemporal weighted neural network are trained simultaneously through the sample set and the regression coefficient vector.

[0080] This step usually needs to determine the types of products to be released and auxiliary variables and regression coefficient vectors that will eventually participate in the release in S202, and then screen the sample set before executing. In this step, it is usually necessary to first randomly divide the above sample set into N parts, select one piece of data as a validation set each time, and the remaining data as a training set. After each piece of data becomes a validation set in turn, an N-fold cross-validation of the entire sample set is completed. The sample set can also be repeatedly randomly divided into N parts to complete multiple rounds of N-fold cross-validation, so as to achieve the simultaneous training of the above-mentioned spatiotemporal neighbor relationship neural network and spatiotemporal weighted neural network. It should be noted that during the training process, the above-mentioned spatiotemporal neighbor relationship neural network, spatiotemporal weighted neural network and the calculation process provided by the above-mentioned S102 and S105 need to be taken as a whole and the whole is trained. During the training process, the above-mentioned calculation process provided by the spatiotemporal neighbor relationship neural network, spatiotemporal weighted neural network, S102 and S105 will be specifically calculated according to the process provided by S102 to S105 provided by the above-mentioned embodiment of the invention. Its specific content can refer to the above-mentioned embodiment of the invention, and will not be repeated here.

[0081] In the specific training process, the embodiments of the present invention usually use mean square root error (MSE) as the loss function, He method as the parameter initialization method, PReLU or ReLU function as the activation function, and small batch stochastic gradient descent algorithm as the parameter update method, and adopt Dropout strategy, batch normalization technology and early stopping method to train the spatiotemporal neighbor relationship neural network and spatiotemporal weighted neural network. For the specific training process, please refer to the existing technology, which will not be repeated here.

[0082] It should be noted that, under normal circumstances, the above model training process and the following model use process are usually carried out in a rolling manner, that is, according to the passage of time, the above training process is interspersed during the use process to ensure the accuracy of the model. Of course, the specific execution order between the above model training process and the following use process needs to be set according to the actual situation, and no specific limitation is made here.

[0083] S204: Obtain the product to be released and the auxiliary variables required for releasing the product to be released and interpolate them to a preset discrete grid.

[0084] In this step, the products to be interpreted and the auxiliary variables need to be interpolated to the required discrete grids. For example, if the numerical model forecast product has a grid point of 9 kilometers, and the required discrete grid is a grid point of 1 kilometer, then the 9×9 km product needs to be interpolated to a grid point of 1×1 km. In this step, the products to be interpreted and the auxiliary variables required for the interpretation of the products to be interpreted need to be interpolated to the required 1×1 km discrete grid, so as to generate the final interpretation results later.

[0085] The rest of the content of this step is basically the same as S101 in the above-mentioned invention embodiment. Please refer to the above-mentioned invention embodiment for details, and will not be described again here.

[0086] S205: According to the space-time position to be interpreted and the pre-established regular space-time control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established.

[0087] S206: Calling the spatiotemporal neighbor relationship neural network, and obtaining the spatiotemporal neighbor relationship cuboid according to the distance cuboid in the u-coordinate direction, the distance cuboid in the v-coordinate direction, and the time distance cuboid.

[0088] S207: Call the spatiotemporal weighted neural network to calculate the spatiotemporal weight vector based on the spatiotemporal proximity relationship cuboid.

[0089] S208: Multiply the spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variables to obtain an intermediate vector, and sum the intermediate vectors to obtain the interpretation result.

[0090] The above S205 to S208 are basically the same as S102 to S105 in the above invention embodiment. Please refer to the above invention embodiment for details, and no further description is given here. The neural network called in the above S206 and S207 is the neural network trained by the above S201 to S202.

[0091] It should be noted that, in the embodiment of the present invention, the area to be used is usually divided into multiple sub-areas. For each sub-area, in the embodiment of the present invention, at least the above S202 to S208 are required to obtain the corresponding interpretation results of each sub-area, and finally the multiple interpretation results are synthesized into the final interpretation result. Dividing the area to be interpreted into multiple sub-areas for separate interpretation can effectively reduce the amount of calculation and further improve the interpretation effect.

[0092] The present invention provides a method for interpreting a numerical model forecast product. First, a space-time neighbor relationship cuboid corresponding to the space-time position to be interpreted is generated through a space-time neighbor relationship neural network. Then, a space-time weight vector corresponding to the space-time position to be interpreted is generated based on the space-time neighbor relationship cuboid and the space-time weighted neural network. Finally, the interpretation result is obtained based on the corresponding space-time weight vector, the regression coefficient vector representing the overall average regression relationship, and the vector synthesized by the product to be interpreted and the auxiliary variable. By using the space-time neighbor relationship neural network to more accurately model the neighbor relationship between space-time positions, and by using the space-time weighted neural network to more accurately model the relationship between the numerical model forecast product and the auxiliary variables and the actual situation that changes with time and space, the interpretation level of the numerical model forecast product is further improved.

[0093] A numerical model forecast product interpretation device provided in an embodiment of the present invention is introduced below. The numerical model forecast product interpretation device described below and the numerical model forecast product interpretation method described above can correspond to each other.

[0094] Please refer to Figure 4 , Figure 4 A structural block diagram of a numerical model forecast product interpretation device provided by an embodiment of the present invention, referring to Figure 4 , the numerical model forecast product interpretation device may include:

[0095] The product acquisition module 100 is used to acquire the product to be released and the auxiliary variables required for releasing the product to be released; the product to be released corresponds to the area to be released;

[0096] Establishing module 200, for establishing a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid according to the space-time position to be released and a pre-established regular space-time control point array;

[0097] The spatiotemporal proximity module 300 is used to call the spatiotemporal proximity relationship neural network to obtain a spatiotemporal proximity relationship cuboid according to the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid and the time distance cuboid;

[0098] A weight module 400 is used to call a spatiotemporal weighted neural network to calculate a spatiotemporal weight vector according to the spatiotemporal neighbor relationship cuboid;

[0099] The interpretation result module 500 is used to multiply the spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variable to obtain an intermediate vector, and sum the intermediate vectors to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model.

[0100] Preferably, in an embodiment of the present invention, the spatiotemporal proximity neural network includes multiple sub-networks, and the weights of the multiple sub-networks are shared.

[0101] Preferably, in an embodiment of the present invention, the spatiotemporal weighted neural network is any one of the following:

[0102] Convolutional neural network, fully connected neural network.

[0103] Preferably, in an embodiment of the present invention, the auxiliary variable may be altitude and / or reporting time error.

[0104] Preferably, in the embodiment of the present invention, the product acquisition module 100 is specifically used for:

[0105] The product to be released and the auxiliary variables required for releasing the product to be released are obtained and interpolated to a preset discrete grid.

[0106] Preferably, in the embodiment of the present invention, it also includes:

[0107] A sample set construction module is used to construct a sample set; the samples in the sample set include observation data, historical numerical model forecast products corresponding to the observation data, and auxiliary variables corresponding to the observation data;

[0108] A regression coefficient module, used to determine the types of the products to be released and auxiliary variables that are ultimately involved in the release and the regression coefficient vector through a preset regression model;

[0109] A training module is used to simultaneously train the spatiotemporal proximity neural network and the spatiotemporal weighted neural network through the sample set.

[0110] Preferably, in an embodiment of the present invention, the regression model includes:

[0111] lasso regression model, lasso-logistic regression model.

[0112] The numerical model forecast product interpretation device of the present embodiment is used to implement the aforementioned numerical model forecast product interpretation method. Therefore, the specific implementation method of the numerical model forecast product interpretation device can be seen in the embodiment part of the numerical model forecast product interpretation method in the previous text. For example, the product acquisition module 100, the establishment module 200, the spatiotemporal proximity module 300, the weight module 400, and the interpretation result module 500 are respectively used to implement steps S101 to S105 in the aforementioned numerical model forecast product interpretation method. Therefore, its specific implementation method can refer to the description of the corresponding various partial embodiments and will not be repeated here.

[0113] A numerical model forecast product interpretation device provided in an embodiment of the present invention is introduced below. The numerical model forecast product interpretation device described below and the numerical model forecast product interpretation method and numerical model forecast product interpretation apparatus described above can be referenced to each other.

[0114] Please refer to Figure 5 , Figure 5 A structural block diagram of a numerical model forecast product interpretation device provided in an embodiment of the present invention.

[0115] Reference Figure 5 The numerical model forecast product interpretation device may include a processor 11 and a memory 12.

[0116] The memory 12 is used to store computer programs; the processor 11 is used to implement the specific content of the numerical model forecast product interpretation method described in the above-mentioned embodiment of the invention when executing the computer program.

[0117] The processor 11 in the numerical model forecast product interpretation device of this embodiment is used to install the numerical model forecast product interpretation device described in the above invention embodiments, and the processor 11 combined with the memory 12 can implement the numerical model forecast product interpretation method described in any of the above invention embodiments. Therefore, the specific implementation of the numerical model forecast product interpretation device can be seen in the embodiment of the numerical model forecast product interpretation method in the above text, and its specific implementation can refer to the description of the corresponding various parts of the embodiment, which will not be repeated here.

[0118] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, a method for interpreting a numerical model forecast product introduced in any of the above-mentioned embodiments of the invention is implemented. The rest of the content can refer to the prior art and will not be described in detail here.

[0119] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0120] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0121] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0122] Finally, 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. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0123] The above is a detailed introduction to a numerical model forecast product interpretation method, a numerical model forecast product interpretation device, a numerical model forecast product interpretation equipment and a computer-readable storage medium provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for interpreting numerical model forecast products, characterized in that: include: Acquire the product to be released and the auxiliary variables required for releasing the product to be released; The product to be released corresponds to the area to be released; According to the space-time position to be interpreted and the pre-established regular space-time control point array, a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid are established; Calling a spatiotemporal neighbor relationship neural network to obtain a spatiotemporal neighbor relationship cuboid according to the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid, and the time distance cuboid; Calling a spatiotemporal weighted neural network to calculate a spatiotemporal weight vector according to the spatiotemporal proximity relationship cuboid; The spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variable are multiplied to obtain an intermediate vector, and the intermediate vectors are summed to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model; The spatiotemporal neighbor relationship neural network includes multiple sub-networks, the number of which is the same as the number of spatiotemporal control points in the regular spatiotemporal control point array, and any sub-network is used to calculate the corresponding elements in the spatiotemporal neighbor relationship cuboid based on the corresponding elements in the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid, and the time distance cuboid of the spatiotemporal position to be interpreted.

2. The method according to claim 1, characterized in that The spatiotemporal proximity neural network includes multiple sub-networks, and the weights of the multiple sub-networks are shared.

3. The method according to claim 1, characterized in that The spatiotemporal weighted neural network is any one of the following: Convolutional neural network, fully connected neural network.

4. The method according to claim 1, characterized in that: The auxiliary variables include the altitude and / or the error of the alarm start time.

5. The method according to claim 1, characterized in that: The acquisition of the product to be released and the auxiliary variables required for releasing the product to be released include: The product to be released and the auxiliary variables required for releasing the product to be released are obtained and interpolated to a preset discrete grid.

6. The method according to any one of claims 1 to 5, characterized in that: Also includes: Construct a sample set; The samples in the sample set include observation data, historical numerical model forecast products corresponding to the observation data, and auxiliary variables corresponding to the observation data; Determine the types of products to be released and auxiliary variables that will eventually be involved in the release, as well as the regression coefficient vector, through a preset regression model; The spatiotemporal proximity neural network and the spatiotemporal weighted neural network are trained simultaneously through the sample set and the regression coefficient vector.

7. The method according to claim 6, characterized in that The regression model includes: lasso regression model, lasso-logistic regression model; The observation data include: site observation data, buoy observation data, multi-source data fusion real-time data, and reanalysis data.

8. A numerical model forecast product interpretation device, characterized in that: include: A product acquisition module, used to acquire the product to be released and the auxiliary variables required for releasing the product to be released; The product to be released corresponds to the area to be released; A building module is used to build a distance cuboid in the u-coordinate direction, a distance cuboid in the v-coordinate direction, and a time distance cuboid according to the space-time position to be released and a pre-established regular space-time control point array; A spatiotemporal proximity module, used to call a spatiotemporal proximity neural network to obtain a spatiotemporal proximity cuboid according to the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid and the time distance cuboid; A weight module, used for calling a spatiotemporal weighted neural network to calculate a spatiotemporal weight vector according to the spatiotemporal proximity cuboid; An interpretation result module is used to multiply the spatiotemporal weight vector, the regression coefficient vector, and the vector synthesized by the product to be interpreted and the auxiliary variable to obtain an intermediate vector, and sum the intermediate vectors to obtain an interpretation result; the regression coefficient vector is a regression coefficient vector determined according to a sample set during training by a preset regression model; The spatiotemporal neighbor relationship neural network includes multiple sub-networks, the number of which is the same as the number of spatiotemporal control points in the regular spatiotemporal control point array, and any sub-network is used to calculate the corresponding elements in the spatiotemporal neighbor relationship cuboid based on the corresponding elements in the u-coordinate direction distance cuboid, the v-coordinate direction distance cuboid, and the time distance cuboid of the spatiotemporal position to be interpreted.

9. A numerical model forecast product interpretation device, characterized in that: The device comprises: Memory: used to store computer programs; Processor: used to implement the steps of the method for interpreting numerical model forecast products as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for interpreting numerical model forecast products as claimed in any one of claims 1 to 7.

Images