A method, device and equipment for processing numerical weather forecast data

Abstract

The present application provides a method, apparatus and processing equipment for processing numerical weather forecast data, focusing on the temperature correction target of numerical weather forecast data. For the numerical weather forecast data processing model, on the one hand, a grid subdivision mechanism is introduced to constrain the model processing unit, reduce the model complexity, reduce the GPU video memory occupied during processing, and effectively increase the sample size during model training. On the other hand, the influence of various meteorological factors and terrain features of adjacent / surrounding areas on the deviation is fully considered, so that the correction target can be achieved efficiently and accurately, and the generated refined temperature grid data can meet the application requirements of high-quality numerical weather forecast data.

Description

Technical Field

[0001] The present application relates to the field of weather forecasting, and in particular to a method, device and processing equipment for numerical weather forecast data. Background Art

[0002] Numerical Weather Prediction (NWP) is the mainstream method of weather forecasting. It solves the state of the atmosphere grid by grid through numerical integration, and has achieved great success in extreme disaster warning, climate change prediction and other fields. However, with the continuous improvement of NWP's temporal and spatial resolution, huge demands have been placed on computing power. At the same time, increasingly complex and imperfect microphysical processes, parameterization schemes and discretization algorithms inevitably introduce systematic errors, and the bottleneck of traditional NWP has become increasingly prominent.

[0003] In response to this, existing technologies have begun to explore new weather forecasting paradigms, such as using pure data-driven machine learning methods to predict future weather. In recent years, with the development of computer science, many new methods have emerged in the field of machine learning, which can mine and analyze data more efficiently. Data-driven weather forecasting has shown great potential and has achieved a series of breakthroughs.

[0004] However, the current data-driven artificial intelligence (AI) meteorological models also exhibit a series of problems such as low physical consistency and too few prediction elements. In addition, the emergence of AI meteorological models is mainly due to high-quality data sets, and these models all use ECMWF's 0.25° resolution ERA5 reanalysis data set, which means that the AI ​​model forecast resolution can only be 0.25°. If the resolution of AI meteorological models needs to be further improved, in addition to being limited by the data set, it will also encounter the same problems as the development of NWP models. The increase in resolution will greatly lead to an exponential increase in computing power when training AI meteorological models.

[0005] Thus, although NWP models and AI meteorological models have made remarkable achievements in forecasting large-scale weather systems, they still have certain limitations in predicting ground temperature at a local fine scale, and these limitations are mainly reflected in the insufficient simulation of smaller-scale meteorological processes such as complex terrain and urban heat island effect. However, many aspects such as agriculture, transportation, energy, and daily life require more accurate and detailed temperature forecasts. NWP models are limited by computing power and microphysical parameterization schemes, and the degree of refinement of temperature forecasts is limited; AI meteorological models are limited by computing power and training data sets, and the temporal and spatial resolution is generally not high.

[0006] In order to quickly and efficiently provide accurate high-temporal and spatial resolution temperature forecast products, the NWP-based model temperature post-processing bias correction scheme and spatiotemporal downscaling scheme are still the most effective ways to solve the current bottleneck problem.

[0007] In order to correct the deviation of NWP model output products, existing technologies have proposed many methods, including MOS linear regression, Bayesian model, Kalman filter, as well as machine learning algorithms represented by tree models such as random forest, XGBoost, LightGBM, and deep learning methods based on neural networks such as CNN and transformer. The main idea of ​​these methods is that NWP model output products may have deviations in some aspects, but these deviations are often predictable. For example, if a model systematically predicts high temperatures under certain conditions, the deviation can be corrected by optimizing the model output by combining historical NWP forecasts and historical site observation data. Practice has proved that these technologies can indeed effectively improve the quality of NWP forecasts.

[0008] In order to obtain NWP model output products with high spatiotemporal resolution, many downscaling schemes have been developed in the existing technology, which are mainly divided into dynamic downscaling and statistical downscaling. The most commonly used dynamic downscaling is a nested high-resolution regional model based on a limited target range. Compared with the global NWP, it can relatively reduce the amount of calculation, but the amount of calculation is still large, and it is also limited by people's understanding of atmospheric physics, and the accuracy is difficult to meet the needs; the statistical downscaling method usually uses statistical models to establish the relationship between low-resolution data and high-resolution observation data. These statistical models mainly include linear regression, decision tree, MLP, etc. With the development of deep learning technology, the existing technology has begun to explore the application of deep learning models to statistical downscaling tasks to take advantage of its advantages in processing complex nonlinear relationships and large-scale data sets. The application of downscaling has developed rapidly and achieved success. For example, someone first used the image super-resolution deep learning method in the downscaling process, integrating high-resolution terrain data to improve the resolution of precipitation data. Since then, more improved deep learning methods have been applied to the field of meteorological element downscaling.

[0009] However, for NWP model output products, the deviation correction methods used in existing technologies are mostly based on station observation data. In practical applications, it is found that there are still problems with low correction accuracy or unstable correction accuracy. This means that certain influencing factors are still ignored in the correction process, which brings deviations to the correction target and affects the correction effect. Summary of the invention

[0010] The present application provides a method, apparatus and processing equipment for processing numerical weather forecast data, focusing on the temperature correction target of numerical weather forecast data. For the numerical weather forecast data processing model, on the one hand, a grid subdivision mechanism is introduced to constrain the model processing unit, reduce the model complexity, reduce the GPU video memory occupied during processing, and effectively increase the sample size during model training. On the other hand, the influence of various meteorological factors and terrain features of adjacent / surrounding areas on the deviation is fully considered, so that the correction target can be achieved efficiently and accurately, and the generated refined temperature grid data can meet the application requirements of high-quality numerical weather forecast data.

[0011] In a first aspect, the present application provides a method for processing numerical weather forecast data, the method comprising:

[0012] Acquiring initial numerical weather forecast data obtained through numerical weather forecast processing in the target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features;

[0013] Input the initial numerical weather forecast data into a preconfigured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used to carry out correction processing on the numerical weather forecast data input into the model, and the numerical weather forecast data processing model is obtained by training the initial deep learning model with labeled sample numerical weather forecast data. During the working process, the numerical weather forecast data processing model carries out temperature correction processing based on multiple subdivided grids, and during the temperature correction processing, the deviation generated by the numerical weather forecast data of the grids surrounding the subdivided grids is considered for correction, and the multiple subdivided grids are obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model according to the preset number of grids or the preset grid size based on the preset resolution;

[0014] Extract the target numerical weather forecast data with subdivided grid granularity output by the numerical weather forecast data processing model

[0015] In a second aspect, the present application provides a device for processing numerical weather forecast data, the device comprising:

[0016] An acquisition unit, used to acquire initial numerical weather forecast data obtained by numerical weather forecast processing in a target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features;

[0017] A processing unit, used for inputting initial numerical weather forecast data into a preconfigured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used for performing correction processing on the numerical weather forecast data input into the model, the numerical weather forecast data processing model is obtained by training an initial deep learning model with labeled sample numerical weather forecast data, and in the working process, the numerical weather forecast data processing model performs temperature correction processing based on a plurality of subdivided grids, and in the temperature correction processing process, the deviation generated by the numerical weather forecast data of the grids surrounding the subdivided grids is considered for correction, and the plurality of subdivided grids are obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model, based on a preset resolution, according to a preset number of grids or a preset grid size;

[0018] The extraction unit is used to extract the target numerical weather forecast data with subdivided grid granularity output by the numerical weather forecast data processing model.

[0019] In a third aspect, the present application provides a processing device, including a processor and a memory, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the method provided in the first aspect of the present application or any possible implementation method of the first aspect of the present application is executed.

[0020] In a fourth aspect, the present application provides a computer-readable storage medium, which stores multiple instructions, and the instructions are suitable for a processor to load to execute the method provided by the first aspect of the present application or any possible implementation of the first aspect of the present application.

[0021] From the above content, it can be concluded that the present application has the following beneficial effects:

[0022] This application focuses on the temperature correction target of numerical weather forecast data. For the numerical weather forecast data processing model, on the one hand, a grid subdivision mechanism is introduced to constrain the model processing unit, reduce the model complexity, and reduce the GPU video memory occupied during processing. At the same time, it can also effectively increase the sample size during model training. On the other hand, it also fully considers the impact of various meteorological factors and terrain features in adjacent / surrounding areas on the deviation, so that delicate and high-precision temperature grid data can be generated to meet the application needs of high-quality numerical weather forecast data. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0024] Figure 1 A flowchart of a method for processing numerical weather forecast data for this application;

[0025] Figure 2 A schematic diagram of the structure of a device for processing numerical weather forecast data in this application;

[0026] Figure 3 A schematic diagram of the structure of the processing equipment of this application. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0028] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices. The naming or numbering of steps in this application does not mean that the steps in the method flow must be executed in the time / logical sequence indicated by the naming or numbering. The process steps that have been named or numbered can change the execution order according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved.

[0029] The division of modules in this application is a logical division. There may be other division methods when it is implemented in actual applications. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between modules can be electrical or other similar forms, which are not limited in this application. In addition, the modules or submodules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules, and some or all of the modules may be selected according to actual needs to achieve the purpose of the present application.

[0030] Before introducing the processing method of numerical weather forecast data provided by this application, the background content involved in this application is first introduced.

[0031] The numerical weather forecast data processing method, device and computer-readable storage medium provided in the present application can be applied to processing equipment. For the numerical weather forecast data processing model, on the one hand, a grid subdivision mechanism is introduced to constrain the model processing unit, reduce the model complexity, reduce the GPU video memory occupied during processing, and effectively increase the sample size during model training. On the other hand, the influence of various meteorological factors and terrain features of adjacent / surrounding areas on the deviation is fully considered, so that refined temperature grid data can be generated to meet the application needs of high-quality numerical weather forecast data.

[0032] The execution subject of the method for processing numerical weather forecast data mentioned in this application can be a processing device for numerical weather forecast data, or a server, physical host or user equipment (UE) and other different types of processing equipment that integrate the processing device for numerical weather forecast data. Among them, the processing device for numerical weather forecast data can be implemented in hardware or software, and the UE can be a terminal device such as a smart phone, tablet computer, laptop computer, desktop computer or personal digital assistant (PDA), and the processing device can be set in the form of a device cluster.

[0033] It can be understood that the present application is a secondary processing based on the numerical weather forecast output products (i.e., NWP output products) provided by the existing numerical weather forecast models. Therefore, the processing equipment that executes the processing method of the numerical weather forecast data provided by the present application, or the processing equipment equipped with the application service corresponding to the processing method of the numerical weather forecast data provided by the present application, only needs to meet the required data processing capabilities, and can be configured as different types of devices such as servers, physical hosts, and even UEs according to actual application scenarios and actual needs, and can also adopt a remote / cloud processing architecture, which is more flexible and can be adapted to actual conditions.

[0034] Next, we will introduce the processing method of numerical weather forecast data provided by this application.

[0035] First, see Figure 1 , Figure 1 A flow chart of a method for processing numerical weather forecast data of the present application is shown. The method for processing numerical weather forecast data provided by the present application may specifically include the following steps S101 to S103:

[0036] Step S101, obtaining initial numerical weather forecast data obtained by numerical weather forecast processing for the target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features;

[0037] It can be understood that weather forecast demand is of regional nature. Therefore, the area of ​​numerical weather forecast data that currently requires secondary processing is recorded as the target area. At the same time, the currently acquired and ready-made numerical weather forecast data is recorded as the initial numerical weather forecast data.

[0038] In actual situations, the secondary processing of numerical weather forecast data is usually initiated in the form of tasks, which can be manually entered tasks, or tasks initiated by the device according to a pre-configured autonomous initiation strategy, or tasks forwarded by other devices. Obviously, the specific form of task initiation is more flexible.

[0039] The acquisition and processing of initial numerical weather forecast data can be either manual entry, local reading, or forwarding by other devices. The specific acquisition method is also relatively flexible.

[0040] It is worth noting that the numerical weather forecast processing involved here, that is, the numerical weather forecast processing that relies on the corresponding numerical weather forecast model (NWP model), belongs to the category of prior art, which corresponds to the situation where the initial numerical weather forecast data is ready-made data.

[0041] In addition, it is also necessary to note that what this application specifically needs to do is to perform secondary processing on the temperature content in order to achieve the effect of correction and downscaling, but other aspects need to be involved in the processing process. Therefore, the initial numerical weather forecast data obtained here not only contains the temperature content, but also other aspects, which naturally include the elements that this application believes can affect the temperature, thereby providing the corresponding data required for data processing.

[0042] As an example, the initial numerical weather forecast data here can be the IFS product of ECMWF, which is output twice a day at 00:00 UTC and 12:00 UTC, with a horizontal resolution of 0.125° for ground meteorological elements, 0.25° for high-altitude pressure layer elements, and a temporal resolution of 3h.

[0043] Step S102, inputting the initial numerical weather forecast data into a pre-configured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used to carry out correction processing on the numerical weather forecast data input into the model, and the numerical weather forecast data processing model is obtained by training the initial deep learning model with labeled sample numerical weather forecast data. During the working process, the numerical weather forecast data processing model carries out temperature correction processing based on a plurality of subdivided grids, and during the temperature correction processing process, the deviation generated by the numerical weather forecast data of the grids surrounding the subdivided grids is considered for correction, and the plurality of subdivided grids are obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model, based on a preset resolution, according to a preset number of grids or a preset grid size;

[0044] It can be understood that, in order to achieve the secondary processing goal of numerical weather forecast data, this application has configured a corresponding processing model based on deep learning technology, namely, a numerical weather forecast data processing model. On this basis, the corresponding model architecture and processing logic of the model are also configured in a targeted manner to implement the numerical weather forecast data processing strategy specially configured in this application.

[0045] In this regard, the method of the present application may also include a training process of the model, namely:

[0046] By training the initial deep learning model with labeled sample numerical weather forecast data, we can obtain the numerical weather forecast data processing model.

[0047] Specifically, the numerical weather forecast data processing model configured in the present application involves the setting of regional subdivision in the process of achieving the primary goal of correcting the temperature data in the present application. By dividing the original entire area into multiple sub-areas and carrying out correction processing in sub-areas, the model complexity can be effectively reduced. While reducing the GPU video memory occupied during processing, it can also effectively increase the sample size during model training, thereby improving processing efficiency and processing accuracy.

[0048] The division of these sub-areas can be completed through a grid subdivision mechanism. For the original, entire area, a preliminary regional division can be performed based on a preset resolution (such as 0.05° resolution), and then the division can be continued according to a preset number of grids or a preset grid size. For example, it can be finally divided into 299 7×7 subdivision grids. If the division is performed with a preset grid size, it can be understood that the number of final subdivision grids varies with the specific area size, but the size of each subdivision grid remains unchanged.

[0049] In addition, this application also believes that for the output products of numerical weather forecast models, the deviation correction methods used in the existing technology are mostly based on station observation data for correction, and fail to fully utilize the influence of various meteorological elements and terrain features in the surrounding areas of the station on the deviation. For example, the statistical downscaling scheme based on deep learning in the existing technology only uses the ground meteorological elements output by the numerical weather forecast model, and does not consider the impact of weather circulation situations in a larger scale range on the ground temperature, and fails to fully utilize the advantage of the relatively accurate forecast of high-altitude meteorological elements of numerical weather forecasts.

[0050] In this regard, the present application fully considers the impact of various meteorological factors and terrain features in adjacent / surrounding areas on the deviation. In this case, in the correction process based on multiple subdivided grids, the model also considers the deviation caused by the numerical weather forecast data of the grids surrounding the subdivided grid (including various meteorological factors and terrain features) and corrects the subdivided grid granularity / spatial resolution. In this way, delicate and high-precision temperature grid data can be corrected. The temperature data of different subdivided grids are spliced ​​together to obtain the temperature data of the entire area, forming the target numerical weather forecast data of the subdivided grid granularity.

[0051] As a supplement, the training process of the numerical weather forecast data processing model, like the conventional model training process, includes:

[0052] Collect sample numerical weather forecast data and mark them (true value, used to assist / guide model training). After marking and configuring the initialized deep learning model, you can start specific model training. In each round of model training, input the sample numerical weather forecast data into the model so that the model can perform corresponding processing to achieve forward propagation. Then, combine the processing results output by the model and the corresponding marking results to calculate the loss function, and reversely optimize the model parameters according to the loss function calculation results to achieve back propagation. When the training requirements such as the number of training times, training time or processing accuracy are met, the model training is completed and a numerical weather forecast data processing model that can be put into practical use is obtained.

[0053] As an example, the sample numerical weather forecast data can be the output product of the Land Data Assimilation System of China Meteorological Administration (CLDAS), with a horizontal resolution of 0.05° and a temporal resolution of 1h. The CLDAS output product is based on the numerical weather forecast analysis field, using observation data from ground meteorological stations, satellites and other sources, and is developed using multi-grid variational assimilation, physical inversion, terrain correction and other technologies, which can be used as training samples for this application.

[0054] After collecting the training samples, it is easy to understand that we can also divide them into training sets, test sets, and validation sets. These are configured according to the specific model training plan. For example, we can take the first 80% sorted by time as the training set and the remaining 20% ​​as the validation set.

[0055] In addition, taking the IFS product as an example, the forecast data of the IFS product is provided twice a day at 00:00 UTC and 12:00 UTC. Due to the time consumption of NWP operation and data distribution, it is often delayed by more than 6 hours to reach the user end. Taking this into account, the test set of this application can be a numerical weather forecast data input model with a forecast time of 12-36 hours each time, so as to simulate the actual business operation environment. In other words, the training samples can also be configured considering the time conditions of the actual business operation environment.

[0056] Step S103, extracting the target numerical weather forecast data with subdivided grid granularity output by the numerical weather forecast data processing model.

[0057] It should be understood that the present application is specifically aimed at the temperature data in the numerical weather forecast data, and after correcting it through the numerical weather forecast data processing model, the corrected temperature data can be obtained. At this time, the target numerical weather forecast data output by the model can include only the content of the corrected temperature data, or can include the corrected temperature data and the content of other original numerical weather forecast data, which can be flexibly configured.

[0058] After obtaining the target numerical weather forecast data, it is easy to understand and can be stored, or its content can be displayed through a visual interface (depending on the device's own display screen, the display screens of other devices or external display devices), or forwarded, or a prompt for completion of correction can be output, or the next step of data processing can be carried out... Obviously, the subsequent specific data applications can be flexibly adjusted according to the pre-configured data application requirements or the real-time configured data application requirements.

[0059] from Figure 1 It can be seen from the illustrated embodiment that the focus is on the temperature correction target of the numerical weather forecast data. For the numerical weather forecast data processing model, on the one hand, a grid subdivision mechanism is introduced to constrain the model processing unit, reduce the model complexity, and reduce the GPU video memory occupied during processing. At the same time, it can also effectively increase the sample size during model training. On the other hand, it also fully considers the influence of various meteorological factors and terrain features in adjacent / surrounding areas on the deviation, so that the correction target can be achieved efficiently and accurately, and the generated refined temperature grid data can meet the application requirements of high-quality numerical weather forecast data.

[0060] Continue to the above Figure 1 Each step of the illustrated embodiment and possible implementation methods thereof in practical applications are described in detail.

[0061] As an exemplary embodiment, corresponding to the processing content of the numerical weather forecast data processing model mentioned above, the preset resolution involved can specifically be 0.05° resolution, which means that the sample numerical weather forecast data and the target numerical weather forecast data output after the model is put into actual use are both 0.05° resolution. Correspondingly, the preset grid size can specifically be 7×7 geographic grid sizes at 0.05° resolution.

[0062] It can be understood that on the basis that the numerical weather forecast data processing model configured in the present application can obtain delicate and high-precision temperature grid data, it can be seen from the embodiment here that it can specifically involve temperature grid data with a resolution of 0.05° and a 7×7 grid, that is, a 7×7 0.05° resolution grid temperature.

[0063] Among them, it is necessary to understand that, corresponding to the preset grid size of the model configuration, such as a 7×7 grid, the previous model input and the features in the model processing process may be adapted to the corresponding size for better processing.

[0064] Furthermore, it can be understood that the initial numerical weather forecast data obtained from the output product of the original numerical weather forecast model involves many different aspects, and the present application focuses on the meteorological factors and terrain features contained therein, or in other words, specifically extracts / extracts the meteorological factors and terrain features required for the present application for the temperature correction processing carried out in the present application.

[0065] In this regard, as another exemplary embodiment, the meteorological factors and terrain features involved in the present application may specifically include 36 pressure layer variables (NWP Pressure Level), 5 ground forecast variables (NWP Surface Level), 3 ground static variables (Others) and 4 horizontal interpolation variables, forming a factor of 7×7×48, wherein:

[0066] 36 barosphere variables, including the gravity potential at 500, 850, 900, 925, 950 and 1000 hPa (Geopotential), the temperature at 500, 700, 800, 850, 900, 925, 950 and 1000 hPa (Temperature), the u component of wind at 500, 700, 850, 925 and 950 hPa (U component of wind), the v component of wind at 500, 700, 850, 925 and 950 hPa (V component of wind), and the relative humidity at 200, 300, 400, 500, 600, 700, 800, 850, 900, 925, 950 and 1000 hPa (RelativeHumidity);

[0067] 5 surface forecast variables, including mean sea level pressure, 2 metre temperature, 2 metre dewpoint temperature, 10 metre U wind component and 10 metre V wind component;

[0068] 3 ground static variables, including altitude, longitude and latitude;

[0069] 4 horizontal interpolation variables, including 2 metre temperature, altitude, longitude and latitude, are horizontally interpolated at a resolution of 0.05°.

[0070] Specifically, this application selects 36 pressure layer variables, among which the gravity potential (Geopotential) is used to characterize the specific height of each variable in the air; the temperature (Temperature) and the u / v component of the wind can characterize the basic conditions of the atmosphere, the evolution of cold and warm advection and weather conditions; the relative humidity (Relative Humidity) is related to the cloud cover condition, and the cloud cover directly affects the solar radiation on the ground and has an important impact on the ground temperature, so the relative humidity (Relative Humidity) variable selects the most pressure levels to express the solar radiation condition in a refined manner.

[0071] At the same time, this application also selects 5 ground forecast variables. This application believes that these ground meteorological elements are directly related to the temperature.

[0072] Next, this application also selected three ground static variables, namely altitude, longitude and latitude, to reflect the geographical differences between different regions / areas.

[0073] In addition, this application also believes that it is necessary to consider the direct correlation between the four factors of 2 meter temperature, altitude, longitude and latitude and the temperature label, and in order to strengthen their weight in the network, these four factors are horizontally interpolated at a resolution of 0.05° and then re-entered into the model, that is, they are horizontally interpolated at a resolution of 0.05° as the new four factors.

[0074] In this way, 48 factors were obtained as specific model input data for the model to use for temperature correction.

[0075] Among them, the temperature at a height of 2 meters (2 metre temperature) is the model processing target of the numerical weather forecast data processing model of this application, that is, it is the target that the numerical weather forecast data processing model needs to correct for the input initial numerical weather forecast data (48 factors), and it is also the target for subsequent downscaling.

[0076] At the same time, in order to better help the model converge quickly during training or in actual use, the present application can also perform corresponding preprocessing operations on the model input data to standardize and simplify the model input data and improve the quality of the model input data.

[0077] Taking the actual use process as an example, as another exemplary embodiment, before the application inputs the initial numerical weather forecast data (48 factors) into the pre-configured numerical weather forecast data processing model, the method of the application may also include:

[0078] The 7×7×48 factor is Z-score standardized using the following formula:

[0079] ,

[0080] in, For output, is the factor that needs to be standardized, is the factor mean, is the factor standard deviation, specifically, and These two parameters can be preset, or they can be determined by samples during the model training process, or they can be determined by all real data input to the model in the actual application of the model, or they can be determined by real data input to the model within a period of time from the current time point to the future in the actual application of the model. The specific setting method is relatively flexible.

[0081] In actual situations, the Z score can also be expressed / called Z-Score or standard score, which can be used to transform features of different scales into features of similar or same scales, achieving a data preprocessing effect.

[0082] It is easy to understand that this application does not make corresponding optimization and improvement to the Z-score standardization processing itself, but points out that the Z-score standardization processing based on Z-score is suitable for the data preprocessing of the first 48 factors in this application. Therefore, no further expansion of the Z-score standardization processing is given here.

[0083] Similarly, other types of specific preprocessing operations may also be involved in actual situations.

[0084] In addition, this application also believes that even if the weather circulation situation is the same, the corresponding model mapping relationship will be different due to the different locations of the sub-regions (subdivided grids). Similarly, the temperature has obvious daily variation characteristics and is directly related to the time, so the location sequence number and hour of the sub-region (subdivided grid) should also be used as factors to input the model.

[0085] If the natural sequence number is used directly, obvious logical errors will occur. For example, 23:00 and 0:00 are close to each other, but they are the farthest in numerical distance. They need to be treated as categorical variables. In this case, traditional machine learning methods usually use one-hot encoding to discretize categorical variables from 0 to 1. However, after one-hot encoding, the vector will be sparse, unable to consider the relationship between categories, and is not suitable for the convolution module of deep learning.

[0086] In this regard, this application adopts the embedding encoding process commonly used in the field of natural language deep learning. The embedding encoding process is actually a way to convert categorical variables into continuous vector representations. It first uses one-hot encoding to sparse the categorical variables, and then increases or reduces their dimensions through neural network training. The neural network training method can find the intrinsic connection between the sparse variables and generate new vectors without treating them as completely independent categorical variables.

[0087] In this way, this application uses embedding encoding processing to automatically encode the location number of the sub-area (subdivided grid) and the 24-hour number (serial number) as factors before inputting them into the model. The encoding size of each categorical variable is 49. After reshaping (Reshaping) to 7×7×1, it is superimposed (spliced) with the aforementioned 48 standardized factors as model input.

[0088] Correspondingly, as another exemplary embodiment, the method of the present application may further include:

[0089] Using embedding coding, multiple subdivided grids and 24 hours are automatically encoded to obtain two factors of size 7×7×1;

[0090] Input the initial numerical weather forecast data into the pre-configured numerical weather forecast data processing model, including:

[0091] According to the preset splicing method that represents the corresponding relationship, the factor of 7×7×48 size and two factors of 7×7×1 encoding size are spliced ​​to obtain a factor of 7×7×50 size;

[0092] A factor of size 7×7×50 is input into a pre-configured numerical weather prediction data processing model.

[0093] Among them, it should be noted that, in practical applications, the subdivided grid is one of the pre-configured contents of the model, that is, the subdivided grid is fixed in advance. In this case, the complete area spliced ​​together by all the subdivided grids can be directly the target area mentioned above (this means that the numerical weather forecast data processing model is in a paired relationship with the target area, and only the target area can be subjected to secondary processing of the numerical weather forecast data), or it can also include the target area (this means that the numerical weather forecast data processing model is in a paired relationship with a certain area that includes the target area, and only the numerical weather forecast data can be subjected to secondary processing of the certain area that includes the target area).

[0094] On the other hand, the subdivision grid may also be divided in real time, for example, according to a strategy pre-adapted to the model / pre-agreed with the model, or directly and autonomously, which is also possible in actual situations.

[0095] At the same time, under the setting of this embodiment, the numerical weather forecast data processing model also has corresponding configuration work. For the model, the characteristic matrix structure of the input data is (H, W, C), where H (height) and W (width) are 7, indicating the number of grids in the sub-region (subdivided grid), and C (convolution channel number) is 50, that is, the number of prediction factors. At this time, the total input size of the model is 7×7×50.

[0096] It should also be noted that the embodiment here also discloses that the present application is processed in units of 1 h (i.e., the time resolution is 1 h, corresponding to a coding setting of 24 hours). Compared with the aforementioned ECMWF IFS product, which has a time interval of 3 h, it has a downscaling effect. In other words, the numerical weather forecast data processing model of the present application has the processing capabilities of bias correction and spatiotemporal downscaling.

[0097] Specifically, with the addition of time coding features, the numerical weather forecast data processing model of this application has a better time downscaling effect, and the stability of the hourly temperature forecast is significantly enhanced, so that the model can minimize the error introduced by direct time interpolation of numerical weather forecast data, and can also show a stronger correction effect during the period when the ECMWF's IFS product forecast is poor.

[0098] In addition, as another exemplary embodiment, the numerical weather forecast data processing model may be configured with the following contents:

[0099] (1) The convolution operation performed by the convolutional layer adopts local connection and weight sharing methods;

[0100] It can be understood that the input data enters the 2D convolution layer, and the convolution operation has two important characteristics: local connections and shared weights. A grid point has a weaker correlation with grid points farther away. Therefore, this application believes that while considering the impact of various meteorological factors and terrain features in adjacent / surrounding areas on the deviation, it is not necessary for each grid point to be connected to the global grid point, only local connections are required.

[0101] In addition, it can be considered that the statistical characteristics between grid points are similar, which means that the learned feature parameters of one part of the grid can also be used on another part. The parameters can be reduced by sharing weights, reducing the complexity of the network. That is, each convolution can be considered as the air temperature, which is closely related to the meteorological elements and geographical features of the surrounding close-range grids (not necessarily directly adjacent), and this correlation can be applied globally, expressing the weights and correlations of each factor through the parameters in the trainable convolution kernel.

[0102] (2) Multiple convolutional layers are stacked to expand the receptive field;

[0103] It can be understood that this application believes that large-scale circulation characteristics will also have an impact on local temperature. Therefore, in order to introduce global features to achieve the unity of local and global, the receptive field can be expanded by superimposing multiple convolutional layers.

[0104] (3) Each convolutional layer does not perform pooling, and only changes the number of convolution kernels to ensure that the output grid size remains unchanged;

[0105] It can be understood that each convolution layer does not need to perform pooling to avoid the loss of refined features. Only the number of convolution kernels is changed to ensure that the output grid size remains unchanged.

[0106] (4) Each convolutional layer is followed by a BatchNorm layer and a softPlus nonlinear activation function layer. The BatchNorm layer is used to maintain data distribution during training to reduce variable offset, solve the gradient vanishing problem during back propagation, and suppress overfitting. The softPlus nonlinear activation function layer is used to avoid generating zero gradient when processing negative inputs.

[0107] It can be understood that each convolutional layer is followed by a BatchNorm layer and a softPlus non-linear activation function layer.

[0108] BatchNorm maintains data distribution during training to reduce variable shift, solves the problem of gradient vanishing during backpropagation, and suppresses overfitting. Its function is similar to the Z-score normalization process mentioned earlier that can be used in pre-data processing. The difference is that its normalization parameters are generated through training rather than statistics.

[0109] The softPlus nonlinear activation function layer introduced after the BatchNorm layer can be regarded as a smooth approximation of the traditional Relu activation function. It retains nonlinearity while introducing continuous differentiability. When processing negative inputs, it does not produce zero gradient like the Relu activation function, which can effectively avoid the phenomenon of neuron "death".

[0110] (5) The number of convolution kernels in the last convolutional layer is 10, and the corresponding output size is 7×7×10. It then passes through the flattening layer and enters the fully connected layer. The output of the fully connected layer is 49, which is restored to the size of 7×7 corresponding to the subdivided grid after reshaping.

[0111] The number of convolution kernels of the last convolution layer in the multi-convolution layer stacking architecture is 10, that is, the final output of the convolution layer is 7×7×10, and then enters the fully connected layer (DenseLayer) through the flatten layer. The final output of the fully connected layer is 49, which is restored to the size of the sub-region 7×7 after reshaping.

[0112] At this point, the original numerical weather forecast data with a 0.125° resolution grid temperature at an interval of 3 hours can be corrected to a 0.05° resolution grid temperature at an interval of 1 hour and more precise, effectively achieving the goals of deviation correction and spatiotemporal downscaling.

[0113] By stitching together all the grid temperatures, we can obviously obtain the corrected and downscaled temperature data of the target area and output it as a model product.

[0114] In order to verify the advantages of the application scheme, the application was compared with the IFS product based on four evaluation indicators: Mean Absolute Error (MAE), Root Mean Absolute Error (RMSE), Accuracy (ACC), and Structural Similarity (SSIM). Among them, the application was ahead of IFS in the MAE, RMSE, ACC and SSIM evaluation indicators, and the MAE was reduced by 35.7% overall in the grid test, showing strong usability, which proves that the application scheme has good application performance.

[0115] The above is an introduction to the processing method of numerical weather forecast data provided by the present application. In order to facilitate better implementation of the processing method of numerical weather forecast data provided by the present application, the present application also provides a processing device for numerical weather forecast data from the perspective of functional modules.

[0116] See also Figure 2 , Figure 2 This is a schematic diagram of a structure of a numerical weather forecast data processing device of the present application. In the present application, the numerical weather forecast data processing device 200 may specifically include the following structure:

[0117] The acquisition unit 201 is used to acquire initial numerical weather forecast data obtained by numerical weather forecast processing in the target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features;

[0118] The processing unit 202 is used to input the initial numerical weather forecast data into a pre-configured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used to perform correction processing on the numerical weather forecast data input into the model, and the numerical weather forecast data processing model is obtained by training the initial deep learning model with labeled sample numerical weather forecast data. During operation, the numerical weather forecast data processing model performs temperature correction processing based on a plurality of subdivided grids, and during the temperature correction processing, the deviation generated by the numerical weather forecast data of the grids surrounding the subdivided grids is considered for correction, and the plurality of subdivided grids are obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model, based on a preset resolution, according to a preset number of grids or a preset grid size;

[0119] The extraction unit 203 is used to extract the target numerical weather forecast data with subdivided grid granularity output by the numerical weather forecast data processing model.

[0120] In an exemplary embodiment, the preset resolution is specifically 0.05° resolution, and the preset grid size is specifically 7×7 geographic grid sizes at 0.05° resolution.

[0121] In another exemplary embodiment, the meteorological factors and terrain features specifically include 36 pressure layer variables, 5 ground forecast variables, 3 ground static variables and 4 horizontal interpolation variables, forming a factor of 7×7×48, wherein:

[0122] 36 barosphere variables, including gravity potential at 500, 850, 900, 925, 950 and 1000 hPa, air temperature at 500, 700, 800, 850, 900, 925, 950 and 1000 hPa, u-component of wind at 500, 700, 850, 925 and 950 hPa, v-component of wind at 500, 700, 850, 925 and 950 hPa, relative humidity at 200, 300, 400, 500, 600, 700, 800, 850, 900, 925, 950 and 1000 hPa;

[0123] 5 surface forecast variables, including mean sea level pressure, temperature at 2 m, dew point temperature at 2 m, u component of wind at 10 m, and v component of wind at 10 m;

[0124] 3 ground static variables, including altitude, longitude and latitude;

[0125] 4 horizontal interpolation variables, including temperature at 2 meters, altitude, longitude and latitude, interpolated horizontally at a resolution of 0.05°;

[0126] Among them, the temperature at a height of 2 meters is specifically the model processing target.

[0127] In yet another exemplary embodiment, the apparatus further includes a preprocessing unit 204, configured to:

[0128] The 7×7×48 factor is Z-score standardized using the following formula:

[0129] ,

[0130] in, For output, is the factor that needs to be standardized, is the factor mean, is the factor standard deviation.

[0131] In yet another exemplary embodiment, the apparatus further includes an encoding unit 205, configured to:

[0132] Using embedding coding, multiple subdivided grids and 24 hours are automatically encoded to obtain two factors of size 7×7×1;

[0133] The processing unit 202 is specifically configured to:

[0134] According to the preset splicing method that represents the corresponding relationship, the factor of 7×7×48 size and two factors of 7×7×1 encoding size are spliced ​​to obtain a factor of 7×7×50 size;

[0135] A factor of size 7×7×50 is input into a pre-configured numerical weather prediction data processing model.

[0136] In another exemplary embodiment, the numerical weather forecast data processing model is configured with the following contents:

[0137] The convolution operation performed by the convolution layer uses local connection and weight sharing;

[0138] Multiple convolutional layers are stacked to expand the receptive field;

[0139] Each convolution layer does not perform pooling, only the number of convolution kernels is changed to ensure that the output grid size remains unchanged;

[0140] Each convolutional layer is followed by a BatchNorm layer and a softPlus nonlinear activation function layer. The BatchNorm layer is used to maintain data distribution during training to reduce variable offset, solve the gradient vanishing problem during back propagation, and suppress overfitting. The softPlus nonlinear activation function layer is used to avoid zero gradient when processing negative inputs.

[0141] The number of convolution kernels in the last convolution layer is 10, and the corresponding output size is 7×7×10. Then it enters the fully connected layer through the flattening layer. The output of the fully connected layer is 49. After reshaping, it is restored to the size of 7×7 corresponding to the subdivided grid.

[0142] In yet another exemplary embodiment, the apparatus further includes a training unit 206, configured to:

[0143] By training the initial deep learning model with labeled sample numerical weather forecast data, we can obtain the numerical weather forecast data processing model.

[0144] This application also provides a processing device from the perspective of hardware structure, see Figure 3 , Figure 3 301, a memory 302, and an input / output device 303. The processor 301 is used to execute the computer program stored in the memory 302 to implement the following Figure 1 The steps of the method for processing numerical weather forecast data in the corresponding embodiment; or, the processor 301 is used to execute the computer program stored in the memory 302 to implement the following Figure 2 Corresponding to the functions of each unit in the embodiment, the memory 302 is used to store the processor 301 executing the above Figure 1 The computer program required by the method for processing numerical weather forecast data in the corresponding embodiment.

[0145] Exemplarily, the computer program may be divided into one or more modules / units, one or more modules / units are stored in the memory 302, and executed by the processor 301 to complete the present application. One or more modules / units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the computer device.

[0146] The processing device may include, but is not limited to, a processor 301, a memory 302, and an input / output device 303. Those skilled in the art will appreciate that the illustration is merely an example of a processing device and does not constitute a limitation on the processing device, and may include more or fewer components than shown in the illustration, or a combination of certain components, or different components. For example, the processing device may also include a network access device, a bus, etc., and the processor 301, the memory 302, the input / output device 303, etc. are connected via a bus.

[0147] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the processing device, and uses various interfaces and lines to connect various parts of the entire device.

[0148] The memory 302 can be used to store computer programs and / or modules. The processor 301 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 302 and calling the data stored in the memory 302. The memory 302 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the processing device, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0149] When the processor 301 is used to execute the computer program stored in the memory 302, the following functions can be implemented:

[0150] Acquiring initial numerical weather forecast data obtained through numerical weather forecast processing in the target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features;

[0151] Input the initial numerical weather forecast data into a preconfigured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used to carry out correction processing on the numerical weather forecast data input into the model, and the numerical weather forecast data processing model is obtained by training the initial deep learning model with labeled sample numerical weather forecast data. During the working process, the numerical weather forecast data processing model carries out temperature correction processing based on multiple subdivided grids, and during the temperature correction processing, the deviation generated by the numerical weather forecast data of the grids surrounding the subdivided grids is considered for correction, and the multiple subdivided grids are obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model according to the preset number of grids or the preset grid size based on the preset resolution;

[0152] Extract the target numerical weather forecast data with subdivided grid granularity output by the numerical weather forecast data processing model.

[0153] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the numerical weather forecast data processing device, processing equipment and corresponding units described above can refer to the following. Figure 1 The description of the method for processing numerical weather forecast data in the corresponding embodiment will not be repeated here in detail.

[0154] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0155] To this end, the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the present application as follows: Figure 1 The steps of the method for processing numerical weather forecast data in the corresponding embodiment, the specific operations can refer to the following Figure 1 The description of the method for processing numerical weather forecast data in the corresponding embodiment will not be repeated here.

[0156] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0157] Due to the instructions stored in the computer-readable storage medium, the present application can be executed. Figure 1 The steps of the method for processing numerical weather forecast data in the corresponding embodiment, therefore, the present application can be implemented as follows Figure 1The beneficial effects that can be achieved by the method for processing numerical weather forecast data in the corresponding embodiment are detailed in the previous description and will not be repeated here.

[0158] The above is a detailed introduction to the processing method, device, processing equipment and computer-readable storage medium for numerical weather forecast data provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for processing numerical weather forecast data, characterized in that: The method comprises: Acquiring initial numerical weather forecast data obtained through numerical weather forecast processing in the target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features; Inputting the initial numerical weather forecast data into a preconfigured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used to carry out correction processing on the numerical weather forecast data input into the model, the numerical weather forecast data processing model is obtained by training an initial deep learning model with labeled sample numerical weather forecast data, and in the working process, the numerical weather forecast data processing model carries out temperature correction processing based on a plurality of subdivided grids, and in the temperature correction processing process, the deviation generated by the numerical weather forecast data of the grids surrounding the subdivided grids is considered for correction, and the plurality of subdivided grids are obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model, based on a preset resolution, according to a preset number of grids or a preset grid size; Extracting the target numerical weather forecast data of the subdivided grid granularity output by the numerical weather forecast data processing model; The preset resolution is specifically 0.05° resolution, and the preset grid size is specifically 7×7 geographic grid sizes at 0.05° resolution; The meteorological factors and the terrain features specifically include 36 pressure layer variables, 5 ground forecast variables, 3 ground static variables and 4 horizontal interpolation variables, forming a factor of 7×7×48, among which: The 36 pressure layer variables include gravity potential at 500, 850, 900, 925, 950 and 1000 hPa, air temperature at 500, 700, 800, 850, 900, 925, 950 and 1000 hPa, u component of wind at 500, 700, 850, 925 and 950 hPa, v component of wind at 500, 700, 850, 925 and 950 hPa, and relative humidity at 200, 300, 400, 500, 600, 700, 800, 850, 900, 925, 950 and 1000 hPa; The five ground forecast variables specifically include mean sea level pressure, temperature at 2 meters, dew point temperature at 2 meters, u component of wind at 10 meters, and v component of wind at 10 meters; The three ground static variables specifically include altitude, longitude and latitude; The four horizontal interpolation variables specifically include the temperature at a height of 2 meters, the altitude, the longitude and the latitude, which are variables after being horizontally interpolated at a resolution of 0.05°; The temperature at a height of 2 meters is specifically the model processing target; The method further comprises: By using embedding encoding processing, the plurality of subdivided grids and 24 hours are automatically encoded to obtain two factors of 7×7×1 size; Inputting the initial numerical weather forecast data into a pre-configured numerical weather forecast data processing model specifically includes: splicing the factor of the size of 7×7×48 and the two factors of the encoding size of 7×7×1 according to a preset splicing method that characterizes the corresponding relationship, to obtain a factor of the size of 7×7×50; Inputting the factor of size 7×7×50 into the pre-configured numerical weather forecast data processing model; The numerical weather forecast data processing model is configured with the following contents: The convolution operation performed by the convolution layer uses local connection and weight sharing; A plurality of convolutional layers are stacked to expand the receptive field; Each of the convolutional layers does not perform pooling, and only changes the number of convolution kernels to ensure that the output grid size remains unchanged; Each of the convolutional layers is followed by a BatchNorm layer and a softPlus nonlinear activation function layer, wherein the BatchNorm layer is used to maintain data distribution during training to reduce variable offset, solve the gradient vanishing problem during back propagation, and suppress overfitting, and the softPlus nonlinear activation function layer is used to avoid generating zero gradient when processing negative inputs; The number of convolution kernels of the last convolution layer is 10, and the corresponding output size is 7×7×10, and then it enters the fully connected layer through the flattening layer. The output of the fully connected layer is 49, and after reshaping, it is restored to the size of 7×7 corresponding to the subdivided grid.

2. The method according to claim 1, characterized in that Before inputting the initial numerical weather forecast data into a preconfigured numerical weather forecast data processing model, the method further comprises: The 7×7×48 factor is subjected to Z-score normalization by the following formula: ， in, For output, is the factor that needs to be standardized, is the factor mean, is the factor standard deviation.

3. The method according to claim 1, characterized in that The method further comprises: The initial deep learning model is trained by using the labeled sample numerical weather forecast data to obtain the numerical weather forecast data processing model.

4. A numerical weather forecast data processing device, characterized in that: The device comprises: An acquisition unit, used to acquire initial numerical weather forecast data obtained by numerical weather forecast processing in a target area, wherein the initial numerical weather forecast data includes meteorological factors and terrain features; A processing unit, used for inputting the initial numerical weather forecast data into a preconfigured numerical weather forecast data processing model, wherein the numerical weather forecast data processing model is used for performing correction processing on the numerical weather forecast data input into the model, the numerical weather forecast data processing model is obtained by training an initial deep learning model with labeled sample numerical weather forecast data, and in the working process, the numerical weather forecast data processing model performs temperature correction processing based on a plurality of subdivided grids, and in the temperature correction processing process, correction is performed by considering deviations generated by numerical weather forecast data of grids surrounding the subdivided grids, the plurality of subdivided grids being obtained by dividing the corresponding area of ​​the numerical weather forecast data input into the model, based on a preset resolution, according to a preset number of grids or a preset grid size; An extraction unit, used for extracting the target numerical weather forecast data of the subdivided grid granularity output by the numerical weather forecast data processing model; The preset resolution is specifically 0.05° resolution, and the preset grid size is specifically 7×7 geographic grid sizes at 0.05° resolution; The meteorological factors and the terrain features specifically include 36 pressure layer variables, 5 ground forecast variables, 3 ground static variables and 4 horizontal interpolation variables, forming a factor of 7×7×48, among which: The 36 pressure layer variables include gravity potential at 500, 850, 900, 925, 950 and 1000 hPa, air temperature at 500, 700, 800, 850, 900, 925, 950 and 1000 hPa, u component of wind at 500, 700, 850, 925 and 950 hPa, v component of wind at 500, 700, 850, 925 and 950 hPa, and relative humidity at 200, 300, 400, 500, 600, 700, 800, 850, 900, 925, 950 and 1000 hPa; The five ground forecast variables specifically include mean sea level pressure, temperature at 2 meters, dew point temperature at 2 meters, u component of wind at 10 meters, and v component of wind at 10 meters; The three ground static variables specifically include altitude, longitude and latitude; The four horizontal interpolation variables specifically include the temperature at a height of 2 meters, the altitude, the longitude and the latitude, which are variables after being horizontally interpolated at a resolution of 0.05°; The temperature at a height of 2 meters is specifically the model processing target; The device also includes an encoding unit, which is used to: By using embedding encoding processing, the plurality of subdivided grids and 24 hours are automatically encoded to obtain two factors of 7×7×1 size; The processing unit is specifically used for: splicing the factor of the size of 7×7×48 and the two factors of the encoding size of 7×7×1 according to a preset splicing method that characterizes the corresponding relationship, to obtain a factor of the size of 7×7×50; Inputting the factor of size 7×7×50 into the pre-configured numerical weather forecast data processing model; The numerical weather forecast data processing model is configured with the following contents: The convolution operation performed by the convolution layer uses local connection and weight sharing; A plurality of convolutional layers are stacked to expand the receptive field; Each of the convolutional layers does not perform pooling, and only changes the number of convolution kernels to ensure that the output grid size remains unchanged; Each of the convolutional layers is followed by a BatchNorm layer and a softPlus nonlinear activation function layer, wherein the BatchNorm layer is used to maintain data distribution during training to reduce variable offset, solve the gradient vanishing problem during back propagation, and suppress overfitting, and the softPlus nonlinear activation function layer is used to avoid generating zero gradient when processing negative inputs; The number of convolution kernels of the last convolution layer is 10, and the corresponding output size is 7×7×10, and then it enters the fully connected layer through the flattening layer. The output of the fully connected layer is 49, and after reshaping, it is restored to the size of 7×7 corresponding to the subdivided grid.

5. A processing device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the processor executes the method according to any one of claims 1 to 3 when calling the computer program in the memory.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the method according to any one of claims 1 to 3.

Images