Optimization Model Construction Method for Improving the Resolution and Forecast Accuracy of Artificial Intelligence Meteorological Large Models

Abstract

The present invention discloses an optimization model construction method and device for improving the resolution and prediction accuracy of an artificial intelligence meteorological large model. The method includes: constructing an initial field of a coarse grid resolution model using a high-resolution training data set, driving a target artificial intelligence meteorological large model with coarse grid resolution to perform continuous prediction for a preset period of time to respectively obtain coarse grid resolution data to be optimized for multiple prediction steps; inputting the data to be optimized corresponding to each prediction step into a pre-constructed artificial intelligence data processing model respectively, and combining with training data to obtain an error value corresponding to each prediction step; summing the error values obtained for each prediction step, and taking minimizing the objective function as the training objective, training the artificial intelligence data processing model based on the sum of the error values to obtain an optimized artificial intelligence data processing model. This method and device can improve the spatial resolution of the meteorological large model, and further improve the accuracy of meteorological prediction.

Description

Technical Field

[0001] The present invention relates to the technical fields of meteorology and artificial intelligence, and particularly relates to an optimization model construction method and device for improving the resolution and prediction accuracy of an artificial intelligence meteorological large model. Background Art

[0002] In recent years, important technological breakthroughs have been made in the application of artificial intelligence in the field of weather forecasting. The artificial intelligence weather forecasting large model (AIWFMs) has become an emerging direction in the development of atmospheric science. Compared with the traditional numerical weather forecasting model, AIWFMs have the outstanding advantages of small resource requirements and fast computing speed, which are convenient for wide and convenient application and ensemble forecasting, and complement the traditional deterministic forecasting.

[0003] Although AIWFMs have obvious advantages, the following problems still need to be solved in their actual business applications: Most AIWFMs use the ERA5 atmospheric reanalysis products of the European Medium-Range Weather Forecast Centre for training, resulting in obvious homogenization of the performance of AIWFMs models. In particular, the spatial resolution is severely limited by the original resolution of the ERA5 atmospheric reanalysis data. Given that the current prediction performance of AIWFMs is comparable to that of traditional numerical weather forecasting, to further improve the accuracy of specific prediction elements through the design and training of new models requires a large amount of additional costs, and the expected benefits are limited. In addition, existing AIWFMs are all global large models and pay insufficient attention to regional forecasting. The sporadic artificial intelligence regional weather forecasting large models (AIRWFM) are also similar to the traditional regional numerical weather models and require the global numerical weather forecasting model or AIGWFMs to provide boundary conditions. The prediction accuracy and computational efficiency are still limited by the prediction accuracy and computational efficiency of the global model, that is, there is a lack of an organic connection between the global and regional forecasting artificial intelligence large models.

[0004] In order to improve the spatial resolution of the artificial intelligence weather forecasting large model, there are mainly the following two existing methods:

[0005] 1) First, produce high-resolution regional reanalysis data and directly use artificial intelligence (especially deep learning) methods to train AIRWFM. This method requires a large amount of costs and the expected benefits are limited. It still needs the global model to provide boundary conditions, resulting in a lack of a close connection between the global and regional forecasting artificial intelligence large models;

[0006] 2) Secondary development of existing AIGWFMs using high-resolution (re)analysis data (or observational data), that is, using global high-resolution analysis data (such as global HRES data: https: / / rda.ucar.edu / datasets / ds113.1 / dataaccess / ) to train and superimpose an artificial intelligence encryption module on it. For example, FengWu-GHR, Aurora, and YanTian all input by converting HRES data into a data format acceptable to low-resolution models (3D Perceiver encoder of Aurora, YanTian is similar to FengWu-GHR). Of course, the processing modules of each model are different (for example, Decompositional and Combinational Transfer Learning of FengWu-GHR, 3DSwin Transformer U-Net of Aurora, and "variance" module of YanTian). However, these several methods of secondary training models do not fully utilize the spatio-temporal prediction evolution contained in the original AIGWFMs model.

[0007] Currently, the solutions to improve the spatial resolution of artificial intelligence weather forecasting large models all have their own drawbacks: directly generating regional high-resolution reanalysis data is extremely costly, and it is impossible to establish an effective connection with the global model, difficult to nest and encrypt, and must rely on the global (traditional or artificial intelligence) model to continuously provide boundary fields; while the secondary development solutions for existing low-resolution AIGWFMs models do not effectively reflect the spatio-temporal prediction evolution contained in the original AIGWFMs model.

[0008] In view of this, an optimized model construction method and device for improving the resolution and prediction accuracy of meteorological large models are provided, in order to at least partially solve the following technical problems existing in the prior art when optimizing the resolution and prediction accuracy of meteorological large models:

[0009] 1) The computational requirements for directly generating global / regional high-resolution reanalysis data for high-resolution global / regional / artificial intelligence training are unbearable;

[0010] 2) There is a lack of effective connection between the artificial intelligence regional meteorological large model AIRWFMs and the artificial intelligence global meteorological large model AIGWFMs, and it is difficult to be organically nested. The former still needs the latter to improve the boundary field;

[0011] 3) The encryption module developed based on the existing low-resolution AIGWFMs model does not fully utilize the spatio-temporal prediction evolution contained in the AIGWFMs model, and currently can only perform global encryption, failing to achieve local (nested) encryption of the global (or regional) large model. Summary of the Invention

[0012] To this end, embodiments of the present invention provide an optimization model construction method and device for improving the resolution and forecasting accuracy of an artificial intelligence meteorological large model to solve at least one of the above technical problems.

[0013] To achieve the above object, embodiments of the present invention provide the following technical solutions:

[0014] The present invention provides an optimization model construction method for improving the resolution and forecasting accuracy of an artificial intelligence meteorological large model, the method comprising:

[0015] Constructing an initial field of a coarse grid resolution model using a high-resolution training data set, and driving a target artificial intelligence meteorological large model with a coarse grid resolution to perform continuous forecasting for a preset period to respectively obtain coarse grid resolution data to be optimized at multiple forecasting steps;

[0016] Inputting the data to be optimized corresponding to each forecasting step into a pre-constructed artificial intelligence data processing model, and combining with the training data set to obtain error values corresponding to each predetermined forecasting step;

[0017] Summing the error values obtained at each forecasting step, and taking minimizing the objective function as the training target, training the artificial intelligence data processing model based on the sum of the error values to obtain an optimized artificial intelligence data processing model;

[0018] Dividing each forecasting step into multiple preset periods, and the artificial intelligence data processing model uses the coarse resolution indication data at the initial moment, the fine resolution indication data at the initial moment, and / or the coarse resolution indication data at the forecasting moment within the preset period as input data, and uses the fine resolution indication data at the forecasting moment of the preset period as output data for training.

[0019] In some embodiments, constructing an initial field of the model with coarse grid resolution using a high-resolution training data set specifically includes:

[0020] Interpolating the high-resolution training data to obtain the coarse grid resolution model field.

[0021] In some embodiments, driving the target artificial intelligence meteorological large model to perform continuous forecasting for a preset period specifically includes:

[0022] Driving the target meteorological large model to perform continuous forecasting from the initial moment t j-1 to the forecasting moment t j where t j-1 = t0 + (j - 1)Δt, t j = t0 + jΔt), j = 1,..., N τ , △t is the time step, τ is the time interval of high-resolution training data, N τ is the number of consecutive forecast moments, and t0 is the initial moment of each time interval.

[0023] In some embodiments, the coarse resolution indicator data at the initial moment is the coarse resolution variable at the initial moment, the fine resolution indicator data at the initial moment is the fine resolution variable at the initial moment, the coarse resolution indicator data at the forecast moment is the coarse resolution variable at the forecast moment, and the fine resolution indicator data at the forecast moment is the fine resolution variable at the forecast moment; accordingly, the artificial intelligence data processing model The expression is:

[0024]

[0025] Among them, t j-1 represents the initial time, represents the time residual between the forecast and the initial coarse-resolution forecast variable, represents the spatial residual between the fine and coarse grid resolution variables at the initial moment, t j Indicates the forecast time.

[0026] In some embodiments, the objective function is:

[0027]

[0028] Where k(=1,…,T) refers to the frequency of high-resolution training data, x h,k represents the kth training data and

[0029] In some embodiments, the network architecture of the artificial intelligence data processing model includes but is not limited to a residual network, a U-type network, a graph neural network, or a neural network based on an attention mechanism.

[0030] The present invention also provides an optimization model construction device for improving the resolution and forecast accuracy of an artificial intelligence meteorological large model, the device comprising:

[0031] A data acquisition unit is used to construct an initial field of a coarse grid resolution model using a high-resolution training data set, and drive a coarse grid resolution target artificial intelligence meteorological large model to perform continuous forecasting for a preset period of time, so as to obtain data to be optimized for multiple forecast steps respectively;

[0032] An outer loop unit is used to input the data to be optimized corresponding to each forecast step into a pre-constructed artificial intelligence data processing model respectively, and combine with the training data to obtain the error values corresponding to each forecast step; sum the error values obtained for each forecast step, and take minimizing the objective function as the training objective, and train the artificial intelligence data processing model based on the sum of the error values to obtain an optimized artificial intelligence data processing model;

[0033] An inner loop unit is used to divide each forecast step into multiple preset time periods. The artificial intelligence data processing model uses the coarse-resolution indication data at the initial moment, the fine-resolution indication data at the initial moment, and / or the coarse-resolution indication data at the forecast moment within the preset time period as input data, and uses the fine-resolution indication data at the forecast moment of the preset time period as output data for training.

[0034] The present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above method are implemented.

[0035] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0036] In one or several of the above specific embodiments, the optimization model construction method and device for improving the resolution and forecast accuracy of a meteorological large model provided by the present invention have the following technical effects:

[0037] Using the global / regional high-precision (re)analysis (or observation, fusion) data in the existing target dataset as training data, realizing the global and local two encryption functions of the artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)), establishing an effective connection between the global and regional artificial intelligence large models, and being able to achieve multiple nested encryptions through repeated encryption operations, and being able to be compatible with traditional numerical weather prediction models;

[0038] In the process of generating the data processing model fully incorporate the spatio-temporal forecast evolution contained in the original coarse-resolution artificial intelligence global (or regional) large model AIG( / R)WFM model (DepM l→l (·)), that is, input the initial and forecast moment coarse-resolution simulation and forecast results of the AIG( / R)WFM model (DepM l→l (·));

[0039] Use artificial intelligence algorithms to train the data processing model to generate a better data optimization model

[0040]

[0041] Generally speaking, the present invention relates to fields such as meteorology, ecology, environment, new energy, and artificial intelligence. In particular, it relates to a method for improving the spatial resolution of an existing artificial intelligence meteorological large model with a coarse grid resolution by superimposing an artificial intelligence model (such as a deep neural network, etc.), and further improving the simulation and prediction accuracy of meteorological, ecological, and environmental variables. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary. For those of ordinary skill in the art, without creative efforts, other implementation drawings can also be obtained based on the provided drawings.

[0043] The structures, proportions, sizes, etc. illustrated in this specification are only used to cooperate with the content disclosed in the specification for those familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention. Therefore, they do not have technical substance. Any modification of the structure, change in the proportional relationship, or adjustment of the size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope covered by the technical content disclosed in the present invention.

[0044] Figure 1 One of the flowcharts of the optimization model construction method for improving the resolution and prediction accuracy of the meteorological large model provided by the present invention;

[0045] Figure 2 Another flowchart of the optimization model construction method for improving the resolution and prediction accuracy of the meteorological large model provided by the present invention;

[0046] Figure 3 The structural schematic diagram of the optimization model construction device for improving the resolution and prediction accuracy of the meteorological large model provided by the present invention;

[0047] Figure 4 The structural block diagram of a computer device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0049] As Figure 1 shown, the optimization model construction method provided by the present invention for improving the pre-resolution and prediction accuracy of the meteorological large model is a new general solution for improving the resolution and prediction accuracy of the AI meteorological large model. The optimization model construction method is applicable to the following two situations:

[0050] First, global encryption: At this time, an artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)) and a high-resolution (re)analysis (or observation, fusion) data set with the same simulation area as it are required. That is, the simulation area of AIG( / R)WFM(DepM l→l (·)) is the same as the coverage area of the high-resolution (re)analysis (or observation, fusion) data. Only the model resolution of AIG( / R)WFM(DepM l→l (·)) is lower than the resolution of the (re)analysis (or observation, fusion) data. This situation realizes the global encryption function of the artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)).

[0051] Second, local encryption: At this time, an artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)) and a high-resolution regional (re)analysis (or observation, fusion) data set are also required. The high-resolution (re)analysis (or observation, fusion) data only partially covers the simulation area of the artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)). This situation realizes the local encryption function of the artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)), that is, a nested encryption.

[0052] In a specific implementation manner, as Figure 2 shown, the optimization model construction method provided by the present invention for improving the resolution and prediction accuracy of the meteorological large model includes the following steps:

[0053] S210: Construct the initial field of the coarse-grid resolution model using a high-resolution training dataset (this dataset can be generated by downscaling existing reanalysis data products, observational data, and driving a high-resolution numerical prediction model with reanalysis data products), and drive the target artificial intelligence meteorological large model with coarse-grid resolution to perform continuous forecasting for a preset period, so as to obtain the data to be optimized with coarse-grid resolution at multiple forecast steps respectively;

[0054] S220: Input the data to be optimized corresponding to each forecast step into the pre-constructed artificial intelligence data processing model respectively, and combine with the training data to obtain the error value corresponding to each forecast step;

[0055] S230: Sum up the error values obtained at each forecast step, and take minimizing the objective function as the training goal, and train the artificial intelligence data processing model based on the sum of the error values to obtain the optimized artificial intelligence data processing model.

[0056] Among them, within each forecast step, it is divided into multiple preset periods. The artificial intelligence data processing model uses the coarse-resolution indicative data at the initial moment within the preset period, the fine-resolution indicative data at the initial moment, and / or the coarse-resolution indicative data at the forecast moment within the preset period as input data, and uses the fine-resolution indicative data at the forecast moment of the preset period as output data for training.

[0057] In step S210, constructing the initial field of the low-grid resolution of the model using the high-resolution training dataset specifically includes:

[0058] Interpolate the high-resolution training data to obtain the coarse-grid resolution model field.

[0059] Among them, driving the target meteorological large model to perform continuous forecasting for a preset period specifically includes:

[0060] Drive the target meteorological large model with coarse-grid resolution from the initial moment t j-1 to the forecast moment t j for continuous forecasting, where t j-1 =t0+(j - 1)Δt, t j =t0 + jΔt), j = 1, …, N τ , Δt is the time step, τ is the time interval of the high-resolution training data, N τ is the number of moments of continuous forecasting, and t0 is the initial moment of each time interval. It should be understood that this is a continuous forecasting process of cyclic iteration, and the data obtained at the forecast moment is used as the data of the initial moment for the next moment.

[0061] In a specific usage scenario, in step S210, first, high-resolution (re)analysis (or observation, fusion) data is simply interpolated into a coarse-grid resolution model field to drive the artificial intelligence large model AIG( / R)WFM(DepM l→l (·)) to perform continuous forecasting from t j-1 (=t0+(j - 1)Δt) to t j (=t0 + jΔt), where Δt is the time step of AIG( / R)WFM(DepM l→l (·)), while τ is the time interval of high-resolution training data.

[0062] It should be understood that the data processing model to be optimized exists at each forecasting step. It takes the initial coarse and fine-grid resolution fields and the coarse-grid forecasting field at the one-step-ahead forecasting moment as inputs, while the fine-grid data at the forecasting moment is the forecasting field, and for the training data, it only exists at relatively large time intervals.

[0063] When designing the data processing model, the inputs and outputs of the model can be various indicator data related to the resolution, such as the time residuals between coarse-resolution variables or the spatial residuals between coarse and fine-resolution variables.

[0064] Specifically, the coarse-resolution indicator data at the initial moment is the coarse-resolution variable at the initial moment, the fine-resolution indicator data at the initial moment is the fine-resolution variable at the initial moment, the coarse-resolution indicator data at the forecasting moment is the coarse-resolution variable at the forecasting moment, and the fine-resolution indicator data at the forecasting moment is the fine-resolution variable at the forecasting moment; at this time, the artificial intelligence data processing model has the expression:[[]]

[0065]

[0066] where t j-1 represents the initial moment, represents the coarse-resolution variable at the initial moment, represents the fine-resolution variable at the initial moment, t j represents the forecasting moment, represents the coarse-resolution variable at the forecasting moment, represents the fine-resolution variable at the forecasting moment. That is to say, the inputs of this data processing model are the coarse and fine-resolution data at time t j-1 and the coarse-resolution data at time t and t j where indicates the AI encryption model ​Incorporated the spatio-temporal prediction evolution contained in the AIG( / R)WFM model (DepM l→l (·)).

[0067] In some other embodiments, the coarse-resolution indication data at the initial moment is the time residual between the coarse resolution at the prediction moment and the coarse-resolution variable at the initial moment, and the fine-resolution indication data at the initial moment is the spatial residual between the fine resolution at the initial moment and the coarse-resolution variable; correspondingly, the artificial intelligence data processing model has the expression of:

[0068]

[0069] where t j-1 represents the initial moment, represents the time residual between the prediction and the coarse-resolution prediction variable at the initial moment, represents the spatial residual between the fine and coarse grid resolution variables at the initial moment, and t j represents the prediction moment.

[0070] That is to say, there can be various variants for the input and output of the above data processing model to be trained. For example, the input can be the residual predicted by the AIG( / R)WFM model (DepM l→l (·)) At time t j-1 the high-resolution prediction and the residual l→l of the AIG( / R)WFM(DepM model's coarse-resolution prediction At this time, it is also necessary to simply interpolate the coarse-resolution result to a high spatial resolution and then calculate the residual, while the output can be the high-resolution prediction at time t j and the residual between the AIG( / R)WFM model (DepM l→l (·)) prediction At this time, it is also necessary to simply interpolate the coarse-resolution result to a high spatial resolution and then calculate the residual. It should be understood that this situation particularly reflects the idea of multi-scale prediction. The AIG( / R)WFM model (DepM l→l (·)) with coarse resolution predicts the long-wave information in the atmosphere, and its residual or increment(·)) is exactly the prediction increment of this long-wave information; while the residual reflects the high-resolution prediction at the initial moment relative to the AIG( / R)WFM model (DepMl→l (·)) Short-wave increment of the coarse-resolution forecast.

[0071] In some embodiments, high-resolution training data in a high-resolution dataset (such as existing high-resolution (re)analysis (or observation, fusion) data) is used, and an artificial intelligence method is adopted for training to obtain a data processing model. The training objective is the following minimization function:

[0072]

[0073] where the subscript k (= 1,..., T) refers to the frequency of occurrence of the high-resolution training data, and x h,k represents the k-th training data and the data processing model to be determined in the above formula can be various suitable artificial intelligence models, including but not limited to residual networks, U-shaped networks, graph neural networks, and neural networks based on the attention mechanism. The model parameters of can be solved by the usual gradient descent method.

[0074] The network architecture of the artificial intelligence data processing model includes but is not limited to residual networks, U-shaped networks, graph neural networks, or neural networks based on the attention mechanism.

[0075] In one or several of the above specific embodiments, the optimization model construction method provided by the present invention for improving the resolution and forecast accuracy of the artificial intelligence meteorological large model has the following technical effects:

[0076] Using the global / regional high-precision (re)analysis (or observation, fusion) data in the existing target dataset as training data, the global and local encryption functions of the artificial intelligence global (or regional) large model AIG( / R)WFM(DepM l→l (·)) are realized, an effective connection between the global and regional artificial intelligence large models is established, and multiple nested encryptions can be achieved through repeated encryption operations, and it can be compatible with traditional numerical prediction models;

[0077] In the process of generating the data processing model the spatio-temporal forecast evolution contained in the original coarse-resolution artificial intelligence global (or regional) large model AIG( / R)WFM model (DepM l→l (·)) is fully incorporated, that is, the input includes the initial and forecast time coarse-resolution simulation and forecast results of the AIG( / R)WFM model (DepM l→l (·));

[0078] An artificial intelligence algorithm is used to train the data processing model to generate a better artificial intelligence data processing model.

[0079] Generally speaking, the present invention relates to the fields of meteorology, ecology, environment, new energy, and artificial intelligence, and particularly relates to a method for improving the spatial resolution of an existing artificial intelligence meteorological large model with a coarse grid resolution by superimposing an artificial intelligence model (such as a deep neural network, etc.), thereby improving the simulation and prediction accuracy of meteorological, ecological, and environmental variables.

[0080] In addition to the above method, the present invention also provides an optimized model construction device for improving the resolution and prediction accuracy of an artificial intelligence meteorological large model, as Figure 3 shown, the device includes:

[0081] A data acquisition unit 310, configured to construct a coarse grid resolution pattern field by using a high-resolution training data set, and drive a target artificial intelligence meteorological large model to perform continuous prediction for a preset period, so as to obtain to-be-optimized data for multiple prediction steps respectively;

[0082] An outer loop unit 320, configured to input the to-be-optimized data corresponding to each prediction step into a pre-constructed artificial intelligence data processing model respectively, and combine with training data to obtain an error value corresponding to each prediction step; sum the error values obtained for each prediction step, and use minimizing the objective function as the training objective, and train the artificial intelligence data processing model based on the sum of the error values to obtain a data optimization model;

[0083] An inner loop unit 330, configured to divide each prediction step into multiple preset periods, and the artificial intelligence data processing model uses the coarse resolution indication data at the initial moment within the preset period, the fine resolution indication data at the initial moment, and / or the coarse resolution indication data at the prediction moment within the preset period as input data, and uses the fine resolution indication data at the prediction moment of the preset period as output data for training.

[0084] In some embodiments, constructing an initial field of the pattern coarse grid resolution by using a high-resolution training data set specifically includes:

[0085] Interpolating the high-resolution training data to obtain the coarse grid resolution pattern field.

[0086] In some embodiments, driving a target artificial intelligence meteorological large model to perform continuous prediction for a preset period specifically includes:

[0087] Driving a target meteorological large model with a coarse grid resolution to perform continuous prediction from an initial moment t j-1 to a prediction moment t j , where t j-1 =t0+(j - 1)Δt, t j =t0 + jΔt), j = 1,..., Nτ , △t is the time step, τ is the time interval of high-resolution training data, N τ is the number of consecutive prediction times, and t0 is the initial time of each time interval.

[0088] In some embodiments, the coarse-resolution indication data at the initial time is the coarse-resolution variable at the initial time, the fine-resolution indication data at the initial time is the fine-resolution variable at the initial time, the coarse-resolution indication data at the prediction time is the coarse-resolution variable at the prediction time, and the fine-resolution indication data at the prediction time is the fine-resolution variable at the prediction time; correspondingly, the artificial intelligence data processing model has the following expression:

[0089]

[0090] where t j-1 represents the initial time, represents the coarse-resolution model variable at the initial time, represents the fine-resolution model variable at the initial time, t j represents the model variable at the prediction time, represents the coarse-resolution model variable at the prediction time, represents the fine-resolution model variable at the prediction time.

[0091] In some embodiments, the coarse-resolution indication data at the initial time is the time residual between the coarse-resolution at the prediction time and the coarse-resolution variable at the initial time, and the fine-resolution indication data at the initial time is the spatial residual between the fine-resolution and the coarse-resolution variable at the initial time; correspondingly, the artificial intelligence data processing model has the following expression:

[0092]

[0093] where t j-1 represents the initial time, represents the spatial residual between the prediction and the prediction of the coarse-resolution variable at the initial time, represents the spatial residual between the fine and coarse grid resolution variables at the initial time, t j represents the prediction time.

[0094] In some embodiments, the objective function is:

[0095]

[0096] where k(=1,…,T) refers to the frequency of occurrence of high-resolution training data, x h,k represents the k-th training data and

[0097] In some embodiments, the network architecture of the artificial intelligence data processing model includes, but is not limited to, residual networks, U-shaped networks, graph neural networks, or neural networks based on attention mechanisms.

[0098] In one or several of the above specific embodiments, the optimization model construction device provided by the present invention for improving the resolution and prediction accuracy of the artificial intelligence meteorological large model has the following technical effects:

[0099] Using the global / regional high-precision (re)analysis (or observation, fusion) data in the existing target dataset as training data, the global (or regional) large model AIG( / R)WFM(DepM l→l (·)) realizes two encryption functions of global and local, establishes an effective connection between the global and regional artificial intelligence large models, and can achieve multiple nested encryption through repeated encryption operations, and is compatible with traditional numerical prediction models;

[0100] In the process of generating the data processing model fully incorporates the spatio-temporal prediction evolution contained in the original low-resolution artificial intelligence global (or regional) large model AIG( / R)WFM model (DepM l→l (·)), that is, the input includes the low-resolution simulation and prediction results of the AIG( / R)WFM model (DepM l→l (·)) at the initial and prediction times;

[0101] Use artificial intelligence algorithms to train the data processing model to generate a better data optimization model

[0102]

[0103] Generally speaking, the present invention relates to fields such as meteorology, ecology, environment, new energy, and artificial intelligence. In particular, it relates to a method for improving the spatial resolution of an existing low-grid-resolution artificial intelligence meteorological large model by superimposing artificial intelligence models (such as deep neural networks, etc.), and then improving the simulation and prediction accuracy of meteorological, ecological, and environmental variables.

[0104] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 4As shown in the figure. The computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a model prediction. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The model prediction of the computer device is used to store static information and dynamic information data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0105] Those skilled in the art can understand that Figure 4 the structure shown in the figure is only a block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different component layout.

[0106] Corresponding to the above embodiments, an embodiment of the present invention further provides a computer storage medium, which contains one or more program instructions. Among them, the one or more program instructions are used to execute the method as described above.

[0107] The present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the above method.

[0108] In the embodiments of the present invention, the processor is a GPU card for AI training, such as NVIDIA H100, Moore Threads, etc.

[0109] It can implement or execute the various methods, steps, and logical block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present invention can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. The processor reads the information in the storage medium and combines its hardware to complete the steps of the above method.

[0110] The storage medium can be a memory, for example, it can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories.

[0111] Among them, the non-volatile memory can be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically Erasable PROM (EEPROM), or a flash memory.

[0112] The volatile memory can be a Random Access Memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM).

[0113] The storage medium described in the embodiments of the present invention is intended to include but not limited to these and any other suitable types of memories.

[0114] Those skilled in the art should be able to realize that in one or more of the above examples, the functions described in the present invention can be implemented by a combination of hardware and software. When applying software, the corresponding functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. The computer-readable medium includes a computer storage medium and a communication medium, where the communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0115] The above specific implementation manners further elaborate on the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above is only the specific implementation manner of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. An optimization model construction method for improving the resolution and prediction accuracy of an artificial intelligence meteorological large model, characterized in that, The method includes: Constructing an initial field of a coarse-grid resolution pattern using a high-resolution training dataset, and driving a target artificial intelligence meteorological large model with coarse-grid resolution to perform continuous forecasting for a preset period to respectively obtain coarse-grid resolution data to be optimized at multiple forecast steps; Inputting the data to be optimized corresponding to each forecast step into a pre-constructed artificial intelligence data processing model respectively, and obtaining the error value corresponding to each forecast step in combination with the training data; Summing the error values obtained at each forecast step, and taking minimizing the objective function as the training objective, training the artificial intelligence data processing model based on the sum of the error values to obtain an optimized artificial intelligence data processing model; Dividing each forecast step into multiple preset periods, and the artificial intelligence data processing model uses the coarse-resolution indication data at the initial moment, the fine-resolution indication data at the initial moment, and / or the coarse-resolution indication data at the forecast moment within the preset period as input data, and uses the fine-resolution indication data at the forecast moment of the preset period as output data for training; Among them, driving the target artificial intelligence meteorological large model with coarse-grid resolution to perform continuous forecasting for a preset period specifically includes: Drive the coarse grid resolution target meteorological large model for the initial time to the forecast time for continuous forecasting, where is the time step, is the time interval of the high-resolution training data, is the number of moments for continuous forecasting, is the initial time of each time interval; Among them, the coarse-resolution indication data at the initial moment is the coarse-resolution variable at the initial moment, the fine-resolution indication data at the initial moment is the fine-resolution variable at the initial moment, the coarse-resolution indication data at the prediction moment is the coarse-resolution variable at the prediction moment, and the fine-resolution indication data at the prediction moment is the fine-resolution variable at the prediction moment; correspondingly, the artificial intelligence data processing model has the following expression: ； Among them, represents the initial time, represents the time residual between the forecast and the coarse-resolution forecast variables at the initial time, represents the spatial residual between the fine- and coarse-grid resolution variables at the initial time, represents the forecast time.

2. The optimization model construction method for improving the resolution and prediction accuracy of the artificial intelligence meteorological large model according to claim 1, characterized in that Constructing an initial field of a coarse-grid resolution pattern using a high-resolution training dataset specifically includes: Interpolating the high-resolution training data to obtain the coarse-grid resolution pattern field.

3. The optimization model construction method for improving the resolution and prediction accuracy of the artificial intelligence meteorological large model according to claim 1, characterized in that The objective function is: ； Among them, refers to the frequency of occurrence of high-resolution training data, represents the th training data and .

4. The optimization model construction method for improving the resolution and prediction accuracy of the artificial intelligence meteorological large model according to claim 3, characterized in that, The network architecture of the artificial intelligence data processing model includes but is not limited to a residual network, a U-shaped network, a graph neural network, or a neural network based on an attention mechanism.

5. An optimization model construction device for improving the resolution and prediction accuracy of an artificial intelligence meteorological large model, based on the method according to any one of claims 1-4, characterized in that, The device includes: A data acquisition unit for constructing an initial field of a coarse-grid resolution pattern using a high-resolution training dataset, and driving a target artificial intelligence meteorological large model with coarse-grid resolution to perform continuous forecasting for a preset period to respectively obtain data to be optimized at multiple forecast steps; An outer loop unit for inputting the data to be optimized corresponding to each forecast step into a pre-constructed artificial intelligence data processing model respectively, obtaining the error value corresponding to each forecast step in combination with the training data; summing the error values obtained at each forecast step, and taking minimizing the objective function as the training objective, training the artificial intelligence data processing model based on the sum of the error values to obtain an optimized artificial intelligence data processing model; An inner loop unit for dividing each forecast step into multiple preset periods, and the artificial intelligence data processing model uses the coarse-resolution indication data at the initial moment, the fine-resolution indication data at the initial moment, and / or the coarse-resolution indication data at the forecast moment within the preset period as input data, and uses the fine-resolution indication data at the forecast moment of the preset period as output data for training.

6. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1-4.

Images