Numerical mode and large model collaborative forecasting method based on multi-dimensional water vapor assimilation

Abstract

The invention discloses a numerical mode and large model collaborative forecasting method based on multi-dimensional water vapor assimilation, and the method comprises the steps: maintaining the global large-scale features, obtained through the training of reanalysis data, of a Pantochaeus large model through the collaborative driving of a WRF numerical mode and the Pantochaeus large model; and meanwhile, the defect of a regional mode in detail description is made up through local fine tuning, a solid technical support is provided for constructing a hybrid forecasting system with generalization ability and regional forecasting precision, and the problem that an existing weather forecasting large model is insufficient in extreme weather event capturing ability is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to a numerical model and a large model cooperative prediction method based on multi-dimensional water vapor assimilation. BACKGROUND

[0002] The current meteorological prediction system still faces multiple severe challenges in achieving high-precision prediction, and a single meteorological numerical prediction model often cannot meet the demand. In view of the close relationship between tropospheric related parameters and meteorological elements, satellite assimilation data has been applied to numerical prediction models since the 1990s to improve prediction accuracy (such as the prediction model disclosed by Lorenc et al. in 2000); Vedel and Huang disclosed in 2004 that the tropospheric zenith delay was assimilated into a high-resolution research area, which improved the ability of heavy rainfall prediction. Similar experiments of assimilating water vapor data into numerical weather prediction models have also been carried out at home and abroad. With the development of water vapor monitoring technology, its advantages in numerical prediction models are becoming more and more significant. Ma et al. carried out a water vapor assimilation experiment in a certain area in 2012, which confirmed that assimilating water vapor data could effectively improve the rainfall prediction accuracy; He Zhixin et al. in 2017 found that assimilating water vapor could improve the prediction effect of rainfall area and intensity; Oigawa et al. in 2018 further assimilated water vapor data from a higher density observation network, successfully improving the prediction accuracy of a heavy rain event in a certain area; Yang et al. in 2020 used data from 66 GNSS ground observation stations to carry out an assimilation experiment for a heavy rainfall event, and the research showed that GPS data assimilation had a significant improvement effect on precipitation prediction; Risanto et al. in 2021 evaluated the impact of water vapor data assimilation on the short-term precipitation prediction of the North American monsoon in the northwest of a certain area, and the results showed that water vapor data assimilation significantly improved the accuracy of precipitation prediction.

[0003] In the traditional assimilation method, three-dimensional variational data assimilation and four-dimensional variational data assimilation techniques have gradually become the focus of meteorological assimilation research due to their strong spatiotemporal optimization capabilities. Ding Jincai et al. used three-dimensional variational system to assimilate water vapor data in 2007 and found that it could improve the rainfall prediction capability; Giannaros et al. used three-dimensional variational assimilation of tropospheric zenith delay data to predict precipitation and water vapor in a certain area in 2020, and the research results showed a positive impact. Some studies have shown that using three-dimensional / four-dimensional variational methods to assimilate external water vapor in the Weather Research and Forecasting Model (WRF) can improve the WRF model prediction performance. Rohm et al. conducted a two-month WRF data assimilation experiment in a certain area in 2019, assimilating tropospheric zenith delay and water vapor data obtained from 106 GNSS stations into the model. The study found that compared with the WRF model without assimilating data, assimilating water vapor data could reduce the water vapor bias from 2.6 kg / m² to 2.5 kg / m² in 24-hour prediction; Singh et al. used the WRF model to study the impact of tropospheric zenith delay and water vapor data assimilation on a certain area in 2019, and the results showed that it had a positive impact on precipitation prediction, as well as water vapor in the lower troposphere, temperature in the upper atmosphere, and wind field in the upper troposphere; Leontiev et al. assimilated GPS water vapor data and GPS and Meteosat satellite image fused water vapor data for prediction in a certain area in 2020, and the results showed that compared with the WRF model without assimilating data, assimilating water vapor data could improve the prediction accuracy by 11% in the first 3 hours. At the same time, based on the WRF model in the north of a certain area, water vapor assimilation experiments were conducted, and it was found that water vapor assimilation could improve the prediction accuracy of rainfall. Although variational assimilation methods have made some progress in local areas and short-term prediction, but in large-scale and complex weather conditions, the effectiveness of these methods is affected by data quality, model accuracy, and boundary condition settings.

[0004] In recent years, the innovation of artificial intelligence technology has injected new energy into the field of weather prediction. AI (Artificial Intelligence) technology represented by machine learning and deep learning has reshaped the weather forecasting technology system by innovating data modeling methods. The Pangu model innovatively uses a three-dimensional neural network architecture, which has shown superior computing performance to traditional numerical models in tasks such as typhoon path and rainfall prediction. Currently, the reliability and practicality of the Pangu weather model have been confirmed by multiple studies. For example, Bi et al. proposed a deep learning-based system, Pangu weather model, to predict global weather in 2022, and the results showed that the Pangu weather model performed well in medium and short-term prediction; Xu et al. used the Pangu weather model to successfully drive the WRF model in 2024, improving the accuracy of extreme precipitation prediction in the northern region of a certain place; Huang et al. used the Pangu model to estimate the surface air temperature from 2016 to 2019 in 2025, and the results showed that the bias and root mean square error of the estimated surface air temperature with respect to the measured surface temperature at the RS station were -0.75 K and 2.54 K, respectively, which can provide reliable high-precision weighted average temperature estimates for real-time GNSS PWV inversion.

[0005] However, there are still some challenges in existing water vapor assimilation and weather prediction technology: on the one hand, existing assimilation methods have certain limitations in capturing the dynamic changes of water vapor, especially how to efficiently assimilate two-dimensional and three-dimensional water vapor data. On the other hand, although the Pangu model has shown significant advantages in weather prediction, current research focuses mainly on the prediction and analysis of meteorological data, and has not yet explored its application in combination with water vapor assimilation. Therefore, it is necessary to further improve the prediction accuracy by assimilating two-dimensional and three-dimensional water vapor data, and to explore how to integrate the Pangu model with numerical prediction models to improve the accuracy of weather prediction. SUMMARY

[0006] In view of the problem that existing weather prediction models have insufficient ability to capture extreme weather events, the present application provides a numerical model and large model collaborative prediction method based on multi-dimensional water vapor assimilation, which cooperatively drives the WRF numerical model and the Pangu large model (i.e. the medium-term global weather prediction model), not only retains the global large-scale features trained on the reanalysis data of the Pangu large model, but also compensates for the deficiencies of regional models in detail description through local fine-tuning, providing solid technical support for building a hybrid prediction system with generalization ability and regional prediction accuracy.

[0007] According to an aspect of the present application, a numerical model and large model collaborative prediction method based on multi-dimensional water vapor assimilation is provided, comprising: Based on the hybrid ensemble-four-dimensional variational assimilation method, combined with an adaptive weight adjustment strategy, a two-dimensional multi-source water vapor initial field is constructed and dynamically assimilated; Based on the conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, the conversion of three-dimensional wet refractive index to temperature, humidity and air pressure is constructed and the three-dimensional water vapor is assimilated; The assimilated multi-dimensional water vapor data is converted to the input format of the medium-term global weather forecast model, and input into the trained WRF numerical mode driven meteorological prediction large model based on the medium-term global weather forecast model, and the prediction result is output. The WRF numerical mode driven meteorological prediction large model introduces a trainable feature adapter and a weighted loss function, and is trained by the "frozen backbone + local fine-tuning" strategy.

[0008] As a further technical solution, based on the hybrid ensemble-four-dimensional variational assimilation method, a two-dimensional multi-source water vapor initial field is constructed and dynamically assimilated, including: Based on the two-dimensional water vapor data, the water vapor initial field of the WRF model is generated; The hybrid ensemble-four-dimensional variational assimilation method is used to dynamically correct the water vapor initial field of the WRF model, wherein the hybrid ensemble-four-dimensional variational assimilation method combines ensemble covariance and static covariance to provide flow-dependent background error statistical estimates, and uses an extended control variable method to merge ensemble-based covariance estimates into the WRF model.

[0009] As a further technical solution, combined with an adaptive weight adjustment strategy, including: The background error covariance of the hybrid ensemble-four-dimensional variational assimilation is defined as: Wherein, is the climatological background error covariance matrix, is the flow-dependent error covariance matrix obtained by ensemble Kalman filter ensemble prediction, is a weight parameter used to balance the contribution of historical statistics and real-time features; The weight ratio between and is dynamically adjusted in different weather systems to obtain the optimal analysis field.

[0010] As a further technical solution, the method further includes optimizing the background error covariance in the following way: The principal component analysis method is used to reduce the dimension of the water vapor field data and extract key time-space features; The localization operation of the background error covariance matrix is defined; The error covariance of different weather systems is dynamically adjusted by the WRF model.

[0011] As a further technical solution, based on the conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, the conversion of three-dimensional wet refractive index to temperature, humidity and air pressure is constructed and the three-dimensional water vapor is assimilated, including: Converting three-dimensional wet refractive index data into relative humidity by using temperature and water vapor pressure; Assimilating the relative humidity, temperature and water vapor pressure data into the WRF model as the initial field of the WRF model.

[0012] As a further technical solution, the training of the WRF numerical mode driven weather forecast large model comprises: Freezing the main network of the medium-term global weather forecast model, and adding a trainable feature adapter at the input end, and only training the weights and biases of the adapter; Defining a loss function based on weighted mean square error to directly use the WRF interpolation result as a supervision signal.

[0013] As a further technical solution, the training of the WRF numerical mode driven weather forecast large model further comprises: After the convergence of the adapter, unfreezing the selected high-level network parameters of the medium-term global weather forecast model, and performing fine-grained optimization on the overall feature representation.

[0014] As a further technical solution, the loss function based on weighted mean square error is: , Wherein, is the prediction result of the medium-term global weather forecast model after being corrected by the adapter, is the WRF label data obtained by coordinate conversion and interpolation, and λ is a weight coefficient set according to the influence degree of each meteorological variable on regional prediction and the uncertainty of the WRF model, is the residual of the predicted value and the true value, is the set of residuals, and P is the predicted value.

[0015] According to an aspect of the present application, a numerical mode and large model cooperative prediction system based on multi-dimensional water vapor assimilation is provided for implementing the method, the system comprising: A first main module, based on a hybrid ensemble-four-dimensional variational assimilation method, combining an adaptive weight adjustment strategy, constructing a two-dimensional multi-source water vapor initial field and performing dynamic assimilation; A second main module, based on a conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, constructing a conversion of three-dimensional wet refractive index to temperature, humidity and pressure and performing three-dimensional water vapor assimilation; The third main module is based on converting the assimilated multi-dimensional water vapor data into the input format of the medium-term global weather forecast model, inputting the trained WRF numerical mode driven meteorological forecast large model, and outputting the forecast result; wherein the WRF numerical mode driven meteorological forecast large model is based on the medium-term global weather forecast model, introduces a trainable feature adapter and a weighted loss function, and is trained by the 'frozen backbone + local fine-tuning' strategy.

[0016] According to an aspect of the present application, a non-transitory computer readable storage medium is provided, which stores computer instructions for executing the numerical mode and large model collaborative forecasting method based on multi-dimensional water vapor assimilation.

[0017] The present application aims at the problem of insufficient extreme weather event capturing ability of the existing meteorological forecast large model, and constructs a numerical mode and large model collaborative forecasting method based on multi-dimensional water vapor assimilation. Compared with the prior art, the present application has the following advantages: 1. Based on the hybrid ensemble-four-dimensional variational assimilation method, combined with the adaptive weight adjustment strategy, a two-dimensional multi-source water vapor initial field optimization and dynamic assimilation method is constructed to improve the response ability of the mesoscale weather forecast model to the temporal and spatial changes of water vapor; 2. Formulate the conversion and assimilation strategy of the three-dimensional wet refractive index data, construct the conversion method of the three-dimensional wet refractive index to temperature, humidity and air pressure, and optimize the application of three-dimensional water vapor information in the mesoscale weather forecast model; 3. Construct the cross-coordinate system data conversion method of the mesoscale weather forecast model data to the Pangu model input format, introduce the trainable feature adapter module and the weighted loss function, and realize the high-precision meteorological forecast of the mesoscale weather forecast model and the Pangu large model collaborative driving by the 'frozen backbone + local fine-tuning' strategy. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 The flowchart of the numerical mode and large model collaborative forecasting method based on multi-dimensional water vapor assimilation provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0020] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application. In addition, the technical features in each embodiment or in a single embodiment provided by the present application can be combined with each other at will to form new technical solutions, and the combination is not restricted by the sequence of steps and / or the mode of structural composition, but should be based on the realization by a person of ordinary skill in the art. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the present application.

[0021] The embodiments of the present application provide a numerical model and large model cooperative prediction method based on multi-dimensional water vapor assimilation, as shown in Figure 1 The method comprises the following steps: based on a hybrid ensemble-four-dimensional variational assimilation method, a two-dimensional multi-source water vapor initial field is constructed and dynamically assimilated by combining an adaptive weight adjustment strategy; based on a conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, a conversion of three-dimensional wet refractive index to temperature, humidity and air pressure is constructed and three-dimensional water vapor is assimilated; the assimilated multi-dimensional water vapor data is converted into a medium-term global weather prediction model input format, input into a trained WRF numerical model driven meteorological prediction large model, and a prediction result is output; wherein the WRF numerical model driven meteorological prediction large model is based on a medium-term global weather prediction model, introduces a trainable feature adapter and a weighted loss function, and is trained by a “frozen backbone + local fine-tuning” strategy.

[0022] Specifically, the method comprises the following steps: (1) Two-dimensional water vapor initial field optimization and dynamic assimilation.

[0023] The accuracy of water vapor is a key factor affecting the performance of WRF. Based on high-resolution two-dimensional water vapor data , the embodiments of the present application use WRF to obtain an initial field , and compare the two-dimensional water vapor data with the model initial field, and use an error function to optimize the initial field: , In the formula: is a spatial grid point, is an error function, and the initial water vapor field of the WRF model is optimized by minimizing the error function.

[0024] This invention further employs a Hybrid Ensemble-Variational Data Assimilation (Hybrid EnVAR) method to dynamically correct the two-dimensional water vapor field of the WRF model, ensuring that the model input more closely reflects actual observations. Unlike three-dimensional and four-dimensional variational assimilation methods that use static covariance models to estimate background forecast errors, the hybrid method combines ensemble covariance with static covariance to provide flow-related statistical estimates of background errors. An extended control variable method is used to incorporate ensemble-based covariance estimates into the WRF. The Hybrid4D-EnVAR method combines the advantages of ensemble methods and four-dimensional variational assimilation, using a weighted fusion of climate statistical error covariance and ensemble error covariance to make assimilation more adaptable to the changing characteristics of different weather systems. The basic variational optimization objective function of Hybrid 4D-EnVAR is defined as follows: , In the formula: x is the analysis field variable, that is, the two-dimensional water vapor field to be updated; y represents the background field variable, i.e., the initial water vapor field obtained through the WRF model; y represents the observation data, i.e., the fused two-dimensional water vapor. is the observation operator; B and R are the background error covariance matrix and the observation error covariance matrix, respectively.

[0025] In Hybrid EnVAR, the flow-dependent error covariance and the climate statistical background error covariance are used together to improve the accuracy of water vapor assimilation. The background error covariance B of Hybrid EnVAR is defined as: , In the formula: The background error covariance matrix for climate statistics is obtained from long-term meteorological data statistics. The flow-dependent error covariance matrix is ​​obtained from the ensemble prediction of the Kalman filter set. The weighting parameters are used to balance the contributions of historical statistics and real-time features. This embodiment of the invention uses WRF to calculate the dynamic error covariance of the background field and optimizes the weight allocation of ensemble members to more accurately describe the error characteristics of the water vapor field. Furthermore, this embodiment of the invention uses an adaptive weight adjustment strategy (such as variance component estimation) to enable Hybrid EnVAR to dynamically adjust under different weather systems. and The weighting ratio between the data ensures optimal assimilation of the water vapor field, yielding the optimal analysis field x. This allows the observational data (two-dimensional water vapor data) to be integrated into the WRF model, correcting the initial water vapor field of the WRF model. This combination enables the assimilation method to consider both long-term climate statistics and the dynamic changes of instantaneous weather systems.

[0026] In order to improve the sensitivity of the Hybrid EnVAR to small and medium scale weather systems, the background error covariance is further optimized in the embodiment of the application. Specifically, principal component analysis is used to reduce the dimensionality of the water vapor field data, extract key time-space features, and reduce the computational complexity of high-dimensional data. At the same time, in order to reduce the unreasonable error propagation between distant grid points, the localization operation is used to limit the error propagation between distant grid points, and the localization operation of the background error covariance matrix is defined as follows: , In the formula: is the global calculated ensemble error covariance; is the distance between grid points , is the localization radius. The localization method can effectively reduce the unreasonable propagation of error covariance between distant grid points and improve the local analysis accuracy of the water vapor field. In addition, the error covariance of different weather systems is dynamically adjusted by using WRF to enhance the adaptability of the Hybrid EnVAR to nonlinear water vapor evolution.

[0027] (2) Three-dimensional water vapor assimilation combined with tomography technology.

[0028] In the process of assimilating two-dimensional water vapor using the WRF-DA model, only the vertical direction of the water vapor total amount information is used, and the vertical profile information of the water vapor is not used, which may limit the effect of data assimilation. The vertical structure of atmospheric water vapor is crucial for the formation and development of precipitation, but its high temporal and spatial resolution and complex vertical structure put higher requirements on the assimilation method. Therefore, the three-dimensional wet refractive index data generated by using the water vapor tomography technology is introduced into the WRF-DA model in the embodiment of the application to enrich the three-dimensional structure information of the water vapor and optimize the initial field of the WRF model.

[0029] It should be noted that in the embodiment of the application, WRF can be regarded as a prediction engine responsible for simulating atmospheric evolution, and the WRF-DA model can be regarded as a pre-processing tool responsible for optimizing the initial field of the WRF model. The output of the WRF-DA model (assimilated analysis field) is directly used as the initial condition of the WRF model. The observation data is first input into the WRF-DA model, and the WRF-DA model generates an analysis field through the variational or ensemble method. The analysis field is used as the initial field of the WRF model to start the prediction simulation.

[0030] ​WRF model itself does not have the ability to generate initial field independently, and it needs to rely on the initial conditions provided by external data sources (such as global model output or reanalysis data). Common sources include: such as GFS (USA), ECMWF (European Center), NCEP FNL (reanalysis data). These data are converted into WRF readable format by WRF preprocessing tools, and initial field and boundary conditions are generated by built-in functions of WRF.

[0031] Three-dimensional wet refractive index can provide water vapor information at different height layers, making up for the lack of vertical distribution information of two-dimensional water vapor data. However, WRF-DA model cannot directly assimilate wet refractive index, but needs to be converted into assimilable basic meteorological variables such as relative humidity (Relative Humidity, RH), temperature (T) and water vapor pressure. Wet refractive index (WetRefractivity, WR) is closely related to water vapor pressure and temperature. And the relationship between RH is as follows: , In the formula, is the saturated water vapor pressure, which is related to temperature, and can be calculated by the following formula: , In the formula, , , is a constant. In the process of converting wet refractive index into temperature, humidity and water vapor pressure combination, first, the temperature and water vapor pressure are used to convert the wet refractive index into relative humidity, and then the RH and temperature and water vapor pressure data are assimilated into the WRF-DA model to generate the initial analysis field. And using the HybridEnVAR method, combined with the ensemble Kalman filter to dynamically calculate the background error covariance. In addition, in order to optimize the adaptability of three-dimensional water vapor information in different weather systems, the background error covariance structure is further adjusted in the embodiment of the application.

[0032] (3) WRF numerical model and disc big model cooperative driving weather forecast.

[0033] After the two-dimensional and three-dimensional water vapor assimilation is completed, the embodiment of the application takes the WRF model output as a high-precision label, performs model fine-tuning on the Pangu large model guided by physics, and constructs a meteorological prediction large model driven by the WRF numerical model. The Pangu large model is a medium-term global weather prediction model constructed based on a 3D neural network. The model unifies and combines meteorological information of different longitudes, latitudes and altitudes to form a standardized atmospheric state field (such as three-dimensional grid data of an isobaric surface or a geopotential height surface) as model input. WRF is a regional high-resolution prediction tool, which uses terrain-following coordinates or hybrid vertical coordinates in the vertical direction, and meteorological variables are stored in eta layers, which leads to a significant difference between the output of WRF and the input format of models such as Pangu. Therefore, it is necessary to convert the vertical coordinate system data of the WRF model to the standard pressure layer.

[0034] For terrain-following coordinates, the eta layer pressure can be expressed as a linear combination of the surface pressure and the pressure at the top of the model . , represents the number of eta layers. For hybrid coordinates, the background pressure profile is superimposed on the terrain disturbance term . By inversely calculating the actual pressure of each eta layer, the original variables (such as potential temperature, humidity, wind field) are projected to the standard isobaric surface using logarithmic linear interpolation.

[0035] The expression of the variable to be interpolated Q on the isobaric surface is: , , respectively represent the meteorological variables to be interpolated. Among them, and are the pressures of adjacent eta layers.

[0036] The interpolation of the geopotential height surface needs to be inverted through the potential height formula: , where PH is the perturbation potential, PHB is the base potential, g is the acceleration of gravity, represents the geodetic height. After interpolation, the three-dimensional grid data needs to be resampled to the resolution matched with the Pangu model, and the scale difference between the WRF model and the Pangu model is eliminated through bilinear interpolation.

[0037] After the data conversion and preprocessing are completed, in order to further realize fine tuning of the Pangu large model, an embodiment of the present application adopts the strategy of freezing the model backbone network and adding a trainable external feature adapter at the input end. Specifically, under the premise of freezing all layer parameters of the Pangu large model, a lightweight adapter module is added before the model input, and the input feature X is output after being converted by the adapter: , Wherein, W1, W2 are weight matrices, b1, b2 are bias terms, and sigma is a Swish activation function. In order to ensure that the model output can fully reflect the high-precision information of the WRF numerical prediction, an embodiment of the present application defines a loss function based on weighted mean square error, so as to directly use the WRF interpolation result as a supervision signal.

[0038] The loss function can be expressed as: (13), Wherein, is the prediction result of the Pangu large model after adapter correction, is the WRF label data obtained by coordinate conversion and interpolation, lambda is a weight coefficient set according to the influence degree of each meteorological variable on regional prediction and the uncertainty of the WRF model, P is a prediction value, is the residual of the prediction value and the true value, is the set of residuals.

[0039] In the model training process, only the weights and biases of the adapter are trained, and all layer parameters of Pangu are set to be untrainable. After the adapter converges, some high layer network parameters of the Pangu large model can be further unfrozen, and the overall feature representation can be further optimized in a more fine-grained manner, so as to realize higher adaptability and prediction accuracy of regional details. In the Pangu weather large model (Pangu-Weather), the high layer network parameters refer to the key learnable parameters in the deep neural network architecture of the model at the top of the model, which are responsible for processing high-level meteorological features. These parameters directly affect the abstraction ability and prediction performance of the model on complex meteorological laws. After the adapter converges, some of its corresponding parameters can also be trained.

[0040] Through the cooperative driving of the WRF numerical model and the Pangu large model, not only the global large-scale features trained on the reanalysis data by the Pangu large model are retained, but also the deficiencies of the regional model in detail description are made up through local fine tuning, thereby providing solid technical support for constructing a hybrid prediction system with generalization ability and regional prediction accuracy.

[0041] The implementation basis of each embodiment of the present application is achieved by a programmed process through a device with processor function. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present application are packaged into various modules. Based on this actual situation, on the basis of each embodiment described above, an embodiment of the present application provides a numerical model and large model collaborative forecasting system based on multi-dimensional water vapor assimilation, which is used to execute the numerical model and large model collaborative forecasting method based on multi-dimensional water vapor assimilation in the method embodiment described above.

[0042] The system comprises: a first main module, based on a hybrid ensemble-four-dimensional variational assimilation method, combined with an adaptive weight adjustment strategy, to construct a two-dimensional multi-source water vapor initial field and perform dynamic assimilation; a second main module, based on a conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, to construct a conversion of three-dimensional wet refractive index to temperature, humidity and air pressure and perform three-dimensional water vapor assimilation; and a third main module, based on conversion of the assimilated multi-dimensional water vapor data to a medium-term global weather forecast model input format, to input a trained WRF numerical mode driven meteorological prediction large model and output a prediction result; wherein the WRF numerical mode driven meteorological prediction large model is based on a medium-term global weather forecast model, introduces a trainable feature adapter and a weighted loss function, and is trained by a "frozen backbone + local fine-tuning" strategy.

[0043] The numerical model and large model collaborative forecasting system based on multi-dimensional water vapor assimilation provided by the embodiment of the present application, aiming at the problem that the existing meteorological prediction large model has insufficient ability to capture extreme weather events, uses the aforementioned several modules, and through the collaborative driving of the WRF numerical mode and the Pangu large model (i.e. the medium-term global weather forecast model), not only retains the global large-scale features trained on the reanalysis data by the Pangu large model, but also compensates for the deficiencies of the regional mode in detail description through local fine-tuning, providing solid technical support for building a hybrid prediction system with generalization ability and regional prediction accuracy.

[0044] It should be noted that the system embodiment provided by the present application is used to implement the method in the method embodiment described above, and is also used to implement the method in other method embodiments provided by the present application. The difference is only that the corresponding function modules are set, and the principle is basically the same as that of the above-mentioned system embodiments provided by the present application. As long as the person skilled in the art improves the modules in the above-mentioned system embodiments on the basis of the above-mentioned system embodiments, refers to the specific technical solutions in other method embodiments, obtains the corresponding technical means by combining technical features, and the technical solutions composed of these technical means, as long as the technical solutions have practicality, the corresponding system class embodiments are obtained, which are used to implement the method in other method class embodiments.

[0045] Based on the same inventive concept as the foregoing embodiments, the embodiments of the present application also provide a non-transitory computer-readable storage medium storing computer instructions, which cause the computer to perform the numerical model and large model collaborative forecasting method based on multi-dimensional water vapor assimilation.

[0046] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) having computer-usable program code embodied therein.

[0047] The present application is described in reference to the flowcharts and / or block diagrams of the methods, apparatus (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by one or more blocks.

[0048] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implement the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by one or more blocks.

[0049] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide functions for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by one or more blocks.

[0050] It should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the embodiments of the present application.

Claims

1. A numerical model and large-scale model collaborative forecasting method based on multidimensional water vapor assimilation, characterized in that, include: Based on the hybrid set-four-dimensional variational assimilation method, combined with an adaptive weight adjustment strategy, a two-dimensional multi-source water vapor initial field is constructed and dynamically assimilated. Based on the conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, a conversion of three-dimensional wet refractive index to temperature, humidity and air pressure is constructed and three-dimensional water vapor is assimilated. The assimilated multidimensional water vapor data is converted to the input format of the medium-term global weather forecast model, input into the trained WRF numerical model-driven meteorological forecast model, and the forecast results are output. The WRF numerical model-driven meteorological forecast model is based on the medium-term global weather forecast model, introduces a trainable feature adapter and a weighted loss function, and is trained through a "frozen trunk + local fine-tuning" strategy.

2. The numerical model and large model co-prediction method based on multidimensional water vapor assimilation according to claim 1, characterized in that, Based on the hybrid ensemble-four-dimensional variational assimilation method, a two-dimensional multi-source water vapor initial field is constructed and dynamically assimilated, including: Based on two-dimensional water vapor data, an initial water vapor field for the WRF model is generated. A hybrid ensemble-four-dimensional variational assimilation method is used to dynamically correct the initial water vapor field of the WRF model. The hybrid ensemble-four-dimensional variational assimilation method combines the ensemble covariance with the static covariance to provide a statistical estimate of the flow-related background error, and uses an extended control variable method to incorporate the ensemble-based covariance estimate into the WRF model.

3. The numerical model and large model collaborative forecasting method based on multidimensional water vapor assimilation according to claim 2, characterized in that, Combined with adaptive weight adjustment strategies, including: Define the background error covariance of the mixed set-four-dimensional variational assimilation as: ,in, The background error covariance matrix for climate statistics. The flow-dependent error covariance matrix is ​​obtained from the ensemble prediction using the Kalman filter. This is a weighting parameter used to balance the contributions of historical statistics and real-time features; Dynamically adjust to different weather systems and The weight ratios between them are determined to obtain the optimal analysis field.

4. The numerical model and large model collaborative forecasting method based on multidimensional water vapor assimilation according to claim 3, characterized in that, The method also includes optimizing the background error covariance in the following manner: Principal component analysis was used to reduce the dimensionality of the water vapor field data and extract key temporal-spatial features. Define the localization operation of the background error covariance matrix; The error covariance of different weather systems is dynamically adjusted using the WRF model.

5. The numerical model and large model collaborative forecasting method based on multidimensional water vapor assimilation according to claim 1, characterized in that, Based on the conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, a conversion mechanism is constructed to transform three-dimensional wet refractive index into temperature, humidity, and air pressure, and three-dimensional water vapor assimilation is performed, including: Three-dimensional wet refractive index data are converted into relative humidity using temperature and water vapor pressure; The relative humidity, temperature, and water vapor pressure data are assimilated into the WRF mode and used as the initial field for the WRF mode.

6. The numerical model and large model co-prediction method based on multidimensional water vapor assimilation according to claim 1, characterized in that, The training of the large-scale weather forecasting model driven by the WRF numerical model includes: Freeze the backbone network of the medium-term global weather forecast model and add a trainable feature adapter at the input, training only the weights and biases of the adapter; Define a loss function based on weighted mean square error to directly utilize the WRF interpolation results as a monitoring signal.

7. The numerical model and large model co-prediction method based on multidimensional water vapor assimilation according to claim 6, characterized in that, The training of the large-scale weather forecasting model driven by the WRF numerical model also includes: After the adapter converges, the selected high-level network parameters of the medium-term global weather forecast model are unfrozen, and the overall feature representation is fine-tuned.

8. The numerical model and large model collaborative forecasting method based on multidimensional water vapor assimilation according to claim 6, characterized in that, The loss function based on weighted mean square error is: ， in, These are the forecast results from the medium-range global weather forecast model after adapter correction. The WRF label data is obtained through coordinate transformation and interpolation, where λ is a weighting coefficient set according to the influence of each meteorological variable on regional forecasts and the uncertainty of the WRF model. It is the residual between the predicted value and the actual value. It is the set of residuals, and P is the predicted value.

9. A numerical model and large-scale model collaborative forecasting system based on multidimensional water vapor assimilation, used to implement the method described in any one of claims 1-8, characterized in that, The system includes: The first main module is based on the hybrid set-four-dimensional variational assimilation method, combined with an adaptive weight adjustment strategy, to construct a two-dimensional multi-source water vapor initial field and perform dynamic assimilation. The second main module, based on the conversion and assimilation strategy of tomographic three-dimensional wet refractive index data, constructs the conversion of three-dimensional wet refractive index to temperature, humidity and air pressure and performs three-dimensional water vapor assimilation. The third main module is based on converting the assimilated multidimensional water vapor data into the input format of the medium-term global weather forecast model, inputting it into the trained WRF numerical model-driven meteorological forecast model, and outputting the forecast results. The WRF numerical model-driven meteorological forecast model is based on the medium-term global weather forecast model, introduces a trainable feature adapter and a weighted loss function, and is trained through a "frozen backbone + local fine-tuning" strategy.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the numerical model and large model co-forecasting method based on multidimensional water vapor assimilation as described in any one of claims 1 to 8.

Images