A meteorological observation correction system, a control method and device thereof, and a storage medium

Abstract

The application belongs to the technical field of meteorological correction, and provides a meteorological observation correction system, a control method and equipment thereof, and a storage medium. Multiple local computing nodes input private meteorological observation data and a numerical weather prediction background field into a meteorological observation correction global model to obtain a meteorological observation bias correction quantity and a corrected wind speed vector field in combination with the background field, calculate a local gradient according to a local loss function including a data loss term and a physical regularization term, and obtain a noise-added gradient through gradient clipping and differential privacy Gaussian noise addition. A central server adopts a secure aggregation mechanism to obtain an aggregation result and generate candidate global model parameters, and performs physical consistency checking based on a candidate corrected wind speed vector field; when a preset physical consistency condition is not met, the update direction or update step is physically constrained before being issued. The application can balance data privacy protection and meteorological physical consistency of correction results.

Description

Technical Field

[0001] This application belongs to the field of meteorological correction technology, specifically relating to a meteorological observation correction system and its control method, equipment, and storage medium. Background Technology

[0002] Meteorological observation data typically includes wind speed, temperature, humidity, etc., and is usually stored in different institutions such as wind farms, weather stations, and satellite centers. It may also involve data security, trade secret, or personal information protection requirements. Because raw meteorological observation data is difficult to store and process centrally, cross-institutional meteorological data fusion and correction face privacy and security limitations.

[0003] In meteorological operations, to improve the quality of initial forecast fields, it is often necessary to fuse meteorological observation data with the background field from numerical weather prediction (NMR) data, making the atmospheric conditions calculated by the model closer to the actual atmospheric conditions. Commonly used data assimilation methods include three-dimensional variational assimilation, four-dimensional variational assimilation, and ensemble Kalman filtering (EnKF). However, these methods typically rely on centralized acquisition, unified quality control, and unified computational processing of the raw meteorological observation data. When meteorological observation data is dispersed across multiple institutions, regions, or operational systems, and direct sharing of the raw data is not possible, traditional centralized data assimilation processes are easily constrained by data ownership, compliance requirements, and privacy and security limitations.

[0004] Chinese invention patent application CN119358016A discloses a distributed model parameter privacy protection method, system, medium, and program. This method involves each node training an initial global model using local data to obtain a local model update. Noise data generated by a differential privacy algorithm is introduced into the local model update to obtain local noise gradients. A central server then aggregates the parameters from multiple local noise gradients and updates the global model parameters. This approach enables joint modeling without sharing local raw data and reduces the risk of privacy leakage during model update transmission.

[0005] However, meteorological observation correction differs from general distributed modeling tasks. Meteorological observation data exhibits significant spatiotemporal correlations and physical constraints. The correction results not only need to match local meteorological observation data but also conform to meteorological physical conservation laws. While using only general federated learning and differential privacy-based noise reduction can reduce the risk of leakage of original data and model update information, the noise introduced by the differential privacy mechanism can perturb local gradients. After aggregating the noisy gradients from multiple local computing nodes, the resulting global model update may still deviate from meteorological physical conservation laws. If the central server directly updates and distributes global model parameters based on the aggregation results, the global model output after aggregation and updating may not conform to the actual atmospheric evolution. Therefore, existing solutions still struggle to simultaneously ensure data privacy protection, multi-source model collaborative updates, and meteorological physical consistency after aggregation and updating in cross-institutional meteorological observation correction scenarios. Summary of the Invention

[0006] This application addresses the technical problem in cross-agency meteorological observation correction where general federated learning privacy protection schemes struggle to simultaneously ensure data privacy protection, multi-source model collaborative updates, and meteorological-physical consistency after aggregated updates without allowing the original meteorological observation data to leave the domain. It provides a meteorological observation correction system and its control method, equipment, and storage medium.

[0007] To achieve the above objectives, this application adopts the following technical solution: In the first aspect, this application proposes a meteorological observation correction system, including a central server and multiple local computing nodes; Multiple local computing nodes are used to receive the current global model parameters issued by the central server, input the private meteorological observation data held by each local computing node and the numerical weather prediction background field into the global meteorological observation correction model to obtain the meteorological observation bias correction amount, and then combine it with the numerical weather prediction background field to obtain the corrected wind speed vector field; calculate the local gradient according to the local loss function including the data loss term and the physical regularization term, wherein the physical regularization term is constructed based on the fluid dynamics equation residuals of the corrected wind speed vector field; prune the local gradient and add Gaussian noise according to the differential privacy mechanism to obtain multiple noisy gradients, which are used to calculate the corresponding local gradients according to the local loss function including the physical regularization term based on the current global model parameters issued by the central server and the private meteorological observation data held by each local computing node, and then add noise to the corresponding local gradients to obtain multiple noisy gradients; wherein the physical regularization term is constructed based on the fluid dynamics equation residuals to constrain the correction results output by the global model to satisfy the meteorological physical conservation law; the global model is used for meteorological observation correction; The central server is used to aggregate multiple noisy gradients using a secure aggregation mechanism to obtain an aggregation result; generate candidate global model parameters based on the aggregation result, and obtain a candidate corrected wind speed vector field based on the candidate global model parameters; calculate the hydrodynamic equation residuals of the candidate corrected wind speed vector field and perform a physical consistency check; when the candidate corrected wind speed vector field does not meet the preset physical consistency condition, constrain the update direction or update step size corresponding to the aggregation result according to the physical conservation law to obtain updated global model parameters; distribute the candidate global model parameters that meet the preset physical consistency condition or the updated global model parameters obtained by the constraint processing to multiple local computing nodes to receive multiple noisy gradients obtained by multiple local computing nodes and aggregate them to obtain an aggregation result; and then update the global model parameters based on the aggregation result.

[0008] Furthermore, the local loss function also includes a data loss term formed from private meteorological observation data; the data loss term is used to characterize the deviation between the corrected wind speed vector field and the local real meteorological observation results, reference values, or operational objectives. The physical regularization term and the data loss term, weighted according to a preset weight, together constitute the local loss function. The deviation between the global model output and local real meteorological observation results, reference values, or operational objectives is also considered in the local loss function, whereby the physical regularization term and the data loss term, weighted according to a preset weight, together constitute the local loss function.

[0009] Furthermore, the local loss function satisfies the L-smoothness condition and the μ-strong convexity property.

[0010] Furthermore, the fluid dynamics equation residuals include at least one of the Navier-Stokes equation residuals, continuity equation residuals, energy conservation equation residuals, and momentum conservation equation residuals.

[0011] Furthermore, in the wind field correction scenario, the residuals of the fluid dynamics equations are the residuals of the Navier-Stokes equations, and the physical regularization term... satisfy:

[0012] in, These are weighting coefficients used to adjust the weights of the residuals in the fluid dynamics equations within the physical regularization term. Velocity in a meteorological field or fluid state. For time, For kinematic viscosity, For density, For air pressure, For gradient operators, For the Laplace operator, This indicates taking the square of the second norm of the Navier-Stokes equation residual vector within the parentheses.

[0013] Further, the local gradient is clipped and Gaussian noise is added according to a differential privacy mechanism. This noise addition results in multiple noisy gradients, including: The local gradient corresponding to each local computing node is clipped according to a preset gradient clipping threshold. The noise scale of Gaussian noise is determined based on the gradient clipping threshold, the preset privacy budget ε, and the preset privacy relaxation parameter δ. Gaussian noise is added to the clipped local gradients according to the noise scale to obtain the noisy gradients corresponding to each local computing node. Gaussian noise is added to the local gradients corresponding to each local computing node according to the differential privacy mechanism to obtain the noisy gradients corresponding to each local computing node.

[0014] Furthermore, the method of aggregating multiple noisy gradients using a secure aggregation mechanism includes: the central server aggregating the multiple noisy gradients based on a federated averaging algorithm framework.

[0015] Secondly, this application proposes a control method for the aforementioned meteorological observation correction system, comprising: Multiple local computing nodes receive the current global model parameters from the central server. The private meteorological observation data held by each local computing node, along with the numerical weather prediction background field, are input into the global meteorological observation correction model to obtain the meteorological observation bias correction amount. This is then combined with the numerical weather prediction background field to obtain the corrected wind speed vector field. Local gradients are calculated using a local loss function that includes a data loss term and a physical regularization term. The physical regularization term is constructed based on the residuals of the fluid dynamics equations of the corrected wind speed vector field. The local gradients are then pruned and Gaussian noise is added according to a differential privacy mechanism to obtain multiple noisy gradients. These gradients are calculated using the local loss function, which includes a physical regularization term, based on the current global model parameters from the central server and the private meteorological observation data held by each local computing node. Noise is then added to each of these local gradients to obtain multiple noisy gradients. The physical regularization term, constructed based on the residuals of the fluid dynamics equations, is used to constrain the correction results output by the global model to satisfy meteorological physical conservation laws. The global model is used for meteorological observation correction. The central server employs a secure aggregation mechanism to aggregate multiple noisy gradients, yielding an aggregation result. Candidate global model parameters are generated based on the aggregation result, and a candidate corrected wind speed vector field is obtained based on these parameters. The hydrodynamic equation residuals of the candidate corrected wind speed vector field are calculated and a physical consistency check is performed. When the candidate corrected wind speed vector field does not meet a preset physical consistency condition, the update direction or update step size corresponding to the aggregation result is constrained according to physical conservation laws, resulting in updated global model parameters. The candidate global model parameters that meet the preset physical consistency condition, or the updated global model parameters obtained through the constraint processing, are distributed to multiple local computing nodes. These local computing nodes receive and aggregate multiple noisy gradients, yielding an aggregation result. Finally, the global model parameters are updated based on the aggregation result.

[0016] Thirdly, this application proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the control method for the meteorological observation correction system described above.

[0017] Fourthly, this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the control method for the meteorological observation correction system described above.

[0018] Compared with the prior art, this application has the following beneficial effects: This application proposes a meteorological observation correction system. Multiple local computing nodes input their respective private meteorological observation data and numerical weather prediction (NMR) background fields into a global meteorological observation correction model to obtain meteorological observation bias correction amounts. These are then combined with the NMR background fields to obtain a corrected wind speed vector field, achieving multi-node collaborative correction without centrally aggregating the original meteorological observation data. By introducing a physical regularization term based on the residuals of the fluid dynamics equations derived from the corrected wind speed vector field into the local loss function, the local gradient is constrained by meteorological physical conservation laws during its formation. Furthermore, by pruning the local gradients and adding Gaussian noise according to a differential privacy mechanism, the risk of inferring private meteorological observation data from gradient information is reduced. Further, the central server employs a secure aggregation mechanism to aggregate multiple noisy gradients, generates candidate global model parameters based on the aggregation results, and performs physical consistency checks based on the candidate corrected wind speed vector fields. When a candidate corrected wind speed vector field does not meet the preset physical consistency conditions, the update direction or update step size corresponding to the aggregation result is constrained based on physical conservation laws. This reduces the possibility of global model updates deviating from meteorological and physical laws due to differential privacy noise and multi-node aggregation, thereby improving the meteorological and physical consistency of correction results while protecting the privacy of cross-agency meteorological observation data.

[0019] This application also proposes a control method for a meteorological observation correction system, an electronic device, and a computer-readable storage medium, which possess all the advantages of the aforementioned meteorological observation correction system. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of a framework for the meteorological observation correction system of this application; Figure 2 This is a flowchart illustrating a control method for the meteorological observation correction system of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0024] Meteorological observation data typically originates from various observation platforms, including ground-based weather stations, automatic weather stations, radar, satellites, buoys, aircraft, ships, and industry-built sensor networks. Data types include wind speed, wind direction, temperature, humidity, air pressure, precipitation, and radiation. Because different observation entities are distributed across different regions, industries, or institutions, meteorological observation data naturally exhibits characteristics such as dispersed sources, large spatial span, strong temporal continuity, and diverse data formats. Furthermore, some meteorological observation data may reflect the environmental conditions of key areas, the operational status of critical infrastructure, patterns of commercial production activities, or information related to individual locations, thus potentially involving trade secrets or personal information. In such cases, centrally aggregating and processing raw meteorological observation data from various sources on a single platform is often constrained by factors such as data ownership, compliance requirements, security reviews, and privacy protection, making it difficult to implement smoothly in actual operations.

[0025] In meteorological operations, to improve the quality of the initial forecast field, it is usually necessary to fuse meteorological observation data with the background field from numerical weather prediction. This process is generally called data assimilation, which involves using observational information to correct the atmospheric state calculated by the model, making it closer to the actual atmospheric state. Commonly used data assimilation methods include three-dimensional variational data assimilation (3D-Var), four-dimensional variational data assimilation (4D-Var), and ensemble Kalman filter (EnKF). These methods are relatively mature in traditional meteorological forecasting systems, but they typically rely on centralized acquisition, unified quality control, and unified computational processing of raw meteorological observation data. When meteorological observation data is distributed across multiple institutions, regions, or business systems, and the original data cannot be directly shared, the traditional data assimilation process will be significantly limited: on the one hand, the central node will have difficulty obtaining a sufficiently complete set of observation samples, affecting the accuracy of the correction results; on the other hand, forcibly centralizing data may lead to problems such as data leakage, unauthorized access, compliance risks, and excessively high costs for cross-institutional collaboration.

[0026] To address the issue of centralized data sharing, federated learning has been proposed in existing technologies. Federated learning is a distributed machine learning method whose basic idea is to retain the original data locally at each participating party, uploading only model parameters, gradients, or intermediate update results. A central endpoint or collaborative mechanism then aggregates the models, thereby achieving joint modeling with data remaining within its domain to a certain extent. When introducing federated learning into meteorological observation data correction or assimilation scenarios, each observation agency can train its model locally using its own meteorological data and then aggregate parameters to form a global model, reducing the need for cross-domain transmission of original meteorological observation data. Furthermore, to further protect model update information, existing solutions may incorporate privacy protection technologies such as Differential Privacy (DP), secure multi-party computation, or homomorphic encryption to reduce the risk of sensitive information being back-engineered during model interaction.

[0027] However, meteorological data, compared to conventional image and text data, exhibits stronger physical constraints and spatiotemporal evolution characteristics, making direct application of traditional federated learning frameworks insufficient. First, meteorological observation data exhibits strong spatiotemporal correlation; data distribution varies significantly across different regions, altitudes, and time periods, often failing to meet the assumption of independent and identically distributed models. This can lead to substantial deviations in the update directions of local models, impacting the convergence speed and stability of the global model. Second, purely data-driven models rely heavily on sample statistical regularities for learning. Without physical constraints such as meteorological dynamics, thermodynamics, or mass conservation, correction results may be inconsistent with actual atmospheric processes. Third, while introducing differential privacy and encrypted computation mechanisms, model updates may be affected by noise disturbances or computational accuracy loss. Although this improves privacy and security, it can also reduce the accuracy of meteorological element corrections, making it difficult to meet the accuracy and stability requirements of operational forecasting and disaster warning. Therefore, current technologies still struggle to simultaneously balance model convergence assurance, physical consistency constraints, and privacy protection with operational utility under conditions of distributed meteorological data.

[0028] Based on the above, this application proposes a meteorological observation correction system and its control method, equipment, and storage medium. The following describes the application in further detail with reference to embodiments and accompanying drawings.

[0029] Example 1 This application proposes a meteorological observation correction system, which includes a central server and multiple local computing nodes. The multiple local computing nodes are deployed on different meteorological observation data holding sides, and each local computing node holds its own corresponding private meteorological observation data. The central server does not centrally store the original private meteorological observation data of each local computing node.

[0030] During a model training or calibration iteration, the central server distributes the current global model parameters to multiple local computing nodes. Each local computing node, upon receiving these parameters, inputs its own private meteorological observation data and the numerical weather prediction (NMR) background field into the global meteorological observation calibration model. The global meteorological observation calibration model uses the NMR background field as the background information to be calibrated and, by combining its local private meteorological observation data, learns the deviation relationship between meteorological observations and the NMR background field to obtain a meteorological observation deviation correction. This correction is then combined with the NMR background field to obtain the calibrated wind speed vector field. Thus, each local computing node can calibrate the NMR background field using its local meteorological observation information without uploading the original private meteorological observation data to the central server.

[0031] Each local computing node calculates its local gradient using a local loss function that includes a data loss term and a physical regularization term. The data loss term is used to match the corrected wind speed vector field with local meteorological observation information; the physical regularization term, constructed based on the residuals of the hydrodynamic equations of the corrected wind speed vector field, introduces the physical constraints required for meteorological observation correction during gradient calculation. Therefore, the local gradient not only reflects the direction of adjustment of the global model parameters for meteorological observation correction by private meteorological observation data, but is also constrained by meteorological physical conservation laws, thereby reducing the possibility that updating the model solely based on data fitting relationships could lead to the corrected wind speed vector field deviating from actual atmospheric physical laws.

[0032] After obtaining their local gradients, each local computing node prunes these gradients and adds Gaussian noise to them according to a differential privacy mechanism, creating multiple noisy gradients. Gradient pruning limits the impact of a single local model update on the uploaded gradients, while Gaussian noise randomly perturbs the pruned gradient information. Since the noisy gradients, processed with pruning and differential privacy, are uploaded to the central server, rather than the original private meteorological observation data or unprotected local gradients, the risk of the central server or other participants inferring local private meteorological observation data from the uploaded information is reduced.

[0033] The central server employs a secure aggregation mechanism to aggregate the noisy gradients uploaded from multiple local computing nodes, yielding an aggregation result. This secure aggregation mechanism allows the central server to obtain aggregated information of the noisy gradients from multiple local computing nodes without needing to obtain the plaintext noisy gradients of individual local computing nodes, thereby further reducing the risk of leakage of model update information from individual participating nodes. Based on the aggregation result, the central server generates candidate global model parameters, enabling model updates from multiple local computing nodes based on different private meteorological observation data to be collectively reflected in the candidate global model.

[0034] Because differential privacy noise is added to the gradients uploaded by each local computing node, and the data distribution and model update direction may differ among different local computing nodes, the candidate global model parameters formed after aggregating multiple noisy gradients may still cause the model output to deviate from the conservation laws of meteorological physics. Therefore, before distributing model parameters to each local computing node, the central server obtains a candidate corrected wind speed vector field based on the candidate global model parameters, calculates the hydrodynamic equation residuals of the candidate corrected wind speed vector field, and performs a physical consistency check to determine whether the candidate global model formed in this round of aggregation maintains the physical consistency required for meteorological observation correction.

[0035] When the candidate corrected wind speed vector field meets the preset physical consistency condition, it indicates that the candidate global model parameters formed based on the current aggregation result can maintain the corresponding meteorological physical consistency of the candidate corrected wind speed vector field. The central server then distributes the candidate global model parameters to multiple local computing nodes. When the candidate corrected wind speed vector field does not meet the preset physical consistency condition, it indicates that the model update corresponding to the current aggregation result may cause the candidate corrected wind speed vector field to deviate from the physical conservation law. The central server constrains the update direction or update step size corresponding to the aggregation result according to the physical conservation law to obtain the updated global model parameters, and then distributes the constrained updated global model parameters to multiple local computing nodes. Thus, physical constraints not only act on the local gradient formation process of each local computing node, but also on the global model update process after aggregation by the central server, thereby reducing the impact of differential privacy noise and multi-node aggregation on the physical consistency of the global model.

[0036] After receiving candidate global model parameters or updated global model parameters from the central server, each local computing node updates its local calibration model accordingly and enters the next round of model training or calibration iteration. Through the above cyclical process, the meteorological observation calibration system can collaboratively train the global meteorological observation calibration model using meteorological observation information from multiple local computing nodes without centrally storing the original private meteorological observation data of each local computing node. At the same time, through local physical regularization, differential privacy protection, secure aggregation, and physical consistency checks and constraint processing on the central server side, it takes into account data privacy protection, multi-source model collaborative updates, and meteorological physical consistency after aggregation updates during the cross-institutional meteorological observation calibration process.

[0037] Example 2 In some embodiments of this application, the meteorological observation correction system is applicable to cross-agency meteorological observation correction scenarios that combine privacy computing with meteorological data assimilation. Private meteorological observation data can originate from ground meteorological stations, wind farm anemometer towers, ocean buoys, satellite remote sensing platforms, radar systems, airborne observation platforms, or industry sensor networks. Private meteorological observation data may include wind speed, wind direction, temperature, humidity, air pressure, precipitation, radiation, wave height, sea surface temperature, cloud cover, visibility, atmospheric boundary layer parameters, or other meteorologically relevant observations. Furthermore, private meteorological observation data may also include observation time information, geographical location information, altitude layer information, and historical meteorological characteristic data corresponding to the numerical weather prediction background field. Since this type of data is typically dispersed across different meteorological observation agencies, wind farms, meteorological stations, satellite centers, or other operational units, and may involve national security, trade secrets, or personal information, the raw data is not suitable for centralized storage and unified processing. Compared to data assimilation methods such as 3D-Var, 4D-Var, and EnKF, which require centralized aggregation of meteorological observation data, this embodiment completes local computation on each data holder side through federated learning. The central server only aggregates the noisy gradients uploaded by each local computing node and updates the model.

[0038] In practical applications, meteorological data exhibits strong spatiotemporal correlation and non-independent identically distributed characteristics. Directly employing ordinary federated learning training methods may lead to difficulties in guaranteeing convergence. Purely data-driven models may also output correction results that do not conform to physical laws. Furthermore, if the gradient uploading process lacks privacy protection mechanisms, there is a risk of inferring the original data from the gradient information. This embodiment addresses these issues by introducing a physical regularization term into the local loss function, pruning the local gradient and adding Gaussian noise according to a differential privacy mechanism, and employing a secure aggregation mechanism on the central server side to perform physical consistency checks on the candidate corrected wind speed vector fields. This approach aims to balance convergence analysis, physical consistency constraints, and privacy protection requirements during model training.

[0039] like Figure 1As shown, in some optional embodiments, the meteorological observation correction system can adopt a three-tier federated architecture. The data layer consists of private meteorological observation data held by multiple meteorological observation agencies, which are stored locally and not uploaded to the central server. The computation layer includes multiple local computing nodes, each equipped with a local data source, a local trainer, an encrypted upload module, and a local correction model. The local trainer uses the private meteorological observation data from the local data source and the current global model parameters issued by the central server to perform gradient calculations. The encrypted upload module is used to upload the processed gradient information. The coordination layer includes a central server, which can be equipped with a model receiving module, a secure aggregation module, a global model updater, a physical consistency check and constraint processing module, and a model broadcasting module. The model receiving module receives the noisy gradients uploaded by each local computing node; the secure aggregation module aggregates multiple noisy gradients using a secure aggregation mechanism; the global model updater generates candidate global model parameters based on the aggregation results; the physical consistency check and constraint processing module obtains candidate corrected wind speed vector fields based on the candidate global model parameters, calculates the hydrodynamic equation residuals of the candidate corrected wind speed vector fields and performs physical consistency checks, and constrains the update direction or update step size corresponding to the aggregation results when the candidate corrected wind speed vector fields do not meet the preset physical consistency conditions; the model broadcasting module distributes candidate global model parameters that meet the preset physical consistency conditions or updated global model parameters obtained after constraint processing to each local computing node. Figure 1 The example uses local nodes A, B, C, and D. In actual deployment, the number of local computing nodes can be determined based on the number of participating organizations.

[0040] During system initialization, the central server initializes the global model parameters θ0 and distributes θ0 to all participating local computing nodes. The global model is used for meteorological observation correction. In some optional embodiments, the global model can be a machine learning model or deep learning model for meteorological observation correction, including but not limited to fully connected neural networks, convolutional neural networks, recurrent neural networks, long short-term memory networks, Transformer models, or hybrid network structures based on spatiotemporal feature modeling. In the wind speed observation correction scenario, the input to the global model for meteorological observation correction includes private meteorological observation data and numerical weather prediction background fields. The private meteorological observation data may include original wind speed observation values, time information, spatial location information, and other meteorological features related to wind speed observation. The global model for meteorological observation correction obtains the meteorological observation deviation correction amount based on the private meteorological observation data and the numerical weather prediction background field, and then combines the meteorological observation deviation correction amount with the numerical weather prediction background field to obtain the corrected wind speed vector field. After receiving the initial global model parameters, each local computing node forms a local copy of the correction model parameters consistent with those of the central server.

[0041] In the During each iteration, the central server distributes the current global model parameters to the local computing nodes participating in the current training round. . No. Local computing nodes receive Then, based on its own private meteorological observation data and the corresponding numerical weather prediction background field, a meteorological observation correction global model loaded with the current global model parameters is input to obtain the meteorological observation deviation correction amount. This meteorological observation deviation correction amount is then combined with the numerical weather prediction background field to obtain the corrected wind speed vector field. Each local computing node performs local training based on the corrected wind speed vector field. Optionally, the first... Local loss function of each local computing node It can be represented as:

[0042] in, Corresponding to the k A private data distribution on each local computing node that does not need to be shared with a central server. The corresponding data loss term, formed from private meteorological observation data, is used to characterize the deviation between the corrected wind speed vector field and the local real meteorological observation results, reference values, or operational objectives. The corresponding physical regularization term is used to constrain the corrected wind speed vector field to meet the physical consistency requirements. This is the weight parameter for the physical regularization term, used to adjust the influence ratio between the data loss term and the physical regularization term. The input meteorological samples include private meteorological observation data and numerical weather prediction background. For the corresponding observation label value or reference value, For data loss items, It is the input meteorological sample The parameters are The meteorological observation bias correction amount output by the global model is used to correct the meteorological observations. This serves as the background field for numerical weather prediction. This is the corrected wind speed vector field. For the first The model parameters used by each local computing node when performing local training with the current global model parameters as initial values. This represents the expectation operation. The above expected loss term is expressed as follows: In the first... Under the private sample distribution of each local computing node, the expected value of the single-sample data loss for the corrected wind speed vector field and its corresponding local real meteorological observation results, reference values, or operational targets is used to measure the average fitting error of the global meteorological observation correction model of that node to its private sample population. The single-sample data loss can be the mean squared error, mean absolute error, or other loss functions suitable for the corresponding meteorological correction task.

[0043] In an optional embodiment, to facilitate theoretical analysis of convergence, it can be assumed that the local loss function satisfies the L-smoothness condition and the μ-strong convexity property. Based on this condition, the local computation nodes calculate the local gradient according to the local loss function. :

[0044] in, This represents the gradient operator. Because... Includes physical regularization terms and local gradients. The update direction is not only driven by local private meteorological observation data, but also constrained by physical regularization terms, thereby reducing the risk of the model learning correction relationships that violate physical laws.

[0045] The physical regularization term is constructed based on the residuals of the fluid dynamics equations of the corrected wind speed vector field, and is used to constrain the corrected wind speed vector field to satisfy meteorological physical conservation laws. Optionally, the fluid dynamics equation residuals may include at least one of the residuals of the Navier-Stokes equations, the continuity equations, the energy conservation equations, and the momentum conservation equations. In wind speed observation correction tasks, the physical regularization term can adopt the fluid dynamics equation residuals used to characterize the corresponding physical conservation laws.

[0046] As an example, the physical regularization term This can be specified as the residuals of the Navier-Stokes equations:

[0047] in, These are weighting coefficients used to adjust the weights of the residuals in the fluid dynamics equations within the physical regularization term. Velocity in a meteorological field or fluid state, measured in m / s. Time, in seconds. Kinematic viscosity, in m² / s. Density, in kg / m³ This refers to atmospheric pressure, measured in Pa. The gradient operator is in units of m. -1 , The Laplace operator, in units of m. -2 The gradient operator and the Laplace operator are both used to characterize the first-order and second-order differential relations in space; the above m -1 and m -2 The dimensional factors used to represent the spatial differentials are not operators that have independent physical units themselves. The value represents the square of the L2 norm of the Navier-Stokes equation residual vector within the parentheses, i.e., the sum of squares of each residual component. This value characterizes the degree of deviation of the model output from the equation; the smaller the value, the higher the physical consistency of the model output. Density and kinematic viscosity can be preset constants, empirical parameters, numerical weather prediction background field parameters, or parameters dynamically determined based on meteorological conditions such as temperature, air pressure, and humidity. The corrected wind speed vector field can be used to calculate the spatial and temporal gradients using finite difference, automatic differentiation, or spatial discretization methods to form the fluid dynamics equation residuals. Furthermore, the physical regularization term can be constructed based on the continuity equation, energy conservation equation, momentum conservation equation, or a combination thereof.

[0048] After obtaining the local gradient Then, the local compute node first prunes the local gradient according to a preset gradient pruning threshold. After gradient pruning, the noise scale σ of the Gaussian noise is determined based on the gradient pruning threshold, the preset privacy budget ε, and the preset privacy relaxation parameter δ. The preset privacy budget ε characterizes the strength of differential privacy protection, and the preset privacy relaxation parameter δ characterizes the preset probability relaxation amount allowed by the differential privacy mechanism. The noise scale σ can be determined based on the adopted Gaussian differential privacy mechanism, the gradient pruning threshold, the preset privacy budget ε, and the preset privacy relaxation parameter δ. The local compute node adds Gaussian noise to the pruned local gradient according to the noise scale, obtaining the... k The noisy gradient corresponding to each local computing node:

[0049] in, For the first k The noisy gradient is formed by the local computing nodes. With a mean of 0 and a variance of Gaussian noise. When the local gradient is a vector, the Gaussian noise can be represented by a covariance matrix of the corresponding dimension. By first clipping the local gradient and then adding Gaussian noise according to the noise scale, the impact of a single local model update on the uploaded gradient can be limited, and the uploaded gradient can be randomly perturbed, thereby reducing the possibility of attackers using gradient information to deduce the original private meteorological observation data. Given a sampling rate and T Given a certain number of iterations, a privacy loss accounting method can be used to calculate the actual cumulative privacy loss and determine whether it meets the privacy constraints corresponding to the preset privacy budget ε and preset privacy relaxation parameter δ. When the noise scale matches the gradient pruning threshold, sampling rate, number of iterations, preset privacy budget ε, and preset privacy relaxation parameter δ, a quantifiable privacy protection boundary can be provided for the gradient uploading process. Optionally, local computing nodes can also upload noisy gradients via an encrypted upload module or a secure communication module. Figure 1 The secure communication module can be used to implement transmission protection for communication encryption, differential privacy, or a combination of both.

[0050] After receiving noisy gradients uploaded by multiple local computing nodes, the central server uses a secure aggregation mechanism to aggregate the multiple noisy gradients, obtaining the aggregated result. This secure aggregation mechanism allows the central server to obtain the aggregation result without exposing the plaintext of each local computing node's individual noisy gradient. During this secure aggregation process, multiple noisy gradients can be aggregated based on a federated averaging algorithm framework. As an example, the central server can calculate candidate global model parameters based on the aggregation result in the following manner:

[0051] in, These are the global model parameters at the t-th iteration. These are the candidate global model parameters calculated based on the aggregation results; is the learning rate. K is the number of local computing nodes participating in the aggregation. This represents the aggregate weight corresponding to the k-th local computing node. This is the noisy gradient uploaded by the k-th local computing node. The central server does not need to obtain the original private meteorological observation data from each local computing node during aggregation, nor does it need to directly access the local data distribution of each local computing node, thus meeting the privacy protection requirements in cross-institutional meteorological observation correction scenarios.

[0052] After the central server calculates the candidate global model parameters according to the above-mentioned federated averaging algorithm framework based on the aggregation results, it does not directly distribute the candidate global model parameters to each local computing node. Instead, it obtains a candidate corrected wind speed vector field based on the candidate global model parameters and calculates the hydrodynamic equation residuals of the candidate corrected wind speed vector field to perform a physical consistency check on the candidate corrected wind speed vector field. Specifically, the central server can configure a numerical weather forecast background field and corresponding meteorological input data for physical consistency checks, obtain meteorological observation bias corrections based on the candidate global model parameters, and then combine them with the corresponding numerical weather forecast background field to obtain the candidate corrected wind speed vector field. The central server determines whether the candidate corrected wind speed vector field meets the preset physical consistency conditions based on the hydrodynamic equation residuals of the candidate corrected wind speed vector field. When the candidate corrected wind speed vector field meets the preset physical consistency conditions, the central server uses the candidate global model parameters as the global model parameters to be distributed in this round. When the candidate corrected wind speed vector field does not meet the preset physical consistency condition, the central server constrains the update direction or update step size corresponding to the aggregation result according to the physical conservation law, so that the update direction of the global model parameters after constraint processing is consistent with the requirements of mass conservation, energy conservation, or momentum conservation, thus obtaining updated global model parameters. The central server distributes the candidate global model parameters that meet the preset physical consistency condition or the updated global model parameters obtained after constraint processing to each local computing node through the model broadcast module. Each local computing node updates its local correction model according to the received global model parameters and enters the next iteration. Thus, the system forms a cyclical process of the central server distributing the current global model parameters, local computing nodes calculating local gradients using private meteorological observation data and physical regularization terms, local computing nodes adding noise and uploading the noisy gradients, the central server aggregating the noisy gradients and updating the global model parameters, and the central server distributing the updated model parameters again.

[0053] Furthermore, the theoretical properties of the system can be explained through convergence analysis of the loss function and physical constraints in gradient updates. As a theoretical analysis scenario, when the local objective function of each local computation node satisfies the L-smoothness condition and μ-strong convexity, and satisfies assumptions such as appropriate learning rate, bounded variance of gradient estimation, and aggregation weights, based on convex optimization theory, the global convergence error can satisfy the following boundary relationship:

[0054] in, Corresponding to the The global objective function after rounds of iteration, The corresponding global objective function under the initial global model parameters, The optimal value corresponding to the global objective function, Corresponding smoothness constant, Corresponding to strong convex parameters, The variance term corresponds to gradient perturbation or randomness. For the number of iteration rounds, To represent taking the square of the 2-norm of the Navier-Stokes equation residual vector within the parentheses, Let this represent the gradient operator. From the above boundary relationships, it can be seen that, under the conditions of L-smoothness, μ-strong convexity, and the corresponding federated optimization assumptions, the gradient operator increases with the number of iterations. As it increases, the upper bound of the error on the right side increases. As the gradient norm decreases, it can approach zero. For finite-round iterations and noisy upload scenarios, the global model parameters can converge to the global optimum. Within the neighborhood of [the data type]. This boundary relationship is used to represent the convergence trend only when the above assumptions hold, and can provide a corresponding theoretical basis for meteorological data scenarios with spatiotemporal correlation and non-independent identical distribution characteristics.

[0055] In this application, physical constraints operate at two levels: local computing nodes and the central server. On one hand, multiple local computing nodes construct physical regularization terms based on the residuals of the hydrodynamic equations of the corrected wind speed vector field, ensuring that the local gradient is constrained by meteorological physical conservation laws during its formation phase. On the other hand, the central server, after obtaining the aggregation result using a secure aggregation mechanism, generates candidate global model parameters based on the aggregation result and obtains candidate corrected wind speed vector fields based on these parameters. Physical consistency checks are then performed by calculating the residuals of the hydrodynamic equations of these candidate corrected wind speed vector fields. When the candidate corrected wind speed vector fields do not meet the preset physical consistency conditions, the update direction or update step size corresponding to the aggregation result is constrained according to physical conservation laws. Thus, local physical regularization reduces the possibility of local computing nodes forming update directions that violate physical laws, while the physical consistency check and constraint processing on the central server side further reduces differential privacy noise and the possibility of global model updates deviating from meteorological physical laws after the aggregation of updates from multiple local computing nodes.

[0056] Through the above methods, this application enables the joint training of a global model for meteorological observation correction using private meteorological observation data from multiple local computing nodes without centralized storage and processing of raw meteorological observation data. The data loss term is used to learn the deviation relationship between the corrected wind speed vector field and local real meteorological observation results, reference values, or operational objectives. The physical regularization term ensures that the local gradient is constrained by meteorological physical conservation laws. Gradient pruning and differential privacy Gaussian noise reduce the risk of privacy leakage during gradient upload. The secure aggregation mechanism allows the central server to complete multi-node update aggregation without exposing the plaintext of individually noisy gradients from each local computing node. Furthermore, by implementing physical consistency checks on candidate corrected wind speed vector fields and imposing necessary update direction or update step size constraints, the update of the aggregated global model is further constrained by meteorological physical conservation laws. Therefore, this application can balance data privacy protection, multi-source model collaborative updates, and meteorological physical consistency after aggregation updates in the process of cross-institutional meteorological observation correction.

[0057] Example 3 like Figure 2 As shown, this application also proposes a control method for the above-mentioned meteorological observation correction system, which may include: S1. Through multiple local computing nodes, the current global model parameters issued by the central server are received. The private meteorological observation data held by each of the multiple local computing nodes and the numerical weather prediction background field are input into the meteorological observation correction global model to obtain the meteorological observation bias correction amount. Then, the corrected wind speed vector field is obtained by combining it with the numerical weather prediction background field. The local gradient is calculated according to the local loss function including the data loss term and the physical regularization term. The physical regularization term is constructed based on the fluid dynamics equation residuals of the corrected wind speed vector field. The local gradient is clipped and Gaussian noise is added according to the differential privacy mechanism to obtain multiple noisy gradients. S2, through a central server, multiple noisy gradients are aggregated using a secure aggregation mechanism to obtain an aggregation result; candidate global model parameters are generated based on the aggregation result, and candidate corrected wind speed vector fields are obtained based on the candidate global model parameters; the hydrodynamic equation residuals of the candidate corrected wind speed vector fields are calculated and physical consistency checks are performed; when the candidate corrected wind speed vector fields do not meet the preset physical consistency conditions, the update direction or update step size corresponding to the aggregation result is constrained according to the physical conservation law to obtain updated global model parameters; the candidate global model parameters that meet the preset physical consistency conditions or the updated global model parameters obtained by the constraint processing are distributed to multiple local computing nodes.

[0058] Example 4 In one specific application example of this application, the meteorological observation correction system is applied to a cross-regional wind farm wind speed observation correction and wind power prediction scenario. Multiple wind farms serve as different local computing nodes, each storing its own historical wind speed observation data, wind direction observation data, temperature observation data, and corresponding numerical weather prediction background field data for the relevant time period. This data can originate from anemometer towers, nacelle sensors, ground meteorological stations, or numerical weather prediction systems. Because different wind farms belong to different operating entities, the original meteorological observation data may involve commercial operation information, power production information, or regional meteorologically sensitive information; therefore, different wind farms do not directly share the original meteorological observation data.

[0059] The central server first initializes a global model for wind speed observation correction and distributes the current global model parameters to the local computing nodes corresponding to multiple wind farms. Each local computing node inputs local historical meteorological observation data and the corresponding time period's numerical weather prediction background field into the global model for meteorological observation correction. This model learns the deviation relationship between the numerical weather prediction background field and the actual observed wind field, obtaining a meteorological observation deviation correction. This correction is then combined with the numerical weather prediction background field to obtain the corrected wind speed vector field. The data loss term in the local loss function characterizes the deviation between the corrected wind speed vector field and the local actual meteorological observation results, while the physical regularization term constrains the corrected wind speed vector field to satisfy fluid dynamics laws.

[0060] In some embodiments, the corrected wind speed vector field may include horizontal and vertical wind speed components, and the physical regularization term is constructed based on the Navier-Stokes equation residuals of the corrected wind speed vector field. The velocity and pressure gradients of the corrected wind speed vector field are calculated using a spatial discretization method to form hydrodynamic residuals, thereby reducing the possibility of local non-physical abrupt changes, anomalous vortices, or energy non-conservation in the corrected wind speed vector field.

[0061] After completing the local gradient calculation, each local computing node clips the local gradient according to a preset gradient clipping threshold. Based on the gradient clipping threshold, the preset privacy budget ε, and the preset privacy relaxation parameter δ, the noise scale of Gaussian noise is determined, and Gaussian noise is added to the clipped local gradient according to the noise scale to form a noisy gradient. The noisy gradient is then uploaded to the central server through the secure communication module.

[0062] The central server employs a secure aggregation mechanism to aggregate noisy gradients uploaded from multiple wind farms, and obtains the aggregation result based on a federated averaging algorithm framework. Candidate global model parameters are then generated based on these aggregation results. The central server uses these candidate global model parameters to obtain candidate corrected wind speed vector fields and calculates the hydrodynamic equation residuals of these fields for physical consistency checks. When a candidate corrected wind speed vector field meets a preset physical consistency condition, the candidate global model parameters are distributed to the local computing nodes corresponding to each wind farm. When a candidate corrected wind speed vector field does not meet the preset physical consistency condition, the update direction or update step size corresponding to the aggregation result is constrained according to physical conservation laws, and the updated global model parameters obtained after constraint processing are distributed to the local computing nodes corresponding to each wind farm. Each local computing node updates its local correction model accordingly and enters the next round of federated training iterations. After multiple iterations, the global model can integrate the wind speed deviation characteristics of wind farms in different regions, learning multi-regional wind speed correction patterns without sharing original meteorological observation data. The corrected wind speed vector field can be further input into the wind power prediction model to improve the accuracy of short-term wind power prediction and the stability of grid dispatch.

[0063] For example, under typhoon, strong wind, or complex terrain conditions, different wind farms may have significantly different wind speed deviation characteristics. This application improves the model's adaptability to complex weather processes and regional differences by collaboratively training a global meteorological observation correction model through multiple local computing nodes. It reduces the risk of local model updates violating atmospheric dynamics through physical regularization terms, reduces the risk of data leakage during gradient upload through gradient pruning and differential privacy mechanisms, and reduces the possibility of global model updates deviating from meteorological physical conservation laws by using a central server to perform physical consistency checks and necessary update direction or update step size constraints on candidate correction wind speed vector fields.

[0064] Example 5 This application also provides an electronic device, which may include one or more processors, a memory, and a communication interface.

[0065] The memory, communication interface, and processor are coupled together. For example, the memory, communication interface, and processor can be coupled together via a bus.

[0066] The communication interface is used for data transmission with other devices. The memory stores computer program code. This computer program code includes computer instructions, which, when executed by the processor, cause the electronic device to perform the steps of the control method for the aforementioned meteorological observation and correction system.

[0067] The processor can be a central processing unit (CPU) or a controller, such as a CPU, a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The processor can be used to support an electronic device in performing the method steps provided in the above embodiments.

[0068] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. These buses can be categorized as address buses, data buses, control buses, etc.

[0069] Example 6 This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the control method for the meteorological observation correction system described above.

[0070] The computer-readable storage media involved in this application include random access memory (RAM), read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage media known in the art.

[0071] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A meteorological observation correction system, characterized in that, Includes a central server and multiple local computing nodes; Multiple local computing nodes are used to receive the current global model parameters issued by the central server, input the private meteorological observation data held by each local computing node and the numerical weather prediction background field into the meteorological observation correction global model to obtain the meteorological observation bias correction amount, and then combine it with the numerical weather prediction background field to obtain the corrected wind speed vector field; calculate the local gradient according to the local loss function including the data loss term and the physical regularization term, wherein the physical regularization term is constructed based on the fluid dynamics equation residuals of the corrected wind speed vector field; prune the local gradient and add Gaussian noise according to the differential privacy mechanism to obtain multiple noisy gradients; The central server is used to aggregate multiple noisy gradients using a secure aggregation mechanism to obtain an aggregation result; generate candidate global model parameters based on the aggregation result, and obtain a candidate corrected wind speed vector field based on the candidate global model parameters; calculate the hydrodynamic equation residuals of the candidate corrected wind speed vector field and perform a physical consistency check; when the candidate corrected wind speed vector field does not meet the preset physical consistency condition, constrain the update direction or update step size corresponding to the aggregation result according to the physical conservation law to obtain updated global model parameters; and distribute the candidate global model parameters that meet the preset physical consistency condition or the updated global model parameters obtained by the constraint processing to multiple local computing nodes.

2. The meteorological observation correction system according to claim 1, characterized in that, The data loss term is used to characterize the deviation between the corrected wind speed vector field and the local real meteorological observation results, reference values, or operational objectives. The physical regularization term and the data loss term together constitute the local loss function according to preset weights.

3. The meteorological observation correction system according to claim 1, characterized in that, The local loss function satisfies the L-smoothness condition and the μ-strong convexity property.

4. The meteorological observation correction system according to claim 1, characterized in that, The fluid dynamics equation residuals include at least one of the Navier-Stokes equation residuals, continuity equation residuals, energy conservation equation residuals, and momentum conservation equation residuals.

5. A meteorological observation correction system according to claim 4, characterized in that, In the wind field correction scenario, the residuals of the fluid dynamics equations are the Navier-Stokes equation residuals, and the physical regularization term... satisfy: in, These are weighting coefficients used to adjust the weights of the residuals in the fluid dynamics equations within the physical regularization term. Velocity in a meteorological field or fluid state. For time, For kinematic viscosity, For density, For air pressure, For gradient operators, For the Laplace operator, This indicates taking the square of the second norm of the Navier-Stokes equation residual vector within the parentheses.

6. A meteorological observation correction system according to claim 1, characterized in that, The local gradient is pruned and Gaussian noise is added according to the differential privacy mechanism, including: The local gradient corresponding to each local computing node is clipped according to a preset gradient clipping threshold. The noise scale of Gaussian noise is determined based on the gradient clipping threshold, the preset privacy budget ε, and the preset privacy relaxation parameter δ. Gaussian noise is added to the clipped local gradients according to the noise scale to obtain the noisy gradient corresponding to each local computing node.

7. A meteorological observation correction system according to claim 1, characterized in that, The method for aggregating multiple noisy gradients using a secure aggregation mechanism includes: the central server aggregating the multiple noisy gradients based on a federated averaging algorithm framework.

8. A control method for a meteorological observation correction system as described in any one of claims 1 to 7, characterized in that, include: Multiple local computing nodes receive the current global model parameters issued by the central server. The private meteorological observation data held by each local computing node, along with the numerical weather prediction background field, are input into the global meteorological observation correction model to obtain the meteorological observation bias correction amount. This is then combined with the numerical weather prediction background field to obtain the corrected wind speed vector field. Local gradients are calculated using a local loss function that includes a data loss term and a physical regularization term. The physical regularization term is constructed based on the residuals of the fluid dynamics equations of the corrected wind speed vector field. The local gradients are then pruned and Gaussian noise is added according to a differential privacy mechanism to obtain multiple noisy gradients. Multiple noisy gradients are aggregated using a secure aggregation mechanism through a central server to obtain an aggregation result. Candidate global model parameters are generated based on the aggregation result, and candidate corrected wind speed vector fields are obtained based on the candidate global model parameters. The hydrodynamic equation residuals of the candidate corrected wind speed vector fields are calculated and physical consistency checks are performed. When the candidate corrected wind speed vector fields do not meet the preset physical consistency conditions, the update direction or update step size corresponding to the aggregation result is constrained according to the physical conservation law to obtain updated global model parameters. The candidate global model parameters that meet the preset physical consistency conditions or the updated global model parameters obtained by the constraint processing are distributed to multiple local computing nodes.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the control method for the meteorological observation correction system as described in claim 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the control method for the meteorological observation correction system as described in claim 8.

Images