A method and system for correcting the bias of forecast data

Abstract

This application discloses a method and system for correcting forecast data bias. The method acquires initial forecast data for a target area; extracts land-sea boundary features from the initial forecast data to obtain the coastal mask and shoreline mask of the target area; inputs the initial forecast data and the shoreline mask of the target area into a bias correction model to obtain the corrected forecast data output by the bias correction model. It can efficiently extract the shoreline mask, and while retaining the effective guiding information of the forecast data by combining a non-complete noise addition strategy, it adds some noise and optimizes features on the initial forecast data, so that the model can be trained by reanalyzing the data. Furthermore, by combining feature extraction and dynamic parameter optimization, it effectively solves the problems of large errors in three-dimensional sea surface temperature forecasts in nearshore waters, long-term error accumulation, and miscorrection caused by land-sea interference, and significantly improves the model's adaptability to complex small-scale environmental factors.

Description

Technical Field

[0001] This application relates to the field of meteorological and oceanographic forecast data processing technology, and in particular to a method and system for correcting forecast data bias. Background Technology

[0002] Currently, large-scale intelligent forecasting models (such as the "Xihe" marine environmental forecasting system) have demonstrated high efficiency in global sea surface temperature (SST) prediction, but significant errors still exist in 3D SST forecasts for nearshore waters. This is because nearshore areas are affected by complex small-scale environmental factors (such as land-sea interaction and abrupt topographic changes), while the training data for existing models is concentrated in the open ocean, resulting in insufficient accuracy in nearshore forecasts. To improve forecast reliability, bias correction processing is needed on the original forecast results.

[0003] Existing methods (such as deterministic models like ConvLSTM and U-Net) achieve correction by optimizing the distance from the true value, but they cannot model the chaotic characteristics of the sea surface temperature system, leading to the accumulation of long-term forecast errors and the gradual blurring of correction results over time. Furthermore, model training requires both historical reanalysis data and sufficient forecast product data as input-label pairs, but the storage of historical forecast data is insufficient or the cost of regenerating it is extremely high, severely limiting the widespread application of these models. In addition, for strip-shaped coastal areas at the land-sea boundary, it is difficult to distinguish between land and sea interference, leading to errors exacerbated by incorrect corrections, resulting in low forecast accuracy. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] The main objective of this disclosure is to propose a method and system for correcting forecast data deviations, which can efficiently extract coastal masks and improve the accuracy and efficiency of forecast data correction by retaining effective guidance information of forecast data and combining feature extraction and dynamic parameter optimization.

[0006] A first aspect of this application provides a method for correcting deviations in forecast data, used in a central controller, the method comprising:

[0007] Obtain initial forecast data for the target area;

[0008] Extract the land-sea boundary features of the target area from the initial forecast data of the target area to obtain the coast mask and the coastal mask of the target area;

[0009] The initial forecast data and the coastal mask of the target area are input into the deviation correction model to obtain the corrected forecast data output by the deviation correction model.

[0010] The correction process of the deviation correction model includes:

[0011] The initial forecast data is partially noise-added to obtain the first forecast data;

[0012] Based on the coast mask and the coastal mask, the first forecast data feature value is extracted from the first forecast data;

[0013] Based on the aforementioned bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data.

[0014] In some embodiments of this application, the training process of the bias correction model includes:

[0015] Preset initial deviation correction model;

[0016] Obtain historical reanalysis data for the target area;

[0017] The initial bias correction model is trained based on the historical reanalysis data to obtain the first bias correction model;

[0018] The mean squared error loss is calculated based on the model output of the first deviation correction model and the historical reanalysis data.

[0019] The first deviation correction model is optimized based on the mean squared error loss to obtain the deviation correction model.

[0020] In some embodiments of this application, the step of extracting the land-sea boundary features of the target area from the initial forecast data of the target area to obtain the coast mask and the coastal mask of the target area includes:

[0021] The initial forecast data for the target area is filtered for extreme values ​​to obtain the land and sea masking code for the target area;

[0022] The land and sea mask is subjected to morphological erosion to obtain the processed mask;

[0023] The coastal mask is obtained by performing a set difference operation between the processed mask and the land and sea mask.

[0024] In some embodiments of this application, the step of adding partial noise to the initial forecast data to obtain the first forecast data includes:

[0025] A noise intensity sequence is generated according to a preset noise intensity scheduler, the noise intensity sequence containing multi-level decreasing noise values ​​from the maximum noise intensity to the minimum noise intensity;

[0026] Find the noise intensity index position in the noise intensity sequence that is closest to the preset target noise intensity value;

[0027] Starting from the index position, a noise signal of corresponding intensity is added to the initial forecast data to generate the first forecast data.

[0028] In some embodiments of this application, the step of extracting first forecast data feature values ​​from the first forecast data based on the coastline mask and the coastal mask includes:

[0029] Based on the coastline mask, the coastal mask, and the first forecast data, ocean area data, coastal area data, and coastal strip area data are generated.

[0030] The ocean area data and the coastal area data are subjected to global mean normalization to generate a first feature value;

[0031] Extract the corresponding sea surface temperature gradient and spatiotemporal rate of change from the data of the coastal strip region;

[0032] The sea surface temperature gradient and the spatiotemporal rate of change are used as the second characteristic value;

[0033] The first feature value and the second feature value are fused to obtain the first forecast data feature value.

[0034] In some embodiments of this application, generating marine area data, coastal area data, and coastal strip area data based on the coastline mask, the coastal mask, and the first forecast data includes:

[0035] The first forecast data is multiplied element-wise with the coastline mask to generate ocean area data.

[0036] The first forecast data is multiplied element-wise with the coastal mask to generate coastal area data.

[0037] The set difference operation is performed on the coast mask and the coastal mask to extract the boundary position of the coastal strip area in the target area from the first forecast data, and obtain the corresponding coastal strip area data.

[0038] In some embodiments of this application, the step of optimizing the first forecast data based on the bias correction model and the feature values ​​of the first forecast data to obtain the bias-corrected forecast data includes:

[0039] The feature values ​​of the first forecast data are input into the encoder part of the bias correction model to generate a feature embedding vector;

[0040] Calculate the gradient change of the feature embedding vector;

[0041] Based on the gradient change, adjust the noise intensity value and the number of denoising iterations of the deviation correction model during the data optimization process;

[0042] Based on the noise intensity value and the number of denoising iterations, the first forecast data is optimized until the gradient change meets the preset convergence condition, at which point the corrected forecast data is output.

[0043] To achieve the above objectives, a second aspect of the present invention provides a forecast data deviation correction system, the system comprising:

[0044] The acquisition module is used to acquire initial forecast data for the target area;

[0045] The extraction module is used to extract the land-sea boundary features of the target area from the initial forecast data of the target area, and obtain the coast mask and the coastal mask of the target area;

[0046] The correction module is used to input the initial forecast data and the coastal mask of the target area into the deviation correction model to obtain the corrected forecast data output by the deviation correction model.

[0047] The correction process of the deviation correction model includes:

[0048] The initial forecast data is partially noise-added to obtain the first forecast data;

[0049] Based on the coast mask and the coastal mask, the first forecast data feature value is extracted from the first forecast data;

[0050] Based on the aforementioned bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data.

[0051] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform the above-described method for correcting deviations in forecast data.

[0052] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the above-described method for correcting deviations in forecast data.

[0053] This application provides a method for correcting the bias of forecast data. It involves acquiring initial forecast data for a target area; extracting the land-sea boundary features of the target area from the initial forecast data to obtain the coastal mask and the coastal mask of the target area; inputting the initial forecast data and the coastal mask of the target area into a bias correction model to obtain the corrected forecast data output by the bias correction model. This method can efficiently extract the coastal mask and, by combining a non-complete noise-addition strategy to retain the effective guiding information of the forecast data, adds some noise and optimizes the features of the initial forecast data, allowing the model to be trained simply by reanalyzing the data. Furthermore, by combining feature extraction and dynamic parameter optimization, this method effectively solves the problems of large errors in nearshore three-dimensional sea surface temperature forecasts, long-term error accumulation, and miscorrection caused by land-sea interference, and significantly improves the model's adaptability to complex small-scale environmental factors.

[0054] It is understood that the beneficial effects of the second to fourth aspects compared with the related technologies are the same as the beneficial effects of the first aspect compared with the related technologies. Please refer to the relevant description in the first aspect above, which will not be repeated here. Attached Figure Description

[0055] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0056] Figure 1 This is a flowchart illustrating a method for correcting deviations in forecast data provided in an embodiment of this application;

[0057] Figure 2 This is a comparative schematic diagram of the acyclic neural network method provided in the embodiments of this application;

[0058] Figure 3 This is a comparative schematic diagram of the recurrent neural network method provided in the embodiments of this application;

[0059] Figure 4 This is a schematic diagram of the structure of a prediction data deviation correction training system provided in an embodiment of this application;

[0060] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0061] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0062] In the description of this application, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of the technical features indicated.

[0063] In the description of this application, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0064] In the description of this application, it should be noted that, unless otherwise explicitly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0065] Sea surface temperature (SST) has a significant impact on human production and daily life. In terms of economy and ecology, accurate nearshore sea surface temperature forecasts provide crucial and effective data support for nearshore aquaculture, offshore fishing ground discovery, and nearshore ecological monitoring. This is of great importance for ensuring nearshore economic development and addressing ecological issues such as red tides. Slow changes in SST also play a significant potential role in the spatiotemporal series prediction of climate variables on short- to medium-term and longer-term scales.

[0066] In recent years, with the development of intelligent technologies, large-scale intelligent weather forecasting models and marine environmental intelligent forecasting models such as "Xihe" have demonstrated performance comparable to or even surpassing traditional numerical weather prediction operational models by mining valuable spatiotemporal patterns from massive amounts of historical meteorological and oceanographic data, while significantly improving forecasting speed. However, due to the uneven spatial distribution of historical marine environmental data, large-scale marine forecasting models such as "Xihe" focus more on discovering spatiotemporal patterns in the relatively large amounts of data from offshore areas. Therefore, in the three-dimensional sea surface temperature forecasting of nearshore areas, compared to offshore areas, the forecasting error usually increases further due to the relatively smaller data volume and the more frequent and complex small-scale changes in nearshore areas, and the forecasting effect urgently needs further improvement. However, there is still relatively little research on the bias correction of large-scale intelligent forecasting models, and the following two major problems remain to be further addressed:

[0067] (1) How to reasonably address the dependence of data-driven bias correction models on forecast products:

[0068] For traditional data-driven bias correction models, while obtaining relevant reanalysis products or observational data as training sample labels, these methods typically rely on sufficient forecast product data as input for training to achieve good correction results. However, in practical applications, relevant operational units may not store or back up sufficient historical forecast product data, and recreating forecast product data that meets the training requirements also requires significant time and hardware costs. Therefore, this reliance on forecast product data limits the widespread application of these models, and this problem urgently needs further resolution.

[0069] (2) How to develop a bias correction model for intelligent large-scale model forecast products:

[0070] For the sea surface temperature (SST) forecast results of large-scale intelligent forecasting models such as "Xihe" in nearshore areas, the error generation mechanism and distribution characteristics are usually different from those of traditional numerical forecasting models. How to develop a bias correction model for the forecast products of large-scale intelligent models and achieve efficient bias correction of the SST forecast results in nearshore areas still needs further research.

[0071] Based on this, embodiments of this application provide a method, system, electronic device, and medium for correcting deviations in forecast data, aiming to efficiently extract coastal masks and improve the accuracy and efficiency of forecast data correction by retaining effective guidance information of forecast data and combining feature extraction and dynamic parameter optimization.

[0072] The forecast data deviation correction method, system, electronic device and medium provided in the embodiments of this application are specifically described through the following embodiments. First, the forecast data deviation correction method in the embodiments of this application is described.

[0073] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0074] Foundational technologies in artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0075] The forecast data deviation correction method provided in this application relates to the field of meteorological and oceanographic forecast data processing technology. The forecast data deviation correction method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the forecast data deviation correction method, but is not limited to the above forms.

[0076] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0077] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0078] Therefore, referring to Figure 1This application provides a method for correcting the deviation of forecast data. This method is applied to a central controller, which can be a server, an electronic device, or a mobile terminal, etc., without specific limitations. The method includes the following steps S110 to S130:

[0079] Step S110: Obtain initial forecast data for the target area;

[0080] Step S120: Extract the land-sea boundary features of the target area from the initial forecast data of the target area to obtain the coast mask and coastal mask of the target area;

[0081] Step S130: Input the initial forecast data and the coastal mask of the target area into the bias correction model to obtain the corrected forecast data output by the bias correction model;

[0082] The correction process of the deviation correction model includes:

[0083] The initial forecast data is partially noise-added to obtain the first forecast data;

[0084] Based on the coastline mask and the coastal mask, the feature values ​​of the first forecast data are extracted from the first forecast data;

[0085] Based on the bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data.

[0086] In this step, the initial forecast data for the target area is first obtained. The initial forecast data refers to the uncorrected original prediction results for the target area, which can be achieved using gridded field data output by the "Xihe" intelligent forecasting model. Then, the land-sea boundary features of the target area are extracted from the initial forecast data to obtain the coastline mask and the coastal mask. By analyzing the geographic information in the initial forecast data, the features at the land-sea boundary are identified, and two different masks are generated to represent the coastline and the coastal area.

[0087] Specifically, land-sea boundary features refer to the spatial distribution characteristics at the junction of land and sea. They can be achieved by extracting the coastline contour through binarization and morphological operations, and are used to distinguish the physical boundaries between land and sea. Coastal mask refers to the spatial location matrix that identifies marine areas. It can be generated by binary segmenting the initial data using a land-sea classification threshold, and is used to isolate data processing of marine areas. Coastal mask refers to the strip-shaped region identifier of the land-sea transition zone. It can be generated through morphological erosion operations and set difference operations, and is used to capture sensitive areas affected by land-sea interaction.

[0088] Furthermore, the initial forecast data and coastal mask are input into the bias correction model. The trained bias correction model, combined with the extracted feature values, is then used to optimize the noisy forecast data, thereby obtaining the final correction result.

[0089] Specifically, the correction process of the bias correction model includes: firstly, partial noise addition processing is performed on the initial forecast data to obtain the first forecast data. The partial noise addition processing refers to superimposing a controllable intensity interference signal at a specific spatial location. Specifically, it can be implemented by using a Gaussian noise generator combined with an intensity scheduling strategy to enhance the robustness of the model to data disturbances. In this way, by introducing a certain degree of random noise into the original data, the uncertainty that may exist in the forecast can be simulated.

[0090] Furthermore, based on the coast mask and the coastal mask, the first forecast data feature value is extracted from the first forecast data. The first forecast data feature value refers to the quantitative indicator that reflects the spatial distribution and dynamic changes of the data. Specifically, it can be generated by regional segmentation, gradient calculation and spatiotemporal change rate statistics fusion, and is used to characterize the thermal anomaly characteristics of the nearshore sea area.

[0091] Furthermore, based on the bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data. Data optimization refers to correcting the prediction bias by iteratively adjusting the model parameters. Specifically, the gradient descent algorithm combined with dynamic convergence conditions can be used to eliminate systematic errors and maintain the continuity of the physical field.

[0092] Specifically, the correction process based on feature extraction and optimization can better capture the influence of complex nearshore environmental factors, reducing forecast bias. Furthermore, the introduction of coastline and coastal masks enables the model to more accurately identify and process forecast data at the land-sea interface. Noise addition enhances the model's adaptability to different forecast scenarios, allowing the bias correction model to more effectively handle complex nearshore forecast situations, improving forecast accuracy. It is also suitable for forecasting nearshore areas affected by land-sea interaction and abrupt topographic changes, effectively overcoming the insufficient forecast accuracy of traditional methods in these regions. Simultaneously, by reducing reliance on historical forecast data, it also improves the deployment flexibility and computational efficiency of the forecast system.

[0093] The following explains the specific training steps of the bias correction model:

[0094] Step S210: Preset the initial deviation correction model;

[0095] Step S220: Obtain historical reanalysis data for the target area;

[0096] Step S230: Train the initial bias correction model based on historical reanalysis data to obtain the first bias correction model;

[0097] Step S240: Calculate the mean squared error loss based on the model output of the first deviation correction model and the historical reanalysis data;

[0098] Step S250: Optimize the first deviation correction model based on the mean square error loss to obtain the deviation correction model.

[0099] In this step, an initial bias correction model is preset. The initial bias correction model can adopt a deep learning-based network structure, such as a convolutional neural network or a residual network. Then, historical reanalysis data of the target area is obtained. The historical reanalysis data is a validated high-precision historical observation data, which is used to replace the forecast product data required for traditional training. The model training uses historical reanalysis data to replace the forecast data input in traditional methods, thus solving the problem of insufficient data.

[0100] Furthermore, an initial bias correction model is trained based on historical reanalysis data to obtain the first bias correction model. During training, historical reanalysis data is used as the model input and labels. The historical reanalysis data, as the label data, directly drives the adjustment of model parameters. The model parameters are optimized through the backpropagation algorithm. This eliminates the need to rely on historical forecast data to generate input-label pairs, and directly corrects errors on the initial forecast data. This reduces the dependence of model training on data storage and generation costs, and improves the applicability of the model.

[0101] Furthermore, the mean squared error loss is calculated based on the model output of the first bias correction model and the historical reanalysis data. The mean squared error loss is calculated by comparing the difference between the model output and the historical reanalysis data. As the basis for optimizing the model, the mean squared error loss reflects the magnitude of the error between the model prediction result and the true value.

[0102] Furthermore, the first bias correction model is optimized based on the mean squared error loss to obtain the trained bias correction model. The optimization process can employ algorithms such as gradient descent, continuously adjusting the model parameters through multiple iterations until the loss function converges. This gradually optimizes the model weights using the gradient descent method, enabling the model to better fit the real data distribution and thus improving the accuracy of the correction results.

[0103] In some embodiments, step S120 involves extracting the land-sea boundary features of the target area from the initial forecast data of the target area to obtain the coast mask and the coastal mask of the target area, including the following steps S310 to S340:

[0104] Step S310: Perform extreme value filtering on the initial forecast data of the target area to obtain the land and sea masking code of the target area;

[0105] Step S320: Perform morphological erosion on the land and sea mask to obtain the processed mask;

[0106] Step S330: Perform a set difference operation between the processed mask and the land and sea masks to obtain the coastal mask.

[0107] In this step, the initial forecast data of the target area is subjected to extreme value filtering to obtain the land-sea mask of the target area. Extreme value filtering distinguishes land and sea areas by setting a threshold, classifying pixels in the initial forecast data that are above or below the threshold as land or sea, and generating a binarized land-sea mask.

[0108] For example, in the extreme value filtering stage, a critical threshold for sea surface temperature is set, and areas in the initial forecast data that are above this value are marked as land, and areas that are below this value are marked as ocean, thus generating an initial land-sea mask.

[0109] Furthermore, the land and sea mask is subjected to morphological erosion processing to obtain the processed mask. The morphological erosion processing uses structuring elements to scan the land and sea mask pixel by pixel to eliminate small protrusions at the mask edges and reduce the mask coverage area. For example, the morphological erosion stage uses rectangular or circular structuring elements to shrink the preset pixel width inward along the mask edge to eliminate isolated points or burrs caused by data noise.

[0110] Furthermore, a set difference operation is performed between the processed mask and the land-sea mask to obtain the coastal mask. This set difference operation involves a logical XOR operation between the original land-sea mask and the eroded mask, preserving the boundary regions removed by the erosion to form the coastal mask. Ultimately, this achieves accurate extraction of the land-sea boundary features of the target area, obtaining both a coastal mask and a coastal mask, effectively distinguishing between ocean, land, and coastal regions. This provides precise spatial information for subsequent bias correction. Moreover, morphological processing and set operations can eliminate noise interference, improving the accuracy and robustness of boundary features. This provides a reliable spatial reference for correcting biases in three-dimensional sea surface temperature forecasts in nearshore waters, contributing to improved forecast accuracy.

[0111] In some embodiments, in step S130, the initial forecast data is partially noise-added to obtain the first forecast data, including the following steps S410 to S430:

[0112] Step S410: Generate a noise intensity sequence according to a preset noise intensity scheduler. The noise intensity sequence contains multi-level decreasing noise values ​​from the maximum noise intensity to the minimum noise intensity.

[0113] Step S420: Find the noise intensity index position in the noise intensity sequence that is closest to the preset target noise intensity value;

[0114] Step S430: Starting from the index position, add a noise signal of corresponding intensity to the initial forecast data to generate the first forecast data.

[0115] In this step, a noise intensity sequence is first generated based on a preset noise intensity scheduler. This sequence contains multiple levels of decreasing noise values, from maximum to minimum, with each level corresponding to a different interference intensity. Then, the noise intensity index position closest to the preset target noise intensity value is found within this sequence. Preferably, the index position is determined by calculating the difference between the current noise intensity and the preset target value, and the position with the smallest difference is selected as the starting point for noise addition.

[0116] Furthermore, starting from the selected index position, a noise signal of corresponding intensity is added to the initial forecast data to generate the first forecast data. The noise addition process is limited to the data region corresponding to the index position to avoid global noise interference. Thus, in the spatial dimension of the initial forecast data, noise is added only to the coastal strip region corresponding to the index position; for example, Gaussian white noise is superimposed at grid points covered by a coastal mask, and the noise amplitude is controlled by the intensity value at the current index position. Through hierarchical noise addition and local region control, controllable noise addition to the initial forecast data can be achieved. The model can introduce controllable interference while preserving the original data characteristics, enhancing its ability to correct errors in the land-sea boundary region.

[0117] In some embodiments, in step S130, the first forecast data features are extracted from the first forecast data based on the coastline mask and the coastal mask, including the following steps S510 to S550:

[0118] Step S510: Generate marine area data, coastal area data, and coastal strip area data based on the coastline mask, the coastal mask, and the first forecast data;

[0119] Step S520: Perform global mean normalization on the marine area data and the coastal area data to generate the first feature value;

[0120] Step S530: Extract the corresponding sea surface temperature gradient and spatiotemporal rate of change from the coastal strip area data;

[0121] Step S540: Use the sea surface temperature gradient and the spatiotemporal rate of change as the second characteristic value;

[0122] Step S550: Fuse the first feature value and the second feature value to obtain the first forecast data feature value.

[0123] In this step, ocean region data, coastal region data, and coastal strip region data are generated based on the coastline mask, the coastal mask, and the first forecast data. Specifically, ocean region data is generated by performing element-wise multiplication between the first forecast data and the coastline mask, coastal region data is generated by performing element-wise multiplication between the first forecast data and the coastal mask, and coastal strip region data is obtained by extracting the boundary positions through set difference operation between the coastline mask and the coastal mask. This effectively extracts the feature information of the ocean, the coast, and the coastal strip region, enabling differentiated processing of data from different regions.

[0124] Furthermore, global mean normalization is performed on the marine and coastal data to generate the first feature value. Global mean normalization uses the average value of the data from each region as a benchmark for standardization; for example, the mean of all grid points in the marine region is used as the normalization parameter.

[0125] Furthermore, the corresponding sea temperature gradient and spatiotemporal change rate are extracted from the data of the coastal strip area. Preferably, the sea temperature gradient can be obtained by calculating the temperature difference between adjacent grid points in the coastal strip area, and the spatiotemporal change rate can be determined by statistically analyzing the data fluctuation amplitude in the region within a continuous time step. Then, the sea temperature gradient and spatiotemporal change rate are used as the second feature value.

[0126] Furthermore, a weighted summation method is preferred to fuse the first feature value and the second feature value to obtain the first forecast data feature value. This can more accurately depict the complex temperature change characteristics at the land-sea interface, providing more comprehensive and refined feature input for subsequent bias correction, thereby improving the accuracy and reliability of three-dimensional sea surface temperature forecasts for nearshore waters.

[0127] In some embodiments, step S510 generates marine area data, coastal area data, and coastal strip area data based on the coastline mask, the coastal mask, and the first forecast data, including the following steps S610 to S630:

[0128] Step S610: Perform element-wise multiplication of the first forecast data and the coastline mask to generate ocean area data;

[0129] Step S620: Perform element-wise multiplication between the first forecast data and the coastal mask to generate coastal area data;

[0130] Step S630: Perform set difference operation on the coast mask and the coastal mask to extract the boundary position of the coastal strip area in the target area from the first forecast data, and obtain the corresponding coastal strip area data.

[0131] In this step, the first forecast data is multiplied element-wise with the coastline mask and the coastal mask respectively. The element-wise multiplication operation directly applies the binary distribution characteristics of the mask to the forecast data by multiplying pixel by pixel, thereby achieving physical isolation between the ocean and the coastal area and generating ocean area data.

[0132] Furthermore, the set difference operation is performed on the coastal mask and the coastal mask. The set difference operation eliminates overlapping areas through mathematical logic, accurately defines the geometric boundary of the coastal strip area, and extracts the boundary position of the coastal strip area in the target area from the first forecast data to obtain the corresponding coastal strip area data.

[0133] Furthermore, by subtracting the coastal mask matrix from the coastal mask matrix, a mask matrix containing only the coastal strip region is obtained. Then, element-wise multiplication is performed between this mask matrix and the first forecast data matrix to extract the data of the coastal strip region. This effectively distinguishes between ocean, coastal, and coastal strip region data, while providing an accurate regional division basis for subsequent feature extraction and data optimization. As a result, differentiated correction strategies can be adopted for different regions to improve the accuracy of nearshore marine forecast data.

[0134] In some embodiments, in step S130, the first forecast data is optimized based on the first forecast data feature values ​​according to the bias correction model to obtain the bias-corrected forecast data, including the following steps S710 to S740:

[0135] Step S710: Input the feature values ​​of the first forecast data into the encoder part of the bias correction model to generate feature embedding vectors;

[0136] Step S720: Calculate the gradient change of the feature embedding vector;

[0137] Step S730: Adjust the noise intensity value and the number of denoising iterations of the bias correction model during the data optimization process according to the gradient change.

[0138] Step S740: Based on the noise intensity value and the number of denoising iterations, optimize the first forecast data until the gradient change meets the preset convergence condition, and then output the corrected forecast data.

[0139] In this step, the feature values ​​of the first forecast data are input into the encoder part of the bias correction model to generate feature embedding vectors. Specifically, the encoder adopts a convolutional neural network structure, which includes multiple convolutional layers and pooling layers. The encoder performs dimensionality reduction and abstraction on the feature values ​​of the first forecast data, captures the global statistical features and local spatiotemporal correlations of the sea surface temperature field, and finally outputs a fixed-dimensional feature embedding vector.

[0140] Furthermore, an automatic differentiation technique is used to calculate the gradient of the feature embedding vector with respect to the model parameters, and the Euclidean distance of the gradient between two adjacent iterations is calculated as the gradient change. The gradient change is used to quantify the convergence state of the data optimization process by calculating the magnitude of change of the feature embedding vector in consecutive iteration steps.

[0141] Furthermore, based on the gradient change, the noise intensity and number of denoising iterations of the bias correction model during data optimization are adjusted. For example, when the gradient change is higher than a preset threshold, the number of iterations is increased and the noise intensity is decreased to accelerate convergence; when the gradient change is lower than the preset threshold, the number of iterations is decreased and the noise intensity is maintained; when the gradient change is large, the noise intensity is increased and the number of iterations is decreased to improve accuracy.

[0142] Furthermore, based on the noise intensity value and the number of denoising iterations, the first forecast data is optimized until the gradient change meets the preset convergence condition, at which point the corrected forecast data is output. Thus, through the iterative denoising process, systematic biases in the forecast data can be effectively removed, improving the accuracy and reliability of nearshore marine forecasts.

[0143] In some embodiments, the method framework of this application is DOLPHIN, which reconstructs the sea surface temperature forecast bias correction as a spatiotemporal data regeneration problem, resorts to denoising methods, and achieves the approximation of the forecast data to the distribution of observation-based reanalysis data, thereby reducing the error.

[0144] 1. Obtain sea surface temperature forecast data:

[0145] Given a regional sea surface temperature forecast field grid data , Represents a three-dimensional tensor space. , Indicates the spatial resolution of grid data. This indicates the number of sea surface layers at different depths, pre-defined. =9, sea surface temperature forecast data is: In diffusion models, the forward diffusion process is considered to occur through a specified... and Noise generation The transformation process can be described by the following formula:

[0146] (1)

[0147] in, It is noise intensity. It is Gaussian noise. Represents the standard normal distribution. This indicates that Gaussian noise follows a standard normal distribution. The mean of noise intensity, Standard deviation of noise intensity The mean is The variance is The normal distribution This indicates that the noise intensity conforms to the above distribution. The reverse process of the model is considered as training a denoiser. ,pass Dependent link prediction data The noise that needs to be added at a specific noise level. Specifically:

[0148] (2)

[0149] in, It is a network that needs to be trained. It is a metric for measuring links. and Scaling input and output amplitudes Map the noise level to a condition as Input.

[0150] Furthermore, by specifying a noise intensity The information obtained is retained in the corresponding Stage diffusion state Through the analysis of Iteratively executing the denoising formula (2) yields a bias-corrected result that approximates the distribution of the reanalyzed data. During training, the mean squared error loss is optimized:

[0151] (3)

[0152] in, It is a truth value. The data was generated using a diffusion model. Indicates noise intensity Next, noise stage From 0 to Mean square error under the given conditions.

[0153] In some embodiments, forecast data with errors is input into the model. First, partial noise is added to obtain data with some information preserved. Then, denoising inference is performed on this data. Under the premise of implicitly guided generation, the model achieves correction. Simultaneously, to reduce the hyperparameter errors introduced by the noise intensity and denoising steps, the encoder part of the diffusion model is extracted, and the data obtained from the first inference is input here. The parameterization scheme of the noise intensity and denoising steps is adaptively improved with the goal of minimizing the feature gradient of the data. Finally, a coastal masking algorithm based on erosion operations is used to focus on key areas, resulting in a corrected result that converges with the reanalysis data distribution and reduces coastal errors.

[0154] 2. Extracting land-sea masking from sea surface temperature forecast data using a land-sea masking filter:

[0155] Sea surface temperature forecast data After input, because As a numerical grid, the forecast data has distinctly different numerical ranges across land and sea regions. First, land and sea masks are obtained through filtering. To facilitate the subsequent extraction of the coastal mask. :

[0156] (4)

[0157] in, This indicates that extreme value filtering is used. Each depth layer of different marine and land structures corresponds to a specific mask.

[0158] 3. Input the land and sea mask into the erosion module to obtain the coastal mask:

[0159] The purpose of segmenting land and sea is to obtain geographical boundaries, while coastal areas require the extraction of the land-sea spatial junction zone. In this embodiment, the erosion operation operator is used. Sea and land shields Perform the operation to extract the coastline area, where Indicates corrosion operation. This represents the entire erosion convolution kernel.

[0160] First, a 3×3 erosion kernel is designed, which acts on the binarized land and sea mask. The etching operation is performed at the location where the numerical anomaly occurs. The etching operation is expressed as:

[0161] (5)

[0162] in, This represents the set of all numerical points in the land-sea mask. This represents the entire erosion convolution kernel. This represents a single element in the erosion convolution kernel. This represents the numerical value of the land-sea mask corresponding to the center position of the erosion convolution kernel. When the erosion kernel slides to a certain position on the land-sea mask... When, if all elements at the corresponding position in the kernel All belong to the marine part Then, the position of the element corresponding to the central element of the erosion nucleus in the land-sea mask is retained. The value is 1, otherwise, The value is set to 0. This operation removes the eroded portion of the coastal zone. Finally, the sea-land mask is compared with the result of formula (5) to extract the coastal area. This process can be described as follows:

[0163]

[0164] Through the above concise corrosion operation, only the following steps are performed: Extracting the coastal zone in terms of time complexity, where It is the computation time complexity. It is the number of numerical grid points. This represents the number of elements in the erosion operator. With GPU-accelerated computation, the time overhead is extremely low.

[0165] 4. Input the sea surface temperature forecast data, land-sea mask, and coastal mask into the trained diffusion model to obtain the processed sea surface temperature forecast data:

[0166] Firstly, for sea surface temperature (SST) forecast data, a controlled partial noise injection (CPNI) strategy is proposed. This strategy utilizes partially retained data for implicit guidance and employs denoising to correct biases. During sampling, the model adjusts the noise level based on the maximum noise intensity. and minimum noise intensity This generates a noise intensity scheduler. The process can be expressed as the following formula:

[0167] ;

[0168] ;

[0169] in, It represents the number of denoising iterations performed during the sampling process. It is a noise intensity scheduler. This represents the noise that the model aims to remove at the corresponding time step. Indicates the numerical scaling scale. The scheduler is used to determine... arrive The expected noise intensity at each denoising step guides the model's evolution over time, progressively denoising from pure Gaussian noise to clear samples. Specifically, this is achieved by specifying the noise intensity. The gridded data is partially denoised. During the denoising process, the model infers the intensity of the added noise. Corresponding expected number of noise addition steps Using this as a priori information, the model is guided in... After a finer denoising step, the prediction error is reduced. Specifically, based on a specified noise intensity, the closest denoising step is found in the list of noise intensities generated by the scheduler. This process can be expressed by the following formula:

[0170] ;

[0171] in, The model represents the Under what circumstances, the noise that you want to remove is... This means in this Find the minimum value among the data. This indicates the index position corresponding to a specified noise intensity. This indicates traversing the entire scheduler and then utilizing the obtained... Construct a scheduler starting from a specified noise level. :

[0172] ;

[0173] Therefore, the model can utilize a scheduler with a specified noise level. In each denoising stage, the second-order Hurn method is used to complete the denoising, and the bias-corrected data is obtained. ,in, Indicates from The entire scheduler is traversed starting from the corresponding index position.

[0174] 5. Extract feature values ​​from the processed sea surface temperature forecast data:

[0175] After the above operations. This has reduced some forecast errors. respectively with and By multiplying and summing the elements, we obtain the feature data that highlights the coastal zone.

[0176] 6. Based on the trained diffusion model, the processed sea surface temperature forecast data is optimized by a preset number of steps according to the feature values ​​to obtain the positively revised sea surface temperature forecast data.

[0177] In the CPNI strategy, the model achieves stochastic modeling of natural scenes. Additionally, the model incorporates starting noise intensity. and denoising iteration steps Two hyperparameters. During the sampling process, the parameters are optimized by focusing on the gradient changes of the corrected data.

[0178] Specifically, the encoder in the model is extracted first. Then the denoised features and zero-value noise intensity vector Input encoder Obtain its hidden variables This process can be described as follows:

[0179] ;

[0180] Subsequently, L1 norm is used to extract key features, enhancing the representational ability of latent variables, and these features are used as a loss term in the calculation. gradient This reveals the difference between the latent variable space and the distribution of the reanalysis data. This process can be represented as:

[0181] ;

[0182] Finally, the PLO optimizer is used to select different combinations of noise intensity and denoising steps in the solution space, aiming to minimize the feature space. Gradient changes reduce the errors introduced by hyperparameters, thereby driving the prediction results to approximate the data distribution characteristics learned by the encoder. By performing a preset number of optimization steps, the corrected result is finally obtained.

[0183] In some embodiments, an experimental dataset is first set up. Referring to the Xihe study, the model of this embodiment is trained using the 1 / 12° resolution GLORYS12 reanalysis dataset. The spatiotemporal range of the data is from January 1, 2012 to December 31, 2020, geographically covering the nearshore areas of 10°N-30°N and 105°E-125°E. Deterministic models are trained using forecast data from Xihe at the same time scale with forecast times of 1, 3, 5, 7, and 9 days. The test set consists of Xihe forecast data with forecast times of 1, 3, 5, 7, and 9 days in 2020.

[0184] Furthermore, this embodiment primarily focuses on bias correction for the large-scale intelligent forecasting model. Referring to current mainstream work, this embodiment selects RMSE as the evaluation metric. This metric is highly sensitive to errors and can intuitively reflect the effectiveness of bias correction.

[0185] Furthermore, an adaptive moment estimation optimizer was trained for 2500 epochs with a learning rate of 0.00075 and a batch size of 4. The parameter settings for the diffusion model were the same as those for the Elucidating the Design Space of Diffusion-Based Generative Models (EDM). , , , , And set noise reduction steps The number is 200. All models are trained fairly on 8 A100 GPUs using the same strategy.

[0186] To demonstrate the advanced nature of the model in this embodiment for correcting biases in 3D sea surface temperature forecast data, it was compared with mainstream solution paradigms, including ConvLSTM, ConvGRU, SAAConvLSTM, TAAConvLSTM, U-Net, and the current best-performing models in the field of computer vision, including EDAE-Net, Swin-Transformer, and MambaOut.

[0187] Table 1 shows a quantitative comparison of model bias correction results under different reporting times. The RMSE performance of different models for sea surface temperature data under different reporting times is shown. The lower the index, the better the performance. Thanks to the progressive error reduction capability of the diffusion model in gradually denoising, the model in this embodiment achieved the minimum RMSE result in the task.

[0188] Table 1

[0189]

[0190] like Figure 2 and Figure 3 The image shows a comparison of the performance of mainstream models for bias correction tasks. Figure 2 In this study, U-Net and its variants, a recurrent neural network methods, achieve good correction results, but as deterministic models, they cannot model natural randomness well. Figure 3 In recurrent neural network models, ConvGRU performs relatively well, while some LSTM and its variants perform relatively poorly. This may be because ConvGRU uses only two gates, which better captures the local mutation characteristics of bias. (Figure 2 and...) Figure 3 This section compares the performance of the current mainstream paradigm for bias correction tasks and the model in this embodiment for forecast data with different start times. The color bars represent the 95% confidence intervals. Figure 2For comparison of recurrent neural network methods, Figure 3 This is a comparative method for recurrent neural networks.

[0191] Therefore, to address the errors in coastal sea surface temperature forecasts, this embodiment proposes a method for bias correction using a diffusion model. By learning the distribution space of the reanalysis data, decoupling from the forecast data is achieved during the training process. Partial noise addition is used to model a natural stochastic process during sampling, achieving non-explicit guidance. Extensive experiments with real forecast data demonstrate that this embodiment performs well.

[0192] like Figure 4 As shown in some embodiments of this application, a forecast data deviation correction system is provided. The system includes an acquisition module 410, an extraction module 420, and a correction module 430. Specifically:

[0193] The acquisition module 410 is used to acquire the initial forecast data of the target area;

[0194] The extraction module 420 is used to extract the land-sea boundary features of the target area from the initial forecast data of the target area, and obtain the coast mask and the coastal mask of the target area.

[0195] The correction module 430 is used to input the initial forecast data and the coastal mask of the target area into the deviation correction model to obtain the corrected forecast data output by the deviation correction model.

[0196] The correction process of the deviation correction model includes:

[0197] The initial forecast data is partially noise-added to obtain the first forecast data;

[0198] Based on the coastline mask and the coastal mask, the feature values ​​of the first forecast data are extracted from the first forecast data;

[0199] Based on the bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data.

[0200] It should be noted that the forecast data deviation correction system provided in this embodiment is based on the same inventive concept as the forecast data deviation correction method described above. Therefore, the relevant content of the forecast data deviation correction method described above also applies to the content of the forecast data deviation correction system, and will not be repeated here.

[0201] To achieve this, the system acquires initial forecast data for the target area; extracts the land-sea boundary features of the target area from the initial forecast data to obtain the coastal mask and shoreline mask; and inputs the initial forecast data and the shoreline mask of the target area into the bias correction model to obtain the corrected forecast data output by the bias correction model. In this way, the system can efficiently extract the shoreline mask, and while retaining the effective guiding information of the forecast data using a non-complete noise-addition strategy, it also adds some noise and optimizes features in the initial forecast data. This allows the model to be trained simply by reanalyzing the data. Furthermore, by combining feature extraction and dynamic parameter optimization, the system effectively solves the problems of large errors in 3D sea surface temperature forecasts in nearshore waters, long-term error accumulation, and miscorrection caused by land-sea interference, and significantly improves the model's adaptability to complex small-scale environmental factors.

[0202] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-mentioned method for correcting the deviation of the forecast data.

[0203] like Figure 5 , Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device includes:

[0204] At least one battery;

[0205] At least one memory;

[0206] At least one processor;

[0207] At least one program;

[0208] The program is stored in memory, and the processor executes at least one program to implement the above-described method for correcting deviations in forecast data.

[0209] This electronic device can be any smart terminal, including mobile phones, tablets, personal digital assistants (PDAs), and in-vehicle computers.

[0210] The electronic devices according to embodiments of this application will now be described in detail.

[0211] The processor 1600 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this disclosure.

[0212] The memory 1700 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1700 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1700 and is called and executed by the processor 1600 to execute a method for correcting deviations in forecast data according to an embodiment of this disclosure.

[0213] The input / output interface 1800 is used to implement information input and output.

[0214] The communication interface 1900 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0215] Bus 2000 transmits information between various components of the device (e.g., processor 1600, memory 1700, input / output interface 1800, and communication interface 1900);

[0216] The processor 1600, memory 1700, input / output interface 1800 and communication interface 1900 are connected to each other within the device via bus 2000.

[0217] This disclosure also provides a storage medium, which is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the aforementioned method for correcting deviations in forecast data.

[0218] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0219] The embodiments described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0220] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this disclosure, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0221] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0222] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0223] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0224] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0225] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0226] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0227] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0228] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0229] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

[0230] The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A method for correcting biases in forecast data, characterized in that, The method includes: Obtain initial forecast data for the target area; Extract the land-sea boundary features of the target area from the initial forecast data of the target area to obtain the coast mask and the coastal mask of the target area; The initial forecast data and the coastal mask of the target area are input into the deviation correction model to obtain the corrected forecast data output by the deviation correction model. The correction process of the deviation correction model includes: The initial forecast data is partially noise-added to obtain the first forecast data; Based on the coast mask and the coastal mask, the first forecast data feature value is extracted from the first forecast data; Based on the aforementioned bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data. The step of optimizing the first forecast data based on the bias correction model and the feature values ​​of the first forecast data to obtain the bias-corrected forecast data includes: The feature values ​​of the first forecast data are input into the encoder part of the bias correction model to generate a feature embedding vector; Calculate the gradient change of the feature embedding vector; Based on the gradient change, adjust the noise intensity value and the number of denoising iterations of the deviation correction model during the data optimization process; Based on the noise intensity value and the number of denoising iterations, the first forecast data is optimized until the gradient change meets the preset convergence condition, at which point the corrected forecast data is output.

2. The method for correcting the deviation of forecast data according to claim 1, characterized in that, The training process of the bias correction model includes: Preset initial deviation correction model; Obtain historical reanalysis data for the target area; The initial bias correction model is trained based on the historical reanalysis data to obtain the first bias correction model; The mean squared error loss is calculated based on the model output of the first deviation correction model and the historical reanalysis data. The first deviation correction model is optimized based on the mean squared error loss to obtain the deviation correction model.

3. The method for correcting the deviation of forecast data according to claim 1, characterized in that, The step of extracting the land-sea boundary features of the target area from the initial forecast data of the target area to obtain the coast mask and coastal mask of the target area includes: The initial forecast data for the target area is filtered for extreme values ​​to obtain the land and sea masking code for the target area; The land and sea mask is subjected to morphological erosion to obtain the processed mask; The coastal mask is obtained by performing a set difference operation between the processed mask and the land and sea mask.

4. The method for correcting the deviation of forecast data according to claim 1, characterized in that, The step of adding some noise to the initial forecast data to obtain the first forecast data includes: A noise intensity sequence is generated according to a preset noise intensity scheduler, the noise intensity sequence containing multi-level decreasing noise values ​​from the maximum noise intensity to the minimum noise intensity; Find the noise intensity index position in the noise intensity sequence that is closest to the preset target noise intensity value; Starting from the index position, a noise signal of corresponding intensity is added to the initial forecast data to generate the first forecast data.

5. The method for correcting the deviation of forecast data according to claim 1, characterized in that, The step of extracting first forecast data feature values ​​from the first forecast data based on the coastline mask and the coastal mask includes: Based on the coastline mask, the coastal mask, and the first forecast data, ocean area data, coastal area data, and coastal strip area data are generated. The ocean area data and the coastal area data are subjected to global mean normalization to generate a first feature value; Extract the corresponding sea surface temperature gradient and spatiotemporal rate of change from the data of the coastal strip region; The sea surface temperature gradient and the spatiotemporal rate of change are used as the second characteristic value; The first feature value and the second feature value are fused to obtain the first forecast data feature value.

6. The method for correcting the deviation of forecast data according to claim 5, characterized in that, The step of generating marine area data, coastal area data, and coastal strip area data based on the coastline mask, the coastal mask, and the first forecast data includes: The first forecast data is multiplied element-wise with the coastline mask to generate ocean area data. The first forecast data is multiplied element-wise with the coastal mask to generate coastal area data. The set difference operation is performed on the coast mask and the coastal mask to extract the boundary position of the coastal strip area in the target area from the first forecast data, and obtain the corresponding coastal strip area data.

7. A forecast data deviation correction system, characterized in that, The system includes: The acquisition module is used to acquire initial forecast data for the target area; The extraction module is used to extract the land-sea boundary features of the target area from the initial forecast data of the target area, and obtain the coast mask and the coastal mask of the target area; The correction module is used to input the initial forecast data and the coastal mask of the target area into the deviation correction model to obtain the corrected forecast data output by the deviation correction model. The correction process of the deviation correction model includes: The initial forecast data is partially noise-added to obtain the first forecast data; Based on the coast mask and the coastal mask, the first forecast data feature value is extracted from the first forecast data; Based on the aforementioned bias correction model, the first forecast data is optimized according to the feature values ​​of the first forecast data to obtain the bias-corrected forecast data. The step of optimizing the first forecast data based on the bias correction model and the feature values ​​of the first forecast data to obtain the bias-corrected forecast data includes: The feature values ​​of the first forecast data are input into the encoder part of the bias correction model to generate a feature embedding vector; Calculate the gradient change of the feature embedding vector; Based on the gradient change, adjust the noise intensity value and the number of denoising iterations of the deviation correction model during the data optimization process; Based on the noise intensity value and the number of denoising iterations, the first forecast data is optimized until the gradient change meets the preset convergence condition, at which point the corrected forecast data is output.

8. An electronic device, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform a forecast data deviation correction method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform a method for correcting deviations in forecast data as described in any one of claims 1 to 6.

Images