Data-driven ocean near-real-time observation rolling integrated forecasting method and device

Abstract

The invention relates to the technical field of weather forecasting, in particular to a data-driven ocean near-real-time observation rolling integrated forecasting method and device, and the method comprises the following steps: obtaining observation and forecasting data, calculating a difference, updating a deviation library, extracting an expected deviation, debiasing an original forecasting field, and generating a candidate integrated temperature and salt field; and calculating the density value by hardware to generate an inconsistent mask, screening unstable points based on the mask, adjusting the weight matrix, iteratively updating until the unstable points are smaller than a threshold value, and outputting a forecast result. According to the method, the difference value of near-real-time observation and forecasting is calculated, a rolling model deviation database is dynamically updated, an expected deviation field is subtracted when new forecasting is generated, a depolarization candidate state field is formed, device deviation is restrained, hardware is used for calculating density in an accelerated mode, vertical adjacent density is compared to generate inconsistent masks, and unstable grid points are screened. And adjusting the integrated weight matrix, and iteratively updating the candidate integrated temperature and salt field until the unstable point reaches the standard, so as to ensure that the result meets the physical rule constraint.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather forecasting, and in particular to a data-driven near-real-time ocean observation rolling integrated prediction method and device. BACKGROUND

[0002] The technical field of weather forecasting involves monitoring, analyzing and predicting the state of the atmosphere and its evolution. The core of this field includes collecting meteorological parameters such as temperature, humidity, air pressure, wind speed, processing these observation data using numerical models and statistical methods, and finally generating inferences about future weather changes. Among them, the traditional data-driven near-real-time ocean observation rolling integrated prediction method refers to a mode that uses observation data to calculate the future state of the ocean environment. This method mainly predicts the dynamic changes of ocean hydrology and meteorological elements. The traditional implementation usually relies on inputting observation data from different sources, such as satellite remote sensing images and sea surface buoy data, into multiple independent numerical prediction models, and then manually selecting or using a simple arithmetic average method to integrate the calculation results of different models as the final prediction conclusion.

[0003] The traditional data-driven prediction method relies on multiple independent models and uses a simple arithmetic average method or manual selection to integrate the results. This approach cannot effectively handle the inherent device bias of the model, leading to cumulative prediction bias over time and continuous decline in prediction accuracy. In addition, this simple result integration method lacks effective constraints on physical laws when fusing data from different sources, which may lead to physically unreasonable phenomena in the prediction conclusion, such as ocean water density inversion. This inconsistency cannot occur in reality, severely affecting the reliability and practical application value of the prediction conclusion. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings in the prior art and to provide a data-driven near-real-time ocean observation rolling integrated prediction method and device.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solution: a data-driven near-real-time ocean observation rolling integrated prediction method, comprising the following steps:

[0006] S1: obtaining near-real-time ocean observation data and prediction data of the last period, calculating the difference value between the near-real-time ocean observation data and the prediction data of the last period, and updating the rolling model bias database;

[0007] S2: obtaining an original temperature-salinity prediction field, extracting an expected bias field from the rolling model bias database, generating a debiased candidate state field set by subtracting the expected bias field from the original temperature-salinity prediction field, and fusing the debiased candidate state field set to generate a candidate integrated temperature-salinity field;

[0008] S3: calculating the grid point density values in the candidate integrated temperature-salinity field through coprocessing hardware, comparing vertically adjacent density values to generate a binary physical inconsistency mask;

[0009] S4: screening unstable grid points in the candidate integrated temperature-salinity field based on the binary physical inconsistency mask, adjusting an integrated weight matrix, and iteratively updating the candidate integrated temperature-salinity field and the binary physical inconsistency mask until the number of unstable grid points in the binary physical inconsistency mask is less than a threshold value, and outputting an integrated prediction result.

[0010] As a further scheme of the present application, the rolling model bias database comprises difference values, the candidate integrated temperature-salinity field comprises grid point density values, the binary physical inconsistency mask comprises unstable grid points, and the integrated prediction result specifically refers to the candidate integrated temperature-salinity field after iterative updating.

[0011] As a further scheme of the present application, the step S1 specifically comprises:

[0012] S11: obtaining near-real-time ocean observation data and prediction data of a previous period, performing spatiotemporal interpolation processing on the near-real-time ocean observation data to make the grid structure of the near-real-time ocean observation data consistent with the grid structure of the prediction data of the previous period, and generating an aligned observation field;

[0013] S12: calculating residuals of the aligned observation field and the prediction data of the previous period at a plurality of corresponding grid points, defining the residuals as difference values, and taking the difference values to represent the degree of systematic bias of the ocean prediction model in a target region and a time window;

[0014] S13: storing the difference values in a rolling model bias database according to their corresponding spatiotemporal coordinate information, and setting a rolling time window; when the current difference values are stored, historical difference values in the rolling model bias database with time stamps earlier than the rolling time window are synchronously searched and removed, and the rolling model bias database is dynamically updated.

[0015] As a further scheme of the present application, the step S2 specifically comprises:

[0016] S21: obtaining an original temperature-salinity prediction field, retrieving and extracting a corresponding set of historical difference values from the rolling model bias database according to the spatiotemporal coordinate range of the original temperature-salinity prediction field, and generating an expected bias field through spatiotemporal weighted average calculation on the set of historical difference values;

[0017] S22: taking the original temperature-salinity prediction field as one of the members, generating a plurality of perturbation prediction fields by using random perturbation fields to perturb the original temperature-salinity prediction field multiple times, and generating a set of debiased candidate state fields by subtracting the expected bias field from the original temperature-salinity prediction field and the plurality of perturbation prediction fields one by one;

[0018] S23: initializing an integrated weight matrix representing the weights of a plurality of member fields in the set of debiased candidate state fields, and generating a candidate integrated temperature-salinity field by weighted fusion of all member fields in the set of debiased candidate state fields according to the integrated weight matrix.

[0019] As a further scheme of the present application, the step S3 is specifically:

[0020] S31: loading the temperature, salinity and pressure data included in the candidate integrated temperature-salinity field into the dedicated video memory of the coprocessing hardware, and using the parallel computing cores of the coprocessing hardware to calculate the grid point density value of each grid point in the candidate integrated temperature-salinity field according to the internationally recognized ocean state equation:

[0021] ;

[0022] and generating a three-dimensional density field;

[0023] wherein, represents the grid point density value, represents the temperature value of the grid point in the candidate integrated temperature-salinity field, represents the salinity value of the grid point in the candidate integrated temperature-salinity field, represents the pressure value of the grid point in the candidate integrated temperature-salinity field;

[0024] S32: in the three-dimensional density field, for each grid point of a non-water bottom boundary, extracting the grid point density value of itself and the grid point density value of its vertically lower adjacent grid point, i.e. the density value, and constructing a to-be-compared density pair with the two density values;

[0025] S33: traversing all the to-be-compared density pairs, if the density value of the grid point above is greater than the density value of its vertically lower adjacent grid point, determining that the grid point above is an unstable grid point and marking it as 1, otherwise marking it as 0, and summarizing the marking results of all grid points to generate a binary physical inconsistency mask.

[0026] As a further scheme of the present application, the step S4 is specifically:

[0027] S41: based on the binary physical inconsistency mask, screening and counting the current total number of the unstable grid points marked as 1 in the candidate integrated temperature-salinity field.

[0028] S42: Determine whether the current total number of unstable grid points is less than the preset threshold. If it is less, determine that the current candidate integrated temperature and salinity field has reached the physical consistency requirement, terminate the iteration, and output the current candidate integrated temperature and salinity field as the integrated prediction result.

[0029] S43: If the current total number of unstable grid points is not less than the threshold, then a penalty function is applied based on the spatial distribution and inconsistency of the unstable grid points:

[0030] ;

[0031] Adjust the integrated weight matrix to reduce the weights of member fields in the debiased candidate state field set that cause physical instability;

[0032] in, This represents the adjusted integrated weight matrix. This represents the integrated weight matrix before adjustment. The Hadamard product represents the product of corresponding elements in a matrix. Representative and A matrix of the same dimension that is entirely identical. This represents the penalty coefficient used to control the magnitude of weight adjustments. A matrix representation of the binary physical inconsistency mask;

[0033] S44: Based on the adjusted integrated weight matrix, re-fuse the debiased candidate state field set and update the candidate integrated temperature and salinity field;

[0034] The grid point density values ​​of the candidate integrated temperature and salinity field are recalculated using the coprocessor hardware, and a new binary physical inconsistency mask is generated. Then, the process returns to S41 to execute the next iteration.

[0035] As a further aspect of the present invention, the coprocessor hardware is a graphics processor, and the process of calculating the grid point density value using the parallel computing core of the coprocessor hardware includes:

[0036] The temperature, salinity, and pressure data of the candidate integrated temperature and salinity fields are constructed into a three-dimensional texture or a unified memory buffer for the coprocessing hardware to retrieve data during parallel computing.

[0037] Write and compile a computation kernel or shader program for the coprocessor hardware architecture, wherein the ocean state equation is embedded in the computation kernel or shader program. The complete computational logic;

[0038] The computing core or the shader program is started to calculate the grid point density values of all grid points in the three-dimensional density field at the same time by using thousands of parallel processing units of the coprocessor hardware, and the calculation results are written back to the dedicated memory of the coprocessor hardware.

[0039] As a further scheme of the present application, the rolling model bias database is a key-value storage device based on a four-dimensional space-time grid index, and the process of updating the rolling model bias database in S13 comprises:

[0040] The spatial latitude and longitude coordinates, the vertical depth level and the observation timestamp corresponding to the difference value are extracted, the spatial latitude and longitude coordinates, the vertical depth level and the observation timestamp are combined to generate a four-dimensional space-time index key;

[0041] The difference value itself is stored in the rolling model bias database as a numerical value associated with the four-dimensional space-time index key.

[0042] The rolling time window is set, and at each update, all the four-dimensional space-time index keys in the rolling model bias database are retrieved, and if the observation timestamp included in the four-dimensional space-time index key is earlier than the start time of the rolling time window, the corresponding four-dimensional space-time index key and its associated difference value are deleted from the rolling model bias database.

[0043] As a further scheme of the present application, the process of weighted fusion calculation of the debiased candidate state field set in S23 comprises:

[0044] The integrated weight matrix is obtained, and the integrated weight matrix stores the current weight coefficient of each member field in the debiased candidate state field set;

[0045] Each grid point in the candidate integrated temperature-salinity field is traversed, and the temperature and salinity values of all member fields in the debiased candidate state field set at the grid point are extracted;

[0046] The temperature and salinity values are multiplied by the corresponding weight coefficients in the integrated weight matrix respectively to perform summation operation, and the operation result is taken as the final temperature and salinity values of the candidate integrated temperature-salinity field at the grid point.

[0047] The data-driven near-real-time ocean observation rolling integration prediction device is used to implement the data-driven near-real-time ocean observation rolling integration prediction method, and the device comprises:

[0048] a bias updating module configured to obtain near real-time ocean observation data and forecast data of a previous period, calculate a difference value between the near real-time ocean observation data and the forecast data of the previous period, update a rolling model bias database, and pass to a temperature-salinity field correction module;

[0049] a temperature-salinity field correction module configured to obtain an original temperature-salinity forecast field, extract an expected bias field from the rolling model bias database, generate a debiased candidate state field set by subtracting the expected bias field from the original temperature-salinity forecast field, fuse the debiased candidate state field set to generate a candidate integrated temperature-salinity field, and pass to a physical consistency checking module;

[0050] a physical consistency checking module configured to calculate grid point density values in the candidate integrated temperature-salinity field through coprocessing hardware, compare vertically adjacent density values to generate a binary physical inconsistency mask, and pass to an integrated weight iteration module;

[0051] an integrated weight iteration module configured to screen unstable grid points in the candidate integrated temperature-salinity field through the binary physical inconsistency mask, adjust an integrated weight matrix, and iteratively update the candidate integrated temperature-salinity field and the binary physical inconsistency mask until the number of unstable grid points in the binary physical inconsistency mask is less than a threshold value, and output an integrated forecast result.

[0052] Compared with the prior art, the application has the advantages and positive effects that:

[0053] In the application, the difference value between near real-time observation data and forecast of a previous period is calculated, the rolling model bias database is dynamically updated, the expected bias field extracted from the database is actively subtracted from the original forecast field when a new forecast is generated, the debiased candidate state field is formed, the accumulation of device bias is effectively inhibited, the forecast accuracy is improved, the grid point density values are calculated using coprocessing hardware, the physical inconsistency mask is generated by comparing vertically adjacent densities, the unstable grid points are screened based on the mask, the integrated weight matrix is adjusted, and the candidate integrated temperature-salinity field is iteratively updated until the number of unstable grid points meets the standard, so that the integrated forecast result output meets the physical law constraint, and the stability and authenticity of the conclusion are improved. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The application is a data-driven ocean near real-time observation rolling integrated forecast method flowchart;

[0055] Figure 2 The application is a rolling model bias database dynamic updating flowchart;

[0056] Figure 3 The application is a candidate integrated temperature-salinity field generation flowchart;

[0057] Figure 4 A flow chart for generating the physical inconsistency mask of the present application is shown in Figure 1.

[0058] Figure 5 A flow chart for integrated forecast result iteration optimization of the present application is shown in Figure 2. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions realized based on software are described in detail below in combination with device architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and do not constitute a limitation on the scope of protection.

[0060] In the description of the present application, the device architecture relationship or data processing flow indicated by the terms "hierarchy", "module", "interface", "data flow", "client", "server" and the like are defined based on the corresponding architecture diagrams or flow charts of the embodiments. This way of expression is only used to clearly explain the logical relationship of each element in the technical solution, and is not limited to the physical deployment form. The "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances or functional components, etc. Extensible elements, and the specific number is determined according to the actual business scenario.

[0061] Please refer to Figure 1 and Figure 2 The present application provides a technical solution: a data-driven near-real-time ocean observation rolling integrated prediction method, comprising the following steps:

[0062] S1: Obtain near-real-time ocean observation data and prediction data of the last period, calculate the difference value between the near-real-time ocean observation data and the prediction data of the last period, and update the rolling model bias database;

[0063] The steps of S1 are specifically:

[0064] S11: Obtain near-real-time ocean observation data and prediction data of the last period, perform spatial and temporal interpolation processing on the near-real-time ocean observation data to make its grid structure consistent with that of the prediction data of the last period, and generate an aligned observation field;

[0065] S12: Calculate the residual of the aligned observation field and the prediction data of the last period at multiple corresponding grid points, define the residual as the difference value, and the difference value represents the degree of device bias of the ocean prediction model in the target area and time window;

[0066] S13: Store the difference value in the rolling model bias database according to its corresponding spatio-temporal coordinate information, and set a rolling time window. When the current difference value is stored, the historical difference values in the rolling model bias database with a time stamp earlier than the rolling time window are retrieved and deleted synchronously, and the rolling model bias database is dynamically updated.

[0067] The rolling model bias database is a key-value storage device based on a four-dimensional spatio-temporal grid index. The process of updating the rolling model bias database in S13 includes:

[0068] Extract the spatial latitude and longitude coordinates, vertical depth level, and observation time stamp corresponding to the difference value, combine the spatial latitude and longitude coordinates, vertical depth level, and observation time stamp to generate a four-dimensional spatio-temporal index key.

[0069] Store the difference value itself as a numerical value associated with the four-dimensional spatio-temporal index key in the rolling model bias database.

[0070] Set a rolling time window. At each update, retrieve all four-dimensional spatio-temporal index keys in the rolling model bias database. If the observation time stamp included in the four-dimensional spatio-temporal index key is earlier than the start time of the rolling time window, delete the corresponding four-dimensional spatio-temporal index key and its associated difference value from the rolling model bias database.

[0071] Obtain near-real-time ocean observation data, for example, through Argo floats, satellite remote sensing, shore-based radars, and other ocean observation platforms. In the period from 00:00 UTC on May 15, 2024, to 06:00 UTC on May 15, 2024, continuous observations of temperature, salinity, and ocean current physical quantities were conducted in the northern South China Sea region (18.0 degrees north latitude to 22.0 degrees north latitude, 112.0 degrees east longitude to 116.0 degrees east longitude). The obtained observation data includes profile data from 15 Argo floats, each profile data contains temperature and salinity data points every 10 meters from the sea surface to a depth of 2000 meters. At the same time, obtain the previous cycle of ocean forecast data, which is released by the global ocean model at 18:00 UTC on May 14, 2024, with a forecast time point of 00:00 UTC on May 15, 2024, to 06:00 UTC on May 15, 2024. Its grid resolution is 0.1 degrees x 0.1 degrees in the horizontal direction and 50 standard depth layers in the vertical direction, such as: surface layer, 5 meters, 10 meters, 25 meters, 50 meters, 75 meters, 100 meters, 125 meters, 150 meters, 200 meters, 250 meters, 300 meters, 400 meters, 500 meters, 600 meters, 700 meters, 800 meters, 900 meters, 1000 meters, 1250 meters, 1500 meters, 2000 meters, etc. Each grid point contains temperature and salinity forecast values.

[0072] The spatio-temporal interpolation is performed on the near-real-time ocean observation data to make its grid structure consistent with that of the forecast data of the last period. First, spatial interpolation is performed to interpolate the data of the discrete observation points of the Argo floats to the horizontal grid points of the forecast model. For example, the inverse distance weighted (IDW) interpolation method is used. Suppose a forecast grid point (latitude , longitude ) needs to be interpolated for temperature value. The nearest 3 Argo float observation points around it are retrieved, whose coordinates and observed temperature values are respectively: , , , , , , , The horizontal distance from each observation point to the forecast grid point is calculated: , , The weight is calculated: The normalized weight , , The interpolated temperature value .

[0073] Then, vertical interpolation is performed to interpolate the data of the non-standard depth layers of the Argo floats to the standard depth layers of the forecast model. For example, linear interpolation is used. Suppose a float observes temperature at 100 m depth and temperature at 200 m depth. The forecast model has a standard depth layer at 150 m. The interpolated temperature value . Finally, temporal interpolation is performed to interpolate the data of different observation times to the time step of the forecast model. Suppose the time step of the forecast model is 3 hours, and the aligned observation field at 03:00 UTC on May 15, 2024 is needed to be generated. If the observation data has data at 01:00 UTC and 04:00 UTC on May 15, 2024, then the linear interpolation is performed using the interpolated results of the two time points to generate the aligned observation field at 03:00 UTC. Through the above spatio-temporal interpolation, the aligned observation field with the same horizontal grid, vertical layer number and time step as the forecast data of the last period is generated.

[0074] ​​The residuals of the aligned observation field and the prediction data of the last cycle on multiple corresponding grid points are calculated. Select the northern South China Sea region, the prediction range is from north latitude 18.0 degrees to 22.0 degrees, east longitude 112.0 degrees to 116.0 degrees, vertical depth 0 to 2000 meters, and the time window is from 00:00 UTC on May 15, 2024 to 06:00 UTC on May 15, 2024. Assuming at a certain grid point (x, y, z, t): The temperature value of the aligned observation field at the grid point is , and the salinity value is . The temperature value of the prediction data of the last cycle at the grid point is , and the salinity value is . The temperature residual (difference value) is calculated as: The salinity residual (difference value) is calculated as: The residual here is the difference value, and its value and positive or negative direction indicate the deviation between the prediction model and the actual observation. For example, the temperature difference value indicates that the temperature prediction value of the prediction model at the grid point is higher than the actual observation value. Store these difference values to form a difference value field with the same dimensions as the aligned observation field and the prediction field. For example, at the grid point, the temperature difference value is , and the salinity difference value is . This difference value represents the degree of systematic deviation of the ocean prediction model in the target region and time window.

[0075] Store the difference value according to its corresponding spatiotemporal coordinate information in the rolling model bias database. The rolling model bias database is a key-value storage device based on spatiotemporal four-dimensional grid index. First, extract the spatial latitude and longitude coordinates, vertical depth level, and observation timestamp corresponding to the difference value. For example, for the difference values (temperature) and (salinity) calculated in the previous step, the corresponding spatial latitude and longitude coordinates are , the vertical depth level is meters, and the observation timestamp is . Combine the spatial latitude and longitude coordinates, vertical depth level, and observation timestamp to generate a four-dimensional spatiotemporal index key. This index key can be represented as a string, for example, N19.5E114.3D50T202405150300. Store the difference value itself as a numerical value associated with the four-dimensional spatiotemporal index key in the rolling model bias database. For example, under the key N19.5E114.3D50T202405150300, store the temperature difference value and the salinity difference value .

[0076] A rolling time window is set. The length of the rolling time window is 30 days, meaning that only the difference data in the last 30 days is retained in the database. The basis for setting this 30-day window is based on the analysis of the time scale of the marine model bias. Through long-term statistics of historical model prediction errors, it is found that most model device biases have good statistical stability within 30 days, and their correlation decreases significantly beyond 30 days. For example, by performing autocorrelation analysis on the prediction errors of the northern South China Sea region in 2023, the error correlation coefficients under different time lags are calculated, and the results are shown in Table 1.

[0077] Table 1 Autocorrelation coefficient table of prediction error

[0078] Time lag (days) Temperature error correlation coefficient Salinity error correlation coefficient 1 0.92 0.88 5 0.85 0.80 10 0.76 0.72 20 0.65 0.60 30 0.52 0.48 45 0.38 0.35 60 0.25 0.22

[0079] As shown in Table 1, when the time lag exceeds 30 days, the correlation coefficients of temperature and salinity errors both decrease to below 0.5, indicating that the effectiveness of historical bias information decreases. Therefore, 30 days is selected as the length of the rolling time window. When the current difference value is stored, for example, the difference value at 03:00 UTC on May 15, 2024, the historical difference values with timestamps earlier than the rolling time window in the rolling model bias database are retrieved and deleted simultaneously. The current timestamp is 03:00 UTC on May 15, 2024, and the start time of the rolling time window is 03:00 UTC on April 15, 2024 (i.e., the current time minus 30 days). All four-dimensional spatiotemporal index keys in the rolling model bias database are retrieved, and if the observation timestamp included in the index key is earlier than 03:00 UTC on April 15, 2024, the corresponding four-dimensional spatiotemporal index key and its associated difference value are deleted from the rolling model bias database. For example, if there is a key N19.5E114.3D50T202404140000 (timestamp is 00:00 UTC on April 14, 2024) in the database, this time is earlier than the start time of the rolling window 03:00 UTC on April 15, 2024, so the key and its associated temperature and salinity difference values are deleted. Through this dynamic updating mechanism, it is ensured that the rolling model bias database always contains the latest and effective model bias information to support the subsequent bias correction process.

[0080] See Figure 1 and Figure 3 , S2: obtaining the original temperature and salinity prediction field, extracting the expected bias field from the rolling model bias database, generating a set of debiased candidate state fields by subtracting the expected bias field from the original temperature and salinity prediction field, and fusing the set of debiased candidate state fields to generate a candidate integrated temperature and salinity field;

[0081] The steps of S2 are specifically:

[0082] S21: Obtain an original temperature-salinity prediction field, retrieve and extract a corresponding set of historical difference values from the rolling model bias database according to the spatiotemporal coordinate range of the original temperature-salinity prediction field, and generate an expected bias field by performing spatiotemporal weighted averaging calculation on the set of historical difference values;

[0083] S22: Use the original temperature-salinity prediction field as one of the members, and generate a plurality of perturbed prediction fields by perturbing the original temperature-salinity prediction field multiple times using a random perturbation field, subtract the expected bias field from the original temperature-salinity prediction field and the plurality of perturbed prediction fields one by one to generate a set of debiased candidate state fields;

[0084] S23: Initialize an integrated weight matrix, which represents the weights of the plurality of member fields in the set of debiased candidate state fields, and perform weighted fusion on all member fields in the set of debiased candidate state fields according to the integrated weight matrix to generate a candidate integrated temperature-salinity field;

[0085] The process of weighted fusion calculation on the set of debiased candidate state fields in S23 includes:

[0086] Obtain the integrated weight matrix, which stores the current weight coefficients of each member field in the set of debiased candidate state fields;

[0087] Iterate through each grid point in the candidate integrated temperature-salinity field, and extract the temperature and salinity values of all member fields in the set of debiased candidate state fields at the grid point;

[0088] Perform product-sum operation on the temperature and salinity values with the corresponding weight coefficients in the integrated weight matrix, and take the operation result as the final temperature and salinity values of the candidate integrated temperature-salinity field at the grid point.

[0089] Obtain the original temperature-salinity prediction field, which is generated by a marine numerical model for a future time (e.g., May 16, 2024, 00:00 UTC), covering the northern South China Sea region (18.0 degrees north latitude to 22.0 degrees north latitude, 112.0 degrees east longitude to 116.0 degrees east longitude), with a horizontal resolution of 0.1 degrees x 0.1 degrees and 50 standard vertical depth layers. For example, at grid point (18.0 degrees north latitude , 112.0 degrees east longitude , depth , time ), the original predicted temperature is , and the original predicted salinity is .

[0090] Retrieve and extract a corresponding set of historical difference values from the rolling model bias database according to the spatiotemporal coordinate range of the original temperature-salinity prediction field. For grid point , retrieve all historical difference values corresponding to the same grid point in the rolling model bias database. Historical difference values that are spatially and temporally close. Spatial closeness can be defined as horizontal distance less than 0.5 degree and vertical depth level difference less than 20 meters. Temporal closeness can be defined as time stamp within a rolling time window (e.g. past 30 days). For example, the following historical temperature difference values are extracted: 1. ): 2.( ): 3.( ): 4.( ):

[0091] The expected bias field is generated by spatio-temporal weighted average calculation on the set of historical difference values. The purpose of the weighted average is to give more weight to the historical biases that are closer in distance to the current forecast point and closer in time. The weight function can be defined as the product of spatial weight and temporal weight. The spatial weight can be a Gaussian function, for example where is the horizontal distance between the historical bias point and the current forecast point, is the vertical distance between the historical bias point and the current forecast point, is the horizontal spatial scale parameter, is the vertical spatial scale parameter. The horizontal spatial scale parameter is set to 0.5 degree, and the vertical spatial scale parameter is set to 20 meters. The temporal weight can also be a Gaussian function, for example where is the time difference between the historical bias time and the current forecast time, is the time scale parameter. The time scale parameter is set to 30 days. The total weight of all historical difference points is . The expected bias .

[0092] The expected temperature bias at grid point is calculated. Historical difference point 1: 0.5 degree, 20 meters, 1.125 days. The total weight

[0093] Historical difference point 2: 0.5 degree. 20 meters. 1.75 days. The total weight

[0094] Historical difference point 3: degrees. meters. hours (2.5 days). Total weight

[0095] Historical difference point 4: degrees, meters, days (624 hours). Total weight

[0096] The expected temperature bias is calculated by weighted average: The expected salinity bias is calculated in the same way, resulting in By repeating this process for all grid points of the original temperature-salinity forecast field, the expected bias field covering the entire forecast area and time window is generated.

[0097] The formula for the above expected bias field is ; where, represents the physical value of the expected bias field; represents the physical value of the th historical difference point; represents the influence weight of the th historical difference point on the current forecast point, and its calculation formula is ; represents the spatial weight of the th historical difference point, and its calculation formula is ; represents the temporal weight of the th historical difference point, and its calculation formula is ; represents the horizontal distance between the th historical difference point and the current forecast point; represents the vertical distance between the th historical difference point and the current forecast point; represents the horizontal spatial scale parameter, which is used to control the decay rate of the horizontal distance on the spatial weight; represents the vertical spatial scale parameter, which is used to control the decay rate of the vertical distance on the spatial weight; represents the time difference between the th historical difference point and the current forecast time; represents the time scale parameter, which is used to control the decay rate of the time difference on the temporal weight; represents the natural logarithm function.

[0098] The original temperature-salinity prediction field is taken as one of the members, as the benchmark member of the integrated prediction. A plurality of perturbed prediction fields are generated by perturbing the original temperature-salinity prediction field using random perturbation fields. The generation of the perturbation field is as follows: at each grid point of the original prediction field, a random number is added, which is subject to a normal distribution with a mean of 0 and a standard deviation of a certain value. The setting of the standard deviation is based on the statistical analysis of the historical prediction error, for example, the standard deviation of the temperature perturbation is set to , and the standard deviation of the salinity perturbation is set to . This standard deviation is obtained by analyzing the residual distribution of the long-term (for example, the past year) prediction model to calculate the spatial average standard deviation. For example, 5 perturbed prediction fields are generated. The temperature , and the salinity of the original prediction field at the grid point . Five random numbers are generated:

[0099] The temperature at of the first perturbed prediction field is , and the salinity is . The temperature at of the second perturbed prediction field is , and the salinity is . In this way, a total of 6 member fields (1 original prediction field and 5 perturbed prediction fields) are generated.

[0100] The expected bias field is subtracted from the original temperature-salinity prediction field and the plurality of perturbed prediction fields one by one to generate a set of debiased candidate state fields. At the grid point , the expected temperature bias , and the expected salinity bias . The temperature of the original prediction field after debiasing is . The salinity of the original prediction field after debiasing is . The temperature of the first perturbed prediction field after debiasing is . The salinity of the first perturbed prediction field after debiasing is . In this way, the process is repeated for each perturbed prediction field to generate a set of debiased candidate state fields containing 6 member fields.

[0101] The integrated weight matrix is initialized. The integrated weight matrix is a matrix, where is the number of member fields, is the number of grid points in the prediction area. The integrated weight matrix represents the weight of the plurality of member fields in the set of debiased candidate state fields. Initially, the weight of each member field is set to be equal, for example, if there are 6 member fields, the initial weight coefficient of each member field is . These weight coefficients are filled into the integrated weight matrix so that for each grid point, the sum of the weights of all member fields is 1. For example, the integrated weight matrix

[0102] The candidate integrated temperature-salinity field is generated by weighted fusion of all member fields in the debiased candidate state field set according to the integrated weight matrix. The integrated weight matrix stores the current weight coefficients of each member field in the debiased candidate state field set. Each grid point in the candidate integrated temperature-salinity field is traversed. For example, at grid point The temperature and salinity values of all member fields in the debiased candidate state field set at grid point are extracted. Assuming that the debiased candidate state field set contains 6 member fields, the temperature values of the member fields at grid point are as follows: The weight coefficients of the initial integrated weight matrix are The temperature and salinity values are multiplied by the corresponding weight coefficients in the integrated weight matrix, respectively, and the results are summed to obtain the final temperature and salinity values of the candidate integrated temperature-salinity field at grid point The candidate integrated temperature-salinity field covering the entire forecast area is obtained by repeating this process for all grid points.

[0103] See Figure 1 and Figure 4 ​​​​​​​​​​​​​​​​​​​​​​​​​​S3: Calculate the grid point density value in the candidate integrated temperature-salinity field by using the coprocessor hardware, and generate a binary physical inconsistency mask by comparing the vertically adjacent density values;

[0104] The steps of S3 are specifically:

[0105] S31: Load the temperature, salinity and pressure data included in the candidate integrated temperature-salinity field into the dedicated video memory of the coprocessor hardware, and use the parallel computing cores of the coprocessor hardware to calculate the grid point density value of each grid point in the candidate integrated temperature-salinity field according to the internationally recognized ocean state equation:

[0106] ;

[0107] Parallelly calculate the grid point density value of each grid point in the candidate integrated temperature-salinity field to generate a three-dimensional density field;

[0108] wherein, represents the grid point density value, represents the temperature value of the grid point in the candidate integrated temperature-salinity field, represents the salinity value of the grid point in the candidate integrated temperature-salinity field, represents the pressure value of the grid point in the candidate integrated temperature-salinity field;

[0109] S32: In the three-dimensional density field, for each grid point on the non-water bottom boundary, extract its own grid point density value and the grid point density value of the vertically adjacent grid point below it, i.e. the density value, and form a to-be-compared density pair with the two density values;

[0110] S33: Traverse all to-be-compared density pairs, if the density value of the grid point above is greater than the density value of the vertically adjacent grid point below, determine that the grid point above is an unstable grid point and mark it as 1, otherwise mark it as 0, and collect the marking results of all grid points to generate a binary physical inconsistency mask;

[0111] The coprocessor hardware is a graphics processor, and the process of calculating the grid point density value by using the parallel computing cores of the coprocessor hardware includes:

[0112] Construct the temperature, salinity and pressure data of the candidate integrated temperature-salinity field into a three-dimensional texture or a unified memory buffer for the coprocessor hardware to grab data when performing parallel calculation;

[0113] Write and compile a calculation kernel or shader program for the architecture of the coprocessor hardware, and the calculation kernel or shader program has the complete calculation logic of the ocean state equation fixed therein;

[0114] Start the calculation kernel or shader program, use the thousands of parallel processing units of the coprocessor hardware to calculate the grid point density value of all grid points in the three-dimensional density field at the same time, and write the calculation results back to the dedicated video memory of the coprocessor hardware.

[0115] The temperature, salinity, and pressure data included in the candidate integrated temperature and salinity field are loaded into the dedicated video memory of the coprocessor hardware. For example, the candidate integrated temperature and salinity field covers the northern South China Sea (18.0°N to 22.0°N, 112.0°E to 116.0°E) with a horizontal grid resolution of 0.1° x 0.1° (i.e., There are 100 horizontal grid points and 50 standard depth layers in the vertical direction. Therefore, there are a total of 1000 horizontal grid points and 50 standard depth layers in the vertical direction. There are 1,000 grid points. For each grid point, there is a temperature. (unit: ),salinity (unit: and pressure (unit: These data are in the form of three-dimensional arrays (e.g.) The data is organized and transferred to the dedicated video memory of the graphics processing unit (GPU). The temperature, salinity, and pressure data of the candidate integrated temperature-salinity field are constructed as a 3D texture or a unified memory buffer for the coprocessor hardware to fetch the data while performing parallel computations. For example, temperature data is stored in the GPU global memory in the form of floatT

[41]

[41]

[50] .

[0116] Utilizing the parallel computing core of coprocessor hardware, based on internationally accepted ocean state equations... The grid point density value of each grid point in the candidate integrated temperature-salinity field is calculated in parallel to generate a three-dimensional density field. The international standard equation of state recommended by TEOS-10 (Thermodynamic Equation of Seawater – 2010) is adopted here. A computational kernel program for the coprocessor hardware architecture is written and compiled, which embeds the ocean equation of state. The complete computational logic is as follows. The computational kernel function takes the longitude, latitude, depth (for pressure calculation), temperature, and salinity of a grid point as input, calculates and returns the seawater density at that point. Specifically, the equation requires converting temperature and salinity to absolute temperature and reference salinity, and considering the effect of pressure. To start the computational kernel program, for example, configure GPU startup. Each thread is responsible for calculating the density value of one grid point. Utilizing thousands of parallel processing units in the coprocessor hardware, the grid point density values ​​of all grid points in the 3D density field are calculated simultaneously, and the results are written back to the dedicated video memory of the coprocessor hardware. For example, at grid points... (north latitude East longitude ,depth Its candidate integration temperature is Salinity is .pressure It is calculated based on depth, for example, The pressure corresponding to a depth of meters is approximately Within the GPU's computing core, the following steps are performed to calculate density. 1. To and As input. 2. Calculate the density using the TEOS-10 library function gsw_rho(SA,CT,p), where... It is absolute salinity. It is a conservative temperature. It's pressure. 3. First, [the following is a separate, unrelated sentence:] Convert to absolute salinity Assume the reference latitude for this region is... Longitude is This is calculated using the `gsw_SA_from_SP(SP,p,long,lat)` function. For example, 4. Convert to conservative temperature Calculated using the `gsw_CT_from_t(SA,t,p)` function. For example, 5. Finally, calculate the density. This density value Grid points The density values ​​of all grid points are calculated and stored in the video memory through parallel computation, forming a complete three-dimensional density field.

[0117] In the generated 3D density field, for each grid point outside the underwater boundary, extract its own grid point density value and the grid point density values ​​of its vertically adjacent grid points. For example, consider a vertical profile at a specific latitude and longitude (19.5 degrees North, 114.3 degrees East). This profile contains 50 depth layers, from the surface (depth...) ) to the deepest layer (depth) For depth layers Its grid point density value is The grid points directly below it represent the depth layer. Its density value is For example, for grid points (north latitude East longitude ,depth Its density value Its vertically adjacent grid point is (north latitude East longitude ,depth ). Assumption The density value is These two density values ​​form a density pair to be compared: For deep layers Grid points that have no adjacent grid points below them (or are underwater boundaries) are not included in the comparison range. The device traverses all grid points that are not underwater boundaries, extracts the density values ​​of the grid point itself and its vertically adjacent grid points, and generates a series of density pairs to be compared. This process is also executed in parallel on the coprocessor hardware, with each thread responsible for extracting the density pair of one grid point.

[0118] Iterate through all the density pairs to be compared. This is a parallel process executed on coprocessor hardware, with each thread processing one density pair. For each density pair to be compared... If the density value of the grid points located above The density value is greater than the density value of the grid point directly below it. If so, the grid points above are determined to be unstable grid points and marked as 1. For example, if and ,but This point is marked as 0 (stable). If another density pair is... ,in , ,at this time Then depth Grid points at a certain location are considered unstable and marked with a value of 1. Otherwise, they are marked with a value of 0. For example, when... When the water body is in a stable or neutral state, the point is marked as 0. The marking results of all grid points are then aggregated to generate a binary physical inconsistency mask. This mask is a binary matrix with the same dimensions as the three-dimensional density field, where each element is either 0 or 1, representing the physical stability state of the corresponding grid point. For example, for all grid points... Its corresponding mask value is This mask will be written back to the dedicated video memory of the coprocessor hardware for access and processing in subsequent steps.

[0119] Please see Figure 1 and Figure 5 S4: Based on the binary physical inconsistency mask, filter unstable grid points in the candidate integrated temperature and salinity field, adjust the integrated weight matrix, iteratively update the candidate integrated temperature and salinity field and the binary physical inconsistency mask until the number of unstable grid points in the binary physical inconsistency mask is less than the threshold, and output the integrated prediction result.

[0120] The specific steps for S4 are as follows:

[0121] S41: Based on the binary physical inconsistency mask, filter and count the current total number of all unstable grid points marked as 1 in the candidate integrated temperature and salinity field;

[0122] S42: Determine whether the current total number of unstable grid points is less than the preset threshold. If it is less, determine that the current candidate integrated temperature and salinity field has reached the physical consistency requirement, terminate the iteration, and output the current candidate integrated temperature and salinity field as the integrated prediction result.

[0123] S43: If the current total number of unstable grid points is not less than the threshold, then a penalty function is applied based on the spatial distribution and inconsistency of the unstable grid points:

[0124] ;

[0125] Adjust the integrated weight matrix to reduce the weights of member fields in the debiased candidate state field set that leads to physical instability;

[0126] in, This represents the adjusted integrated weight matrix. This represents the integrated weight matrix before adjustment. The Hadamard product represents the product of corresponding elements in a matrix. Representative and A matrix of all one dimensions This represents the penalty coefficient used to control the magnitude of weight adjustments. A matrix representation of a binary physical inconsistency mask;

[0127] S44: Based on the adjusted integrated weight matrix, re-fuse the debiased candidate state field set and update the candidate integrated temperature and salinity field;

[0128] The grid point density values ​​of the candidate integrated temperature and salinity field are recalculated using coprocessor hardware, and a new binary physical inconsistency mask is generated. Then, return to S41 to execute the next iteration.

[0129] The rolling model deviation database includes difference values, the candidate integrated temperature and salinity field includes grid point density values, the binary physical inconsistency mask includes unstable grid points, and the integrated forecast result specifically refers to the candidate integrated temperature and salinity field after iterative updates.

[0130] Based on a binary physical inconsistency mask, the current total number of unstable grid points marked as 1 in the candidate integrated temperature-salinity field is filtered and counted. For example, assuming that after step S3, the coprocessor hardware generates a binary physical inconsistency mask. Its dimensions are (Because the bottom layer has no grid points vertically below it, the vertical depth is 50-1=49). Now, through the parallel reduction operation of the coprocessor hardware, the number of all grid points marked as 1 is counted. Each thread checks the mask element it is responsible for. If the value is 1, the counter is incremented by 1. Finally, the counts of all threads are summed to obtain the current total number of unstable grid points. For example, statistical results show that among the candidate integrated thermo-salinity fields in the entire northern South China Sea region, there are a total of... Unstable grid points.

[0131] Determine if the current total number of unstable grid points is less than a preset threshold. Preset threshold Set to 0.1% of the total number of grid points, that is Rounded down to The threshold is set based on empirical and physical oceanographic requirements for model stability. Completely eliminating gravitational instability is difficult in ocean numerical models, but a small number of localized, weakly unstable regions are permissible. Setting the threshold to 0.1% indicates that we accept a very small proportion of unstable points.

[0132] Compare (The total number of unstable grid points currently counted) and (Preset threshold). In this example, ,and . The judgment result is "No" ( Not less than Therefore, the current candidate integrated temperature and salt field has not yet met the physical consistency requirements.

[0133] If the current total number of unstable grid points is not less than the threshold (as in this example) Then, based on the spatial distribution and inconsistency of the unstable grid points, a penalty function is applied to adjust the integrated weight matrix. The penalty function is... In this formula, This represents the adjusted integrated weight matrix. This represents the integrated weight matrix before adjustment. The Hadamard product represents the product of corresponding elements in a matrix. Representative and A matrix of all one dimensions This represents the penalty coefficient used to control the magnitude of weight adjustments. A matrix representation of a binary physical inconsistency mask.

[0134] Penalty coefficient Used to control the magnitude of weight adjustment, its value ranges from 0 to 1, for example, setting... This value is set based on a balance between iterative convergence speed and the improvement in physical consistency. Experiments have verified that when... In this case, the device can converge within a relatively small number of iterations.

[0135] The specific adjustment process is as follows: for each grid point and each member's field Its adjusted weight The calculation is as follows: If grid points It is a stable point, that is ,but The weights remain unchanged. If the grid points It is an unstable point, that is ,but This means that at this unstable grid point, the weights of all member fields are reduced by 20%. This reduces the weights of member fields in the set of debiased candidate state fields that lead to physical instability. For example, after the first iteration, at a certain grid point... Marked as unstable ( Assume that at this grid point, the current weight of member field 1 (original field debiased) is... After adjustment, its new weight is: The weights of all member fields are similarly adjusted at this unstable point, and the sum of all weights needs to be renormalized at each grid point. For example, if a certain unstable grid point... The weights of all 6 member fields were changed from Become The normalized weights are: This normalization ensures that at each grid point, the sum of the weights of all member fields remains 1.

[0136] The advantage of the formula lies in the introduction of a binary physical inconsistency mask. and penalty coefficient This formula enables adaptive adjustment of the ensemble weight matrix. When a local physical instability is detected, the formula directly and quantitatively reduces the contribution of the weights of all member fields that caused the instability to that unstable region. This mechanism allows the ensemble forecasting device to actively learn and correct its physical inconsistencies.

[0137] Based on the adjusted ensemble weight matrix, the debiased candidate state field set is re-fused, and the candidate ensemble temperature-salinity field is updated. This process is the same as the weighted fusion process in S23, but uses the new weight matrix adjusted in S43. For example, for grid points After the first iteration, if it is not marked as unstable, its weight coefficients remain unchanged. If it is marked as unstable, its weights will be adjusted and normalized as described in S43. Assume that in the first iteration, the average weight of all grid points is... (After normalization). Candidate integrated temperature and salinity fields in Update temperature : Candidate integrated temperature and salinity field in Updated salinity :

[0138] The updated grid point density values ​​of the candidate integrated temperature-salinity field are recalculated using the coprocessor hardware, generating a new binary physical inconsistency mask. This process is identical to steps S31, S32, and S33. The updated temperature, salinity, and pressure data of the candidate integrated temperature-salinity field are then loaded back into the dedicated video memory of the coprocessor hardware. The coprocessor hardware calculates the density values ​​of all grid points in parallel, generating a new three-dimensional density field. Then, based on the new three-dimensional density field, a vertical density comparison is performed again to generate an updated binary physical inconsistency mask. For example, after the first iteration, if Still greater than If the condition is not met, return to S41 and continue with the next iteration. The entire iteration process will continue until the current total number of unstable grid points is less than or equal to a preset threshold. Until then. For example, after the 5th iteration, the total number of unstable grid points counted decreased to One, at this time If the judgment result is "yes", then it is determined that the current candidate integrated temperature and salinity field has met the physical consistency requirements, the iteration is terminated, and the current candidate integrated temperature and salinity field is output as the final integrated prediction result.

[0139] A data-driven near-real-time ocean observation rolling integrated forecasting device, used to execute the aforementioned data-driven near-real-time ocean observation rolling integrated forecasting method, the device comprising:

[0140] The deviation update module is configured to acquire near-real-time ocean observation data and forecast data from the previous period, calculate the difference between the near-real-time ocean observation data and the forecast data from the previous period, update the rolling model deviation database, and pass it to the temperature and salinity field correction module.

[0141] The temperature and salinity field correction module is configured to acquire the original temperature and salinity forecast field, extract the expected deviation field from the rolling model deviation database, subtract the expected deviation field from the original temperature and salinity forecast field to generate a debiased candidate state field set, fuse the debiased candidate state field set to generate a candidate integrated temperature and salinity field, and pass it to the physical consistency verification module.

[0142] The physical consistency verification module is configured to calculate the grid point density values ​​in the candidate integrated temperature and salinity field through coprocessor hardware, compare the vertically adjacent density values ​​to generate a binary physical inconsistency mask, and pass it to the integration weight iteration module.

[0143] The integrated weight iteration module is configured to filter unstable grid points in the candidate integrated temperature-M field through a binary physical inconsistency mask, adjust the integrated weight matrix, iteratively update the candidate integrated temperature-salt field and the binary physical inconsistency mask until the number of unstable grid points in the binary physical inconsistency mask is less than a threshold, and output the integrated prediction result.

[0144] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or device architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the claims of the present invention.

Claims

1. A data-driven rolling ensemble forecasting method for near-real-time ocean observations, characterized in that, Includes the following steps: S1: Obtain near-real-time ocean observation data and forecast data from the previous period, calculate the difference between the near-real-time ocean observation data and the forecast data from the previous period, and update the rolling model bias database; S2: Obtain the original temperature and salinity forecast field, extract the expected deviation field from the rolling model deviation database, subtract the expected deviation field from the original temperature and salinity forecast field to generate a debiased candidate state field set, and fuse the debiased candidate state field set to generate a candidate integrated temperature and salinity field. S3: Calculate the grid point density values ​​in the candidate integrated temperature and salinity field through coprocessing hardware, and generate a binary physical inconsistency mask by comparing the vertically adjacent density values. S4: Based on the binary physical inconsistency mask, filter unstable grid points in the candidate integrated temperature and salinity field, adjust the integrated weight matrix, iteratively update the candidate integrated temperature and salinity field and the binary physical inconsistency mask until the number of unstable grid points in the binary physical inconsistency mask is less than the threshold, and output the integrated prediction result.

2. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 1, characterized in that, The rolling model deviation database includes difference values, the candidate integrated temperature and salinity field includes grid point density values, the binary physical inconsistency mask includes unstable grid points, and the integrated forecast result specifically refers to the candidate integrated temperature and salinity field after iterative updates.

3. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Acquire near-real-time ocean observation data and forecast data from the previous period, perform spatiotemporal interpolation on the near-real-time ocean observation data to make its grid structure consistent with the grid structure of the forecast data from the previous period, and generate an aligned observation field; S12: Calculate the residuals between the aligned observation field and the forecast data of the previous period at multiple corresponding grid points, and define the residuals as the difference values, which characterize the degree of device-related offset of the ocean forecast model in the target area and time window. S13: Store the difference value into the rolling model deviation database according to its corresponding spatiotemporal coordinate information, and set a rolling time window. When storing the current difference value, synchronously retrieve and remove historical difference values ​​in the rolling model deviation database whose timestamps are earlier than the rolling time window, and dynamically update the rolling model deviation database.

4. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 1, characterized in that, The specific steps of S2 are as follows: S21: Obtain the original temperature and salinity forecast field; based on the spatiotemporal coordinate range of the original temperature and salinity forecast field, retrieve and extract the corresponding historical difference value set from the rolling model deviation database; and generate the expected deviation field by performing spatiotemporal weighted average calculation on the historical difference value set. S22: Take the original temperature and salinity forecast field as one of the members, and use the random perturbation field to perturb the original temperature and salinity forecast field multiple times to generate multiple perturbation forecast fields. Subtract the expected deviation field from the original temperature and salinity forecast field and the multiple perturbation forecast fields one by one to generate a set of debiased candidate state fields. S23: Initialize the integrated weight matrix, which represents the weights of multiple member fields in the debiased candidate state field set. Based on the integrated weight matrix, perform weighted fusion on all member fields in the debiased candidate state field set to generate a candidate integrated temperature-salinity field.

5. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 1, characterized in that, The specific steps in S3 are as follows: S31: Load the temperature, salinity, and pressure data included in the candidate integrated temperature and salinity field into the dedicated video memory of the coprocessor hardware, and utilize the parallel computing core of the coprocessor hardware according to the internationally accepted ocean state equation: ; The grid point density value of each grid point in the candidate integrated temperature and salinity field is calculated in parallel to generate a three-dimensional density field. in, Represents the grid point density value. The temperature values ​​of grid points in the candidate integrated temperature-salinity field are represented. This represents the salinity values ​​of grid points in the candidate integrated temperature and salinity field. The pressure values ​​at grid points in the candidate integrated temperature-salinity field; S32: In the three-dimensional density field, for each grid point that is not at the bottom boundary, extract its own grid point density value and the grid point density value of its vertically adjacent grid point, i.e., the density value, and form a density pair to be compared between the two density values. S33: Traverse all the density pairs to be compared. If the density value of the grid point above is greater than the density value of its vertically adjacent grid point below, then the grid point above is determined to be an unstable grid point and marked as 1. Otherwise, it is marked as 0. Summarize the marking results of all grid points to generate a binary physical inconsistency mask.

6. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Based on the binary physical inconsistency mask, filter and count the current total number of all unstable grid points marked as 1 in the candidate integrated temperature and salinity field; S42: Determine whether the current total number of unstable grid points is less than the preset threshold. If it is less, determine that the current candidate integrated temperature and salinity field has reached the physical consistency requirement, terminate the iteration, and output the current candidate integrated temperature and salinity field as the integrated prediction result. S43: If the current total number of unstable grid points is not less than the threshold, then a penalty function is applied based on the spatial distribution and inconsistency of the unstable grid points: ; Adjust the integrated weight matrix to reduce the weights of member fields in the debiased candidate state field set that cause physical instability; in, This represents the adjusted integrated weight matrix. This represents the integrated weight matrix before adjustment. The Hadamard product represents the product of corresponding elements in a matrix. Representative and A matrix of the same dimension that is entirely identical. This represents the penalty coefficient used to control the magnitude of weight adjustments. A matrix representation of the binary physical inconsistency mask; S44: Based on the adjusted integrated weight matrix, re-fuse the debiased candidate state field set and update the candidate integrated temperature and salinity field; The grid point density values ​​of the candidate integrated temperature and salinity field are recalculated using the coprocessor hardware, and a new binary physical inconsistency mask is generated. Then, the process returns to S41 to execute the next iteration.

7. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 5, characterized in that, The coprocessor hardware is a graphics processing unit (GPU), and the process of calculating the grid point density value using the parallel computing core of the coprocessor hardware includes: The temperature, salinity, and pressure data of the candidate integrated temperature and salinity fields are constructed into a three-dimensional texture or a unified memory buffer for the coprocessing hardware to retrieve data during parallel computing. Write and compile a computation kernel or shader program for the coprocessor hardware architecture, wherein the ocean state equation is embedded in the computation kernel or shader program. The complete computational logic; The computational kernel or the shader program is started, and the thousands of parallel processing units of the coprocessor hardware are used to calculate the grid point density value of all grid points in the three-dimensional density field at the same time, and the calculation results are written back to the dedicated video memory of the coprocessor hardware.

8. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 3, characterized in that, The rolling model deviation database is a key-value storage device based on a spatiotemporal four-dimensional grid index. The process of updating the rolling model deviation database in S13 includes: Extract the spatial latitude and longitude coordinates, vertical depth level, and observation timestamp corresponding to the difference value, and combine the spatial latitude and longitude coordinates, the vertical depth level, and the observation timestamp to generate a four-dimensional spatiotemporal index key; The difference value itself is stored in the rolling model deviation database as a value associated with the four-dimensional spatiotemporal index key; A rolling time window is set. During each update, all four-dimensional spatiotemporal index keys in the rolling model deviation database are retrieved. If the observation timestamp included in the four-dimensional spatiotemporal index key is earlier than the start time of the rolling time window, the corresponding four-dimensional spatiotemporal index key and its associated difference value are deleted from the rolling model deviation database.

9. The data-driven near-real-time ocean observation rolling integrated forecasting method according to claim 4, characterized in that, The process of performing weighted fusion calculation on the debiased candidate state field set in S23 includes: Obtain the integrated weight matrix, which stores the current weight coefficients of each member field in the debiased candidate state field set; Traverse each grid point in the candidate integrated temperature and salinity field and extract the temperature and salinity values ​​of all member fields in the debiased candidate state field set at the grid point. The temperature and salinity values ​​are multiplied and summed with the corresponding weight coefficients in the integrated weight matrix, and the results are used as the final temperature and salinity values ​​of the candidate integrated temperature-salinity field at the grid points.

10. A data-driven, near-real-time ocean observation and rolling integrated forecasting device, characterized in that, The apparatus is used to implement the data-driven near-real-time ocean observation rolling integrated forecasting method according to any one of claims 1-9, and the apparatus comprises: The deviation update module is configured to acquire near-real-time ocean observation data and forecast data from the previous period, calculate the difference between the near-real-time ocean observation data and the forecast data from the previous period, update the rolling model deviation database, and transmit it to the temperature and salinity field correction module. The temperature and salinity field correction module is configured to acquire the original temperature and salinity forecast field, extract the expected deviation field from the rolling model deviation database, subtract the expected deviation field from the original temperature and salinity forecast field to generate a debiased candidate state field set, fuse the debiased candidate state field set to generate a candidate integrated temperature and salinity field, and pass it to the physical consistency verification module. The physical consistency verification module is configured to calculate the grid point density values ​​in the candidate integrated temperature and salinity field through coprocessing hardware, compare the vertically adjacent density values ​​to generate a binary physical inconsistency mask, and pass it to the integration weight iteration module. The integrated weight iteration module is configured to filter unstable grid points in the candidate integrated temperature-m field through the binary physical inconsistency mask, adjust the integrated weight matrix, iteratively update the candidate integrated temperature-salt field and the binary physical inconsistency mask until the number of unstable grid points in the binary physical inconsistency mask is less than a threshold, and output the integrated prediction result.

Images