A high-time-frequency typhoon forecasting method, device, storage medium and processor

Abstract

This invention discloses a high-frequency typhoon forecasting method, device, storage medium, and processor, specifically relating to the field of high-frequency typhoon forecasting. The method includes acquiring a multi-source forecast dataset and a real-time dataset of the target typhoon; performing field unification and format conversion on the multi-source forecast dataset and removing data sources that do not meet the requirements of continuity and accuracy according to filtering rules; outputting a basic forecast sequence and a set of release time identifiers; performing high-frequency lag interpolation at the release time boundary on the edge computing side and generating a traceable interpolation interpretation record set which is then uploaded to the cloud side; and using an append-write method with both interpolation and correction layers to form a publishable forecast set, externally adjudicating visible values ​​and imposing a non-rollback restriction on adjudicated historical data.

Description

Technical Field

[0001] This invention relates to the field of high-frequency typhoon forecasting technology, and more specifically, to a high-frequency typhoon forecasting method, apparatus, storage medium, and processor. Background Technology

[0002] In the existing typhoon forecasting service, the common practice in the industry is to centrally aggregate multi-source forecasts and real-time data to the central side, generate forecast products according to a fixed release schedule, and interpolate the forecast time series within the release interval to solve the problem that users need continuous updates but the data release is not continuous. Taking the application in the near impact stage as an example, when a typhoon enters the 48-hour warning line, the business side often requires rolling updates of the forecast results every 15 minutes. At the same time, there are objectively minutes-level delays in the release of news bulletins or forecasts. This forces the system to continue to produce high-frequency forecast points during the delay period and form data files for external release as required. Moreover, these files usually require that each release retain the full history since the beginning of the typhoon for easy archiving, backtracking and verification. Under such hard constraints, once the mainstream practice introduces more realistic high-frequency location and intensity data or real-time data to correct the output high-frequency time-frequency forecasts, a stable and observable phenomenon will emerge: the historical forecast values ​​of the same typhoon at the same time point will repeatedly change in subsequent releases. The curves seen by downstream systems over the past fifteen minutes, thirty minutes, or one hour will keep jumping, making it impossible for the business side to reproduce the results of the previous release, and it is also difficult to explain the reasons for the differences. The fundamental reason is that corrections can usually only be completed by overwriting existing outputs, while full traceability requires that the history remain consistent and readable over a long period of time. Based on this, the technical problem that this application aims to solve is how to implement the correction processing of high-frequency time-frequency forecasts in the release link of edge computing and end-cloud collaboration under the condition that the forecast has a minute-level lag and needs to be updated at high frequency according to the warning line, while avoiding repeated changes in historical data and ensuring that the results of each release can be stably reproduced and verified. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a high-frequency typhoon forecasting method, apparatus, storage medium, and processor. By performing high-frequency lag interpolation at the release time boundary on the edge computing side and generating a traceable interpolation interpretation record set which is then sent to the cloud side, the cloud side forms a publishable forecast set by appending and writing in a manner that combines interpolation and correction layers, and decides visible values ​​externally while imposing a non-rollback restriction on the decided history.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a high-frequency typhoon forecasting method, comprising: S1. Obtain the multi-source forecast dataset and real-time dataset of the target typhoon. Perform field unification and format conversion on the multi-source forecast dataset and remove data sources that do not meet the requirements of continuity and accuracy according to the filtering rules. Output the basic forecast sequence and release time identifier set. S2. Read the basic forecast sequence and the release time identifier set on the edge computing side, perform time alignment on the basic forecast sequence according to the preset high time frequency time granularity, generate interpolated forecast points from adjacent valid forecast points for the uncovered time according to the lag interpolation rule, and output the interpolated forecast point set; S3. Generate evidence information for each point in the interpolated forecast point set. The evidence information includes the referenced input forecast point identifier, release time identifier, forecast time identifier, and lag time parameter. Then, organize the interpolated forecast point set and evidence information into an interpolation interpretation record set and send it to the cloud to complete end-cloud collaboration. S4. Receive the interpolation interpretation record set on the cloud side and write it into the interpolation layer according to the release time identifier to generate a publishable forecast set. The interpolation layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. S5. After obtaining the correction dataset on the cloud side and performing time alignment at a high time frequency granularity, it is matched point by point with the publishable forecast set to generate a correction record set and written into the correction layer to coexist with the interpolation layer. Then, the visible value adjudication is performed on the interpolation layer record and the correction layer record to output a visible forecast point set. The visible value corresponding to the adjudicated historical forecast time identifier is not allowed to be rolled back under the subsequent release time identifier.

[0005] In a preferred embodiment, S1 includes: S1-1. Obtain the source forecast sequences from each forecast data source, and extract the typhoon identifier, forecast time identifier, and release time field from the source forecast sequences. Output the source forecast sequence set carrying the release time field. S1-2. Sort each source forecast sequence in the source forecast sequence set according to the forecast time identifier and calculate the continuity index. The continuity index includes the time interval between adjacent forecast time identifiers and the length of the missing segment. When the continuity index exceeds the preset continuity constraint, output the removal flag and remove the corresponding source forecast sequence. Otherwise, output the continuous qualified forecast sequence and output the continuous qualified forecast sequence set. S1-3. For each source forecast sequence in the continuous qualified forecast sequence set, perform time alignment with the actual dataset according to the forecast time identifier and calculate the accuracy index. The accuracy index is whether the deviation after alignment is not greater than the preset accuracy boundary. When the accuracy index meets the preset accuracy boundary, perform field unification and format conversion on the source forecast sequence to generate candidate basic forecast sequences and write the release time field of the corresponding source forecast sequence into the release time identifier set. Otherwise, remove the corresponding source forecast sequence and output the basic forecast sequence and the release time identifier set.

[0006] In a preferred embodiment, S2 includes: S2-1. Read the basic forecast sequence and release time identifier set on the edge computing side, determine the upper limit of the available forecast time using the release time identifier set, and generate a target time raster that does not exceed the upper limit of the available forecast time according to the preset high time frequency time granularity, and output the target time raster. S2-2. Perform time alignment on the basic forecast sequence according to the target time grid. Mark the time without a corresponding forecast point in the target time grid as an uncovered time. For each uncovered time, select the previous and next valid forecast points in the basic forecast sequence that are time-adjacent as adjacent valid forecast points. Output the set of uncovered times and the pair of adjacent valid forecast points. S2-3. For each uncovered time point, perform interpolation calculation according to the lag interpolation rule. The lag interpolation rule includes determining the lag lead time parameter based on the release time identifier and restricting the interpolation calculation to use only adjacent valid forecast points that are no later than the upper limit of the available forecast time corresponding to the release time identifier, and outputting the interpolated forecast point set.

[0007] In a preferred embodiment, S3 includes: S3-1. On the edge computing side, extract the adjacent valid forecast points from the interpolation forecast point set point by point to generate the adjacent valid forecast points used, and use the sequence index identifier of the adjacent valid forecast points in the basic forecast sequence as the input forecast point identifier, and output the correspondence between the interpolation forecast points and the input forecast point identifiers. S3-2. Write the release time identifier, forecast time identifier, and lag time parameter for each interpolated forecast point, and use the typhoon identifier and forecast time identifier as record identifiers. Combine the interpolated forecast points with the corresponding input forecast point identifier, release time identifier, forecast time identifier, and lag time parameter to generate evidence information and output the evidence information set. S3-3. Bind the interpolation forecast point set and the evidence information set at the record level according to the typhoon identifier and the forecast time identifier, organize them into an interpolation interpretation record set, and send the interpolation interpretation record set to the cloud side through the end-to-cloud collaborative link, and output the sending result.

[0008] In a preferred embodiment, S4 includes: S4-1. Receive the interpolation interpretation record set sent from the edge computing side via the end-cloud collaborative link on the cloud side, group the interpolation interpretation record set according to the release time identifier and the typhoon identifier, and output the release group set; S4-2. Generate an interpolation layer write sequence for each release group in the release group set. The interpolation layer write sequence is arranged in ascending order according to the forecast time identifier and a write identifier is generated for each interpolation interpretation record. Output the interpolation layer write sequence. S4-3. Perform an append write judgment on the interpolation layer write sequence. When there is no interpolation layer record corresponding to the write identifier in the cloud-side interpolation layer, append the corresponding interpolation interpretation record to the interpolation layer and output a write success flag. When there is already an interpolation layer record corresponding to the write identifier in the cloud-side interpolation layer, keep the existing interpolation layer record unchanged and output a rejection overwrite flag. S4-4. After the interpolation layers have been appended, they are aggregated according to the release time identifier to generate a publishable forecast set. The publishable forecast set consists of interpolation layer records corresponding to the same release time identifier. The publishable forecast set is then output.

[0009] In a preferred embodiment, S5 includes: S5-1. Obtain the correction dataset of the target typhoon on the cloud side, and perform time alignment on the correction dataset at a high time frequency granularity to generate a correction point set. S5-2. Perform point-by-point matching between the correction point set and the publishable forecast set according to the typhoon identifier and the forecast time identifier. Generate correction records for points with matching relationships and generate empty correction records for points without matching relationships. Output the correction record set. S5-3. Write the correction record set into the correction layer storage area according to the release time identifier so that it can coexist with the interpolation layer storage area. The correction layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. Output the correction layer writing result. S5-4. Align the interpolation layer records corresponding to the release time identifier in the interpolation layer with the correction layer records corresponding to the release time identifier in the correction layer according to the forecast time identifier, and output the record pair to be decided. S5-5. Perform visible value adjudication for each pair of records to be adjudicated. When the correction layer record is a non-empty correction record, output the forecast value corresponding to the correction layer record as the visible value. When the correction layer record is an empty correction record, output the forecast value corresponding to the interpolation layer record as the visible value. Output the set of visible forecast points. S5-6. Perform historical rollback restriction judgment on the visible forecast point set. When there is a historical release time identifier corresponding to the adjudicated visible value that is earlier than the release time identifier on the cloud side, keep the adjudicated visible value unchanged and only output the visible value for the forecast time identifier that has not yet been adjudicated. Output the visible forecast point set without rollback.

[0010] In a preferred embodiment, the process of outputting the set of visible forecast points in S5 further includes: S5-51. Obtain the interpolation layer record and correction layer record corresponding to the same typhoon identifier and the same forecast time identifier, and extract the forecast value, evidence information and release time identifier from them. Construct a decision problem with the forecast value consistency cost and evidence integrity cost as the objective function and the upper limit of available information corresponding to the release time identifier as the constraint set. Output the decision objective function and constraint set. S5-52. Under the adjudication problem, calculate the confidence scores for the interpolation layer records and the correction layer records respectively, and perform confidence updates. The confidence update uses the input forecast point identifier and lag time parameter referenced in the evidence information to perform attenuation correction on the confidence score of the interpolation layer records, and uses the time alignment residual and the reality consistency residual of the correction layer records to perform convergence correction on the confidence score of the correction layer records. Output the interpolation confidence score and the correction confidence score. S5-53. Perform multi-view consistency verification on the interpolation confidence and correction confidence and output conflict flags. The multi-view consistency verification includes at least performing continuity constraint verification on the interpolation layer records and correction layer records with the adjudicated visible values ​​corresponding to adjacent forecast time identifiers and calculating the continuity default amount. S5-54. For records to be decided with a conflict mark of no, output visible values ​​according to the non-empty priority rule of the correction layer records. For records to be decided with a conflict mark of yes, perform constraint solving to select visible values ​​by minimizing the objective function and satisfying the constraint set as the stopping condition, and output the set of visible forecast points.

[0011] A high-frequency typhoon forecasting device includes a multi-source screening module, a high-frequency supplementation module, an evidence solidification and uploading module, an interpolation layer addition and retention module, and a decision-making and non-return module. The multi-source filtering module is used to acquire the multi-source forecast dataset and the real-time dataset of the target typhoon. It performs field unification and format conversion on the multi-source forecast dataset and removes data sources that do not meet the requirements of continuity and accuracy according to the filtering rules. It outputs the basic forecast sequence and the set of release time identifiers. The high-frequency interpolation module is used to read the basic forecast sequence and the release time identifier set on the edge computing side, perform time alignment on the basic forecast sequence according to the preset high time frequency granularity, generate interpolated forecast points from adjacent valid forecast points for the uncovered time according to the lag interpolation rule, and output the interpolated forecast point set; The evidence solidification and uploading module is used to generate evidence information point by point from the interpolated forecast point set. The evidence information includes the referenced input forecast point identifier, release time identifier, forecast time identifier and lag time parameter. After organizing the interpolated forecast point set and evidence information into an interpolation interpretation record set, it is uploaded to the cloud to complete end-cloud collaboration. The interpolation layer appending and recording module is used to receive the interpolation interpretation record set on the cloud side and write it into the interpolation layer according to the release time identifier to generate a publishable forecast set. The interpolation layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. The decision-making non-rollback module is used to obtain the correction dataset, perform time alignment at a high time frequency granularity, and then match it point by point with the publishable forecast set to generate a correction record set, which is written into the correction layer to coexist with the interpolation layer. Subsequently, a visible value decision is performed on the interpolation layer records and the correction layer records to output a visible forecast point set. The visible values ​​corresponding to the historical forecast time identifiers that have been decided are not allowed to be rolled back under the subsequent release time identifiers.

[0012] A storage medium, comprising: The storage medium stores a computer program, which, when executed by a processor, implements the various steps of the high-frequency typhoon forecasting method.

[0013] A processor, comprising: A processor, and a memory connected to the processor; the memory is used to store computer programs; the processor is used to call and execute the computer programs in the memory to perform a high-frequency typhoon forecasting method.

[0014] The technical effects and advantages of this invention are as follows: 1. This scheme is based on the coexistence of interpolation layer and correction layer and implements non-rollback restriction on the adjudicated historical visible values. By decoupling the correction writing from the adjudication of external visible values, it can implement corrections with a minute-level lag while relatively suppressing repeated changes in historical forecast values, so that the full set of historical results released each time remains reproducible and verifiable. 2. Based on the screening of continuity and accuracy indicators of multi-source forecasts and the implementation of field unification and format conversion, by eliminating data sources that do not meet the requirements of continuity and accuracy, the impact of time-related missing data and caliber differences on subsequent high-frequency data supplementation can be relatively reduced, making the basic forecast sequence more stable and usable in terms of time and element caliber. 3. The edge computing side aligns the time according to the preset high time frequency granularity and generates interpolation forecast points according to the lag interpolation rules. By filling the uncovered time with adjacent valid forecast points, it can continuously produce high time frequency points and maintain the rolling update rhythm during the release lag period, which can relatively alleviate the output gap caused by the discontinuity of data release. 4. Generate evidence information containing the input forecast point identifier and lag time parameters for each interpolation forecast point and upload it to the cloud. By establishing traceable generation basis for each interpolation point, the verification capability and difference interpretation capability of the interpolation results can be improved, making the results in the end-to-cloud collaborative link easier to trace and verify. 5. Based on the cloud side, the interpolation interpretation record set is written to the interpolation layer according to the release time identifier, and only append writing is allowed and overwriting is not allowed. By solidifying the interpolation batch trace of each release, the traceability of historical versions can be improved and the interpolation records can be avoided from being overwritten by subsequent writes, making the rules for forming the publishable forecast set more stable and consistent. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention; Figure 2 This is a schematic diagram of the forecasting device of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1-2 The present invention provides a high-frequency typhoon forecasting method, comprising: S1. Obtain the multi-source forecast dataset and real-time dataset of the target typhoon. Perform field unification and format conversion on the multi-source forecast dataset and remove data sources that do not meet the requirements of continuity and accuracy according to the filtering rules. Output the basic forecast sequence and release time identifier set. This embodiment addresses the issues that arise in typhoon forecasting services when multiple forecast sources are accessed in parallel and their release times are asynchronous. These issues often include inconsistent field definitions, missing forecast time series data, and uncontrollable deviations from actual data. A reproducible multi-source forecast access and filtering process is presented to ensure that subsequent high-frequency interpolation on the edge computing side is based on time-continuous data consistent with actual conditions. Furthermore, the release time field is solidified into a set of release time identifiers for subsequent determination of the upper limit of usable forecast time, thereby reducing interpolation anomalies and result fluctuations caused by data latency and quality differences. This implementation process includes the following steps: The raw forecast content from each accessed forecast data source is first parsed to extract the typhoon identifier, forecast time identifier, and release time fields, forming a unified indexable set of source forecast sequences and providing input for subsequent filtering and boundary determination. The input consists of source files or data streams from each forecast data source, which must include at least two of the following: CCTV, JTWC, and ECMWF. Each source contains a typhoon identifier field, a set of forecast points, and a file generation time field or an operational release time field. The processing actions include parsing and reading each source, mapping the typhoon identifier to a typhoon identifier under the same encoding rule, and extracting the forecast time identifier point by point from the set of forecast points as a point time index. The generation or release time field in the file header is normalized to a release time field. Normalization includes converting time strings of different formats into timestamps in the same time zone and retaining them to the minute level. The output is a set of source forecast sequences carrying the release time field. Each source forecast sequence contains at least a typhoon identifier, a release time field, and a list of forecast points organized by forecast time identifier. The set of source forecast sequences is written to the source forecast buffer for the next step. Anomaly handling includes that when a source is missing the release time field, a replacement release time field is first obtained by parsing the file name. If it still cannot be obtained, a missing flag is output and the source is removed to avoid being unable to determine the upper limit of available information in the future. After the source forecast sequence set is formed, it is then quantified into a continuity index from a temporal integrity perspective and compared with a preset continuity constraint to eliminate data sources that cannot support high-frequency interpolation and ensure that the basic forecast sequence has adjacent valid forecast points that can be used for interpolation. The input is the source forecast sequence set. The processing actions include sorting each source forecast sequence in ascending order by forecast time identifier, calculating the time interval between adjacent forecast time identifiers to form an interval sequence, identifying continuous segments in the interval sequence that are greater than the preset upper limit of the time interval as missing segments, and calculating the length of the missing segments. The continuity index includes at least the maximum time interval and the maximum length of the missing segment. The preset continuity constraint is based on a preset configuration, specifically according to the nominal value of the source. The update frequency is set with an upper limit for the time interval. The nominal hourly time interval is set to an upper limit of two hours, and the nominal three-hourly time interval is set to an upper limit of six hours. At the same time, the upper limit for the length of the missing segment is set to twice the upper limit of the time interval to tolerate a single delay but exclude consecutive missing measurements. When the continuity index exceeds the preset continuity constraint, a rejection flag is output and the corresponding source forecast sequence is rejected. Otherwise, a retention flag is output and a continuous qualified forecast sequence is output to form a continuous qualified forecast sequence set and written to the continuity screening result area for the next step to read. Anomaly handling includes directly outputting a rejection flag and rejecting the source when the source forecast sequence contains only a single forecast point, which makes it impossible to form an interval sequence, to avoid the inability to select adjacent valid forecast points in subsequent interpolation. After obtaining the continuous qualified forecast sequence set, the continuous qualified forecast sequence set is further aligned with the actual data set by execution time, and the deviation is calculated to form an accuracy index. The data source is filtered by the accuracy boundary, and field unification and format conversion are completed, thereby outputting the basic forecast sequence and the set of release time identifiers. The input is the continuous qualified forecast sequence set and the actual data set corresponding to the target typhoon. The actual data set contains the actual time identifier and at least one of the actual elements including latitude and longitude, central pressure, and maximum wind speed. The processing actions include aligning each continuous qualified forecast sequence and the actual data set by execution time according to the forecast time identifier. The time alignment criterion is to select the nearest actual point with a time difference not exceeding a preset alignment window as the alignment actual point at the time corresponding to the forecast time identifier. The alignment window is set to 30 minutes. The system uses a preset configuration caliber. When no real-time point exists within the alignment window, a missing alignment marker is output and the forecast time identifier is skipped. After alignment is completed, the deviation is calculated for each alignment point. The deviation includes at least one of path deviation and intensity deviation. The path deviation is the spherical distance between the forecast latitude and longitude and the actual latitude and longitude, and the intensity deviation is the absolute difference between the forecast maximum wind speed and the actual maximum wind speed. The maximum value of the deviation for each alignment point is used as the accuracy index of the data source. The preset accuracy boundary is a statistical estimation caliber. Specifically, the deviation distribution of the data source under the same lead time is statistically analyzed in historical samples of similar typhoons, and the threshold corresponding to the 95th percentile is used as the boundary. The path deviation boundary is the distance threshold, and the intensity deviation boundary is the wind speed difference threshold. When the accuracy index is not greater than the corresponding boundary, field unification and format conversion are performed on the source forecast sequence to generate candidate basic forecast sequences. Field unification includes mapping latitude and longitude fields, intensity fields, and typhoon status fields from different sources to unified field names and units. Format conversion includes converting the point list into a unified serialized structure and creating an index according to the forecast time identifier. At the same time, the release time field of the source forecast sequence is written into the release time identifier set, which is a deduplicated set of release time fields from each retained source and arranged in ascending order of time. When the accuracy index is greater than the corresponding boundary, a removal flag is output and the source forecast sequence is removed. The output is the basic forecast sequence and the release time identifier set. The basic forecast sequence is written into the basic forecast storage area for S2 to read, and the release time identifier set is written into the release time index area for S2 to determine the upper limit of the available forecast time. Anomaly handling includes directly outputting a removal flag and removing the source when the alignment missing ratio of a certain source exceeds a preset ratio threshold. The ratio threshold is 50% and is a preset configuration caliber, used to exclude systematic unusable sources caused by long-term lack of real-time measurements. Through the above implementation process, a feasible data source filtering mechanism can be formed with continuity constraints and accuracy boundaries under the condition of parallel input of multiple forecasts with varying quality. This ensures that the output basic forecast sequence has stable and usable adjacent effective forecast points on the time axis and maintains verifiable consistency with the actual situation. At the same time, the published time identifier set provides an upper limit constraint on the available information for subsequent high time-frequency interpolation on the edge computing side, thereby reducing the risk of interpolation anomalies and result jumps caused by data delays, missing measurements, or large errors. In practical applications: when the business side simultaneously accesses multiple forecast sources and needs high-frequency output when the typhoon enters the warning line stage, the basic forecast sequence and published time identifier set can be generated first according to this embodiment. Then, the edge computing side restricts the available forecast time range based on the published time identifier set and performs lag interpolation to generate high time-frequency points. After the correction data arrives, the cloud side performs visible value adjudication on the interpolation layer and the correction layer and does not roll back the adjudicated history, thereby realizing a landing process that satisfies both high time-frequency continuous output and ensures that historical results are reproducible and verifiable.

[0018] S2. Read the basic forecast sequence and the release time identifier set on the edge computing side, perform time alignment on the basic forecast sequence according to the preset high time frequency time granularity, generate interpolated forecast points from adjacent valid forecast points for the uncovered time according to the lag interpolation rule, and output the interpolated forecast point set; This embodiment addresses the issue of inconsistency between time-crossing interpolation and time raster in high-frequency typhoon forecasts under minute-level lag and multi-source asynchronous release conditions. It presents a reproducible time raster generation and lag interpolation process on the edge computing side, ensuring that interpolation calculations are strictly constrained by the release time identifier and only fill in uncovered times within the available forecast time range. This guarantees clear and verifiable time boundaries for interpolated forecast points in subsequent edge-cloud collaborative links. The implementation process includes the following steps: The basic forecast sequence and release time identifier set read from the edge computing side are first used to explicitly quantify the usable information boundary into an upper limit of available forecast time and generate a unified target time raster to ensure that subsequent time alignment and interpolation are performed based on a unique time reference. The input is the basic forecast sequence and release time identifier set, where the basic forecast sequence includes typhoon identifiers, a list of forecast points indexed by forecast time identifiers, and forecast element fields for each point. The release time identifier set is a deduplicated set of release time identifiers corresponding to the data source, filtered and retained by S1, and sorted in ascending order of time. The processing actions include determining the release time identifier in the release time identifier set, using the latest release time identifier no later than the current processing time on the edge computing side as the release time identifier, and using it as the reference for the usable information boundary. Then, forecast points with forecast time identifiers no later than the release time identifier in the basic forecast sequence are selected as the set of usable forecast points, and the largest forecast time identifier in the set of usable forecast points is determined as the upper limit of available forecast time for this time. After obtaining the upper limit of available forecast time, target time grids are generated incrementally from the starting alignment time corresponding to the current release time identifier, according to the preset high-frequency time granularity, until the upper limit of available forecast time is reached. The starting alignment time is determined by rounding down the current release time identifier to the nearest integer point of the high-frequency time granularity to ensure that grids generated in different batches can be aligned. The preset high-frequency time granularity is a preset configuration, specifically configured as 15 minutes within the 48-hour warning line, and as one hour or three hours outside the 48-hour warning line, as given by the business configuration table. The output is the target time grid not exceeding the upper limit of available forecast time, and the target time grid is written to the time grid buffer on the edge computing side for the next step to read. Anomaly handling includes: when the set of available forecast points is empty, making it impossible to determine the upper limit of available forecast time, the current release time identifier is backed up to the previous release time identifier in the set of release time identifiers and the filtering is re-executed. If it is still empty after backing up, no available point marker is output and the current interpolation process is stopped to avoid generating unverifiable interpolation results. After the target time raster is generated, the base forecast sequence is mapped to the target time raster, and uncovered times are explicitly identified. Simultaneously, a unique pair of adjacent valid forecast points is constructed for each uncovered time, ensuring that interpolation calculations have a definite input source and reproducible neighbor selection rules. The inputs are the base forecast sequence and the target time raster. Processing actions include establishing a time index mapping table for the base forecast sequence based on forecast time identifiers, and traversing the target time raster time-by-time. When a raster time in the target time raster has a completely equal forecast time identifier in the time index mapping table, the corresponding forecast point is mapped as an aligned point and written to the alignment result set. When no completely equal forecast time identifier exists, the raster time is marked as an uncovered time and written to the uncovered time set. For each uncovered time, the forecast point with the smallest time difference from the base forecast sequence that is less than the uncovered time is selected as the previous valid forecast point. The forecast point with the smallest time difference and a time greater than the uncovered time is selected as the next valid forecast point. The selection rule uses the smallest difference in forecast time identifiers as the sole criterion to avoid multiple solutions. When there is no previous valid forecast point or no next valid forecast point for the uncovered time, the uncovered time is marked as boundary uncovered and removed from the subsequent interpolation set. The output is the set of uncovered times and the pairs of adjacent valid forecast points that correspond to them. The set of uncovered times and the pairs of adjacent valid forecast points are written to the interpolation task queue on the edge computing side for the next step to read. Anomaly handling includes selecting candidate points according to the data source priority rule when multiple forecast points with the same time difference exist for the same uncovered time, resulting in candidate points being listed in parallel. The value of the data source priority rule is a preset configuration and is consistent with the accuracy ranking of the data source retained in the S1 filtering results, so as to ensure that the neighbor selection can still be reproduced in different operating environments. After obtaining the set of uncovered time points and adjacent valid forecast point pairs, lag interpolation is performed on the uncovered time points under the constraint of the release time identifier, and the interpolated forecast points are output. This ensures that the interpolation results can both fill in the high-frequency raster and avoid introducing information later than the release time, thus preventing historical value jumps caused by time travel during the lag period. The inputs are the set of uncovered time points and adjacent valid forecast point pairs, as well as the release time identifier and the upper limit of available forecast time. The processing actions include reading the previous and next valid forecast points for each uncovered time point, and determining the lag lead parameter based on the release time identifier. The value of the lag lead parameter is based on a preset configuration, specifically configuring the maximum lag lead in the business configuration table according to the data source type and forecast lead interval. The maximum lag lead can be configured as six hours, twelve hours, or twenty-four hours, but not exceeding seventy-two hours. When performing interpolation, it is first determined whether the forecast time identifier of the next valid forecast point is later than the upper limit of available forecast time or later than the release time identifier plus the maximum lag lead. If any condition is met, the uncovered time is marked as non-interpolable and no interpolation forecast points are generated, ensuring that interpolation only uses adjacent valid forecast points available within the release time boundary; when the interpolable mark is met, interpolation coefficients are calculated according to the lag interpolation rule. The interpolation coefficients are the proportion of the time difference between the uncovered time and the previous valid forecast point to the time difference between the previous valid forecast point and the next valid forecast point. Linear interpolation is then performed on each forecast element, such as latitude and longitude, maximum wind speed, and central pressure, to generate interpolation forecast points. When a forecast element is a discrete state field, the state field of the previous valid forecast point is used as the interpolation output to avoid interpolation distortion of discrete fields; the output is the interpolated forecast point set, and the interpolated forecast point set is written into the interpolation result area on the edge computing side for S3 to generate evidence information for reading; the anomaly handling includes when the time difference between adjacent valid forecast points is zero, causing the interpolation coefficient to be uncalculated, the previous valid forecast point is directly copied as the interpolated forecast point and written into the copy mark to ensure that the output is calculable and to facilitate subsequent verification of the source of the anomaly; Through the above implementation process, the release time identifier can be converted into an upper limit of available forecast time and a unified high-frequency target time raster can be generated on the edge computing side. Then, the uncovered time is identified by a reproducible neighbor selection rule, and interpolation calculation is performed under the constraint of lag timeliness parameters. This ensures that the high-frequency interpolation results are strictly limited by the release time boundary and avoids the introduction of information later than the release time, reducing the risk of historical curve jumps and unreproducible results caused by time travel interpolation during the lag period. In practical applications: after receiving the release time identifier from the cloud side or selecting the latest available identifier from the release time identifier set, the edge computing side... First, select available forecast points in the basic forecast sequence that are no later than the release time and determine the upper limit of available forecast time. Then, generate target time grids with a 15-minute time granularity and identify uncovered times. Subsequently, construct the previous and next valid forecast points for each uncovered time and perform element-wise interpolation to generate an interpolated forecast point set under the constraint that the maximum lag time does not exceed 72 hours and write it into the interpolation result area. Subsequent steps directly read the interpolated forecast point set to generate evidence information and send it to the cloud side. In this way, even when there is a release lag and data asynchrony in real business, the interpolation process is still executable, verifiable and reproducible.

[0019] S3. Generate evidence information for each point in the interpolated forecast point set. The evidence information includes the referenced input forecast point identifier, release time identifier, forecast time identifier, and lag time parameter. Then, organize the interpolated forecast point set and evidence information into an interpolation interpretation record set and send it to the cloud to complete end-cloud collaboration. This embodiment addresses the problem that after high-frequency typhoon forecasts complete lag interpolation at the edge computing side, if only the interpolated forecast points are uploaded without verifiable generation evidence, the cloud side will find it difficult to determine whether the interpolated forecast points exceed the release time boundary, whether the correct adjacent valid forecast points are used, and whether the lag timeliness parameter constraints are met. It provides a processing flow that generates evidence information for the interpolated forecast points at the edge computing side and organizes it into an interpolation interpretation record set. This ensures that the objects transmitted on the edge-cloud collaborative link simultaneously include the interpolated forecast points and their generation evidence, thereby providing traceable, reproducible, and verifiable input for the cloud side's interpolation layer appending, correction layer coexistence, and subsequent visible value adjudication. The implementation process includes the following steps: On the edge computing side, the adjacent valid forecast points used are extracted point by point from the interpolation forecast point set and used to form input forecast point identifiers. This explicitly solidifies the source of the adjacent valid forecast points on which each interpolation forecast point depends into a reproducible index, thereby supporting the subsequent generation and verification of evidence information. The inputs are the interpolation forecast point set and the base forecast sequence. The interpolation forecast point set contains interpolation forecast points arranged by forecast time identifiers and their forecast element fields. The base forecast sequence contains a list of valid forecast points indexed by forecast time identifiers and the sequence position of each valid forecast point in the base forecast sequence. The processing actions include traversing the interpolation forecast point set point by point, reading the previous and next valid forecast points determined in S2 for each interpolation forecast point, locating the sequence positions of the previous and next valid forecast points in the base forecast sequence, and then... The sequence index identifier of a valid forecast point and the sequence index identifier of the next valid forecast point are concatenated in a fixed order to generate the input forecast point identifier corresponding to the interpolated forecast point. The fixed order is determined by writing the sequence index identifier of the previous valid forecast point first and then the sequence index identifier of the next valid forecast point to avoid ambiguity. The output is the correspondence between the interpolated forecast point and the input forecast point identifier, and the correspondence between the interpolated forecast point and the input forecast point identifier is written to the input forecast point identifier mapping area on the edge computing side for the next step to read. Anomaly handling includes marking the interpolated forecast point as having an index missing and removing it from the set that can be uploaded when a certain adjacent valid forecast point cannot be located in the basic forecast sequence, resulting in a missing sequence index identifier. At the same time, the reason for the index missing is written to the index missing log for operation and maintenance to investigate the index construction anomaly of the basic forecast sequence. The process involves writing the release time identifier, forecast time identifier, and lag time parameter to each interpolated forecast point and generating an evidence information set. This solidifies the time boundary and interpolation constraint of the interpolated forecast points into an indexable set of evidence fields, supporting the cloud side's verification and grouping of the interpolated forecast points' legitimacy. The inputs include the interpolated forecast point set, the correspondence between the interpolated forecast points and the input forecast point identifiers, and the release time identifier and lag time parameter. The release time identifier is the current release time identifier determined by S2, and the lag time parameter is the maximum lag time adopted by S2. The processing includes writing the release time identifier and the forecast time identifier of each interpolated forecast point to the set of interpolated forecast points, while simultaneously writing the lag time parameter and recording it as the interpolation constraint of that interpolated forecast point. The lag time parameter's value is a preset configuration and consistent with S2 to ensure consistent understanding of the interpolable boundary on both the edge and cloud sides. Subsequently, a typhoon identifier is used. Evidence information is generated by using the forecast time identifier as a record identifier. The generation method is to combine the forecast element fields of the interpolated forecast point with the corresponding input forecast point identifier, release time identifier, forecast time identifier, and lag time parameter into the same evidence information record, and supplement the evidence information record with an evidence information sequence number. The value of the evidence information sequence number is generated in ascending order of forecast time identifier within the same typhoon identifier and the same release time identifier to maintain the consistency of the upload order. The output is an evidence information set, which is written to the evidence information cache area on the edge computing side for the next step to read. Anomaly handling includes marking the interpolated forecast point as missing record identifier and removing it from the evidence information set when the interpolated forecast point is missing forecast time identifier or the release time identifier is empty, which causes the record identifier to be unable to be generated. At the same time, the reason for the missing record identifier is written to the record identifier missing log to ensure that the cloud side will not receive unindexable evidence information records in the future. After the interpolation forecast point set and the evidence information set are bound at the record level and organized into an interpolation interpretation record set, they are sent to the cloud side via an edge-cloud collaborative link. This ensures that the cloud-side receiving object contains both interpolation forecast points and evidence information at the record granularity and can be matched one-to-one with typhoon identifiers and forecast time identifiers. This supports traceable processing where cloud-side interpolation layer appending and subsequent correction layers coexist. The inputs are the interpolation forecast point set and the evidence information set. The processing actions include binding the interpolation forecast point set and the evidence information set at the record level according to the typhoon identifier and forecast time identifier. The binding criteria are: for each evidence information record, find interpolation forecast points with the same typhoon identifier and the same forecast time identifier in the interpolation forecast point set and merge them into an interpolation interpretation record. When multiple evidence information records correspond to the same interpolation forecast point, the one with the smallest evidence information sequence number is taken as the valid evidence information record. The remaining evidence information is marked as duplicate evidence. After binding is completed, the interpolation interpretation records are grouped by release time identifier and sorted by evidence information sequence number to form an interpolation interpretation record set. An upload packet is generated for the interpolation interpretation record set. The upload packet includes at least the typhoon identifier, release time identifier, number of interpolation interpretation records, and list of interpolation interpretation records. Then, the upload is executed through the end-cloud collaborative link. The upload packet is written to the end-cloud collaborative message queue and retransmission control is triggered. The retransmission control value is a preset configuration value and a maximum number of retransmissions is set. When the upload fails and the maximum number of retransmissions is not exceeded, the upload is re-uploaded at an exponential backoff interval. When the maximum number of retransmissions is exceeded, the upload failure result is output and the upload packet is written to the failure queue for manual retransmission. The output is the upload result, and the upload result is written to the upload result log on the edge computing side for reconciliation between the cloud side and the edge side. Through the above implementation process, evidence information regarding the generation source, time boundary, and lag constraint of interpolation forecast points can be formed on the edge computing side and bound to the interpolation forecast points at the record level. This ensures that the interpolation interpretation record set sent on the edge-cloud collaborative link simultaneously carries traceable input forecast point identifiers, release time identifiers, forecast time identifiers, and lag time parameters. This allows the cloud side to verify the legality of interpolation forecast points before the interpolation layer adds data and the correction layer coexists, and to locate the specific source and constraint criteria of adjacent valid forecast points in case of anomalies. This reduces the risk of inconsistencies in cloud-side processing and unreproducible results due to the untraceability of the edge interpolation process. In practical applications: the edge computing side completes the interpolation of the target typhoon... After the forecast point set is generated, the corresponding previous and next valid forecast points are extracted point by point, and their sequence index identifiers are located in the basic forecast sequence to form input forecast point identifiers. Then, for each interpolated forecast point, the current release time identifier, forecast time identifier, and maximum lag time are written, and an evidence information set is generated. The interpolated forecast points and evidence information are then bound according to the typhoon identifier and forecast time identifier to form an interpolation interpretation record set, which is sent to the cloud side through the end-to-cloud collaborative link. After receiving the data, the cloud side can write the data into the interpolation layer in groups according to the release time identifier and execute the visible value adjudication after the correction data arrives. The adjudicated history is not rolled back, thus realizing a feasible end-to-cloud collaborative high-frequency forecast process under the conditions of multi-source asynchronous and release lag.

[0020] S4. Receive the interpolation interpretation record set on the cloud side and write it into the interpolation layer according to the release time identifier to generate a publishable forecast set. The interpolation layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. This embodiment addresses the issue that if the interpolation interpretation record set sent by the cloud side to the edge computing side is directly overwritten, it will cause the interpolation results of the same typhoon identifier and the same forecast time identifier to be rewritten untraceably under different release time identifiers, thus causing historical curve jumps and unreproducible results. It proposes a processing flow that groups the interpolation layer by release time identifier on the cloud side and writes it to the interpolation layer in an append-only manner. This ensures that the interpolation layer records remain unrewriteable at the record granularity and can form a stable set of publishable forecasts by release time identifier, thereby providing consistent interpolation layer input for subsequent correction layer coexistence and visible value adjudication. The implementation process includes the following steps: After receiving the interpolation interpretation record set from the edge computing side via the end-cloud collaborative link, the cloud side first groups the interpolation interpretation record set according to the release time identifier and typhoon identifier to form a release group set. This ensures that subsequent writing and aggregation processes are performed within the same release time boundary and avoids cross-release time mixed writing. The input is the interpolation interpretation record set sent by the end-cloud collaborative link, which includes the typhoon identifier, the release time identifier, and a list of interpolation interpretation records formed by record-level binding. Each interpolation interpretation record contains at least the interpolation forecast point and its corresponding input forecast point identifier, forecast time identifier, and lag time parameter. The processing actions include the cloud side reading the uploaded packet from the end-cloud collaborative message queue and verifying the integrity of the typhoon identifier and release time identifier fields in the uploaded packet. The verification criteria are that the uploaded packet must contain both the typhoon identifier and the release time identifier. The number of interpolation interpretation records is consistent with the length of the interpolation interpretation record list; otherwise, an incomplete field flag is displayed in the output upload packet, and the upload packet is written to the exception queue. After successful verification, the cloud side groups the interpolation interpretation records according to the typhoon identifier and the release time identifier to generate a release group set, and sorts them in ascending order by the forecast time identifier within each release group to maintain a deterministic order. The output is the release group set, which is written to the cloud side's group buffer for the next step to read. Anomaly handling includes marking duplicate interpolation interpretation records with the same typhoon identifier and the same release time identifier as duplicate records in the release group, and retaining only the interpolation interpretation record with the smallest evidence information sequence number as a valid record. The evidence information sequence number comes from the evidence information field in the interpolation interpretation record to ensure the uniqueness of the subsequent sequence. For each release group in the release group set, an interpolation layer write sequence is generated, and an append write determination is performed on the interpolation layer write sequence. This writes the interpolation interpretation records into the interpolation layer in a way that prevents them from being overwritten or modified, forming a traceable set of interpolation layer records. The inputs are the release group set and the cloud-side interpolation layer storage area. The interpolation layer storage area is used to store the write results of the interpolation interpretation records and supports searching for whether a corresponding interpolation layer record already exists by the write identifier. The processing actions include generating an interpolation layer write sequence for each release group. The interpolation layer write sequence is generated by generating a write identifier for each interpolation interpretation record in the release group and binding it to the interpolation interpretation record to form a write item. The write identifier is generated by concatenating the typhoon identifier, the forecast time identifier, and the release time identifier to ensure that the same typhoon identifier and the same forecast time identifier can be appended to form multiple interpolation layer records under different release time identifiers, and these records cannot be overwritten or modified. After generating the interpolation layer write sequence, an append write determination is performed on each item in the interpolation layer write sequence. The determination criteria are to search the interpolation layer storage area for an interpolation layer record that is completely consistent with the write identifier. If it does not exist, the corresponding interpolation interpretation record is appended to the interpolation layer storage area as a new interpolation layer record and a write success flag is output. If it exists, the existing interpolation layer record is kept unchanged and a rejection overwrite flag is output. The output is a set of write results with write success flags or rejection overwrite flags, and the set of write results is written to the interpolation layer write log on the cloud side for subsequent verification of whether the interpolation layer write has been duplicated or duplicated. Anomaly handling includes marking the corresponding interpolation interpretation record as missing write identifier and writing it to the failure queue when the write identifier field is missing, so as to trace back whether there is a missing field in the generation of evidence information on the edge computing side. After the interpolation layer is appended, the interpolation layer is aggregated according to the release time identifier to generate a publishable forecast set, which serves as the input for subsequent point-by-point matching and visibility value adjudication in the correction layer. This ensures that the interpolation results used in the correction process and visibility value adjudication are consistent with the release time boundary. The input consists of interpolation layer records and the set of write results in the interpolation layer storage area. The processing actions include reading the set of interpolation layer records with the same release time identifier from the interpolation layer storage area according to the target release time identifier, and aggregating the interpolation layer record set in ascending order of forecast time identifier to generate a publishable forecast set. The publishable forecast set is composed of a list of interpolation forecast points indexed by forecast time identifier, and retains the input forecast point identifier and lag time parameter corresponding to each interpolation forecast point for subsequent correction matching and adjudication. When the write result set is... When a rejection-overwrite flag exists, the cloud side will include the rejection-overwrite flag in the duplicate write count and write the duplicate write count into the publishable forecast set's publishable metadata field for subsequent judgment of whether there is duplicate delivery in the uplink; the output is the publishable forecast set and the publishable forecast set is written into the cloud side's publishable forecast cache for S5 point-by-point matching and reading; the exception handling includes when the interpolation layer record set is empty, resulting in the inability to form a publishable forecast set, the cloud side will backtrack the publishable group identified by the corresponding publishable time in the group cache and check whether there is a missing uplink packet. If there is a missing uplink packet, an uplink packet missing flag will be output and a cloud-end collaborative re-pull request will be triggered. If there is no missing uplink packet, an interpolation layer record missing flag will be output and the generation of this publishable forecast set will be stopped to avoid outputting an empty set to the subsequent correction link and causing erroneous decisions; Through the above implementation process, the interpolation interpretation record set can be grouped by release time identifier and typhoon identifier on the cloud side, and append write judgment can be performed with the write identifier as a constraint. This enables the interpolation layer records to be appended across release time identifiers under the same typhoon identifier and the same forecast time identifier dimension, and prohibits overwriting. This ensures that the publishable forecast set is stably generated within the release time boundary and can be used for subsequent correction layer coexistence and visible value adjudication. It reduces the risk of curve jumps and unreproducible results caused by the overwriting of historical interpolation results. In practical applications: the cloud side reads the interpolation solution sent from the edge computing side from the edge-cloud collaborative message queue. The system interprets the record set and generates a release group set according to the typhoon identifier and release time identifier. Then, it generates a write identifier for each interpolated interpretation record, which is formed by concatenating the typhoon identifier, forecast time identifier, and release time identifier. The interpolation layer write sequence is constructed in ascending order of forecast time identifier. Subsequently, the interpolation layer storage area is searched according to the write identifier and an append write judgment is performed to reject any overwrite write. Finally, the system is aggregated from the interpolation layer according to the release time identifier to generate a publishable forecast set and writes it to the publishable forecast cache for subsequent correction point set point matching and visible value adjudication. This ensures that the interpolation results remain traceable, reproducible, and verifiable in multi-source asynchronous release and high-frequency update scenarios.

[0021] S5. After obtaining the correction dataset and performing time alignment at the high time frequency granularity, it is matched point by point with the publishable forecast set to generate a correction record set and written into the correction layer to coexist with the interpolation layer. Then, the visible value adjudication is performed on the interpolation layer record and the correction layer record to output the visible forecast point set. The visible value corresponding to the adjudicated historical forecast time identifier is not allowed to be rolled back under the subsequent release time identifier. This embodiment addresses the problem that in high-frequency typhoon forecasts, where correction data arrives continuously and is released frequently, using overlay correction can lead to interpolation results and correction results overwriting each other at different release times, causing repeated changes in historical visible results. It proposes a process where the correction record set is written into the correction layer and coexisted with the interpolation layer on the cloud side, and then visible value adjudication is performed within the release time boundary, with a non-rollback restriction applied to adjudicated historical data. Simultaneously, a computable adjudication link is provided for outputting the visible forecast point set, ensuring that visible value selection satisfies both the operational requirement of prioritizing correction and the constraints of the upper limit of available information at the release time and the continuity consistency constraint. This guarantees that high-frequency output remains traceable, reproducible, and verifiable under real operational conditions. The implementation process includes the following steps: After acquiring the corrected dataset of the target typhoon and generating a corrected record set on the cloud side, the corrected record set is written to the correction layer storage area to coexist with the interpolation layer storage area. This ensures that the corrected data can be appended at the record granularity without overwriting or rewriting the interpolation layer records, thus providing coexisting input for subsequent visible value adjudication. The inputs are the corrected dataset of the target typhoon, the publishable forecast set, and the release time identifier. The corrected dataset contains the correction time identifier and correction feature fields of the corrected points. The publishable forecast set contains the interpolated forecast points indexed by the forecast time identifier and their forecast feature fields. The release time identifier is the release time identifier corresponding to the current processing batch on the cloud side. The processing actions include performing time alignment on the corrected dataset at a high time frequency granularity to generate a set of corrected points. The time alignment method involves rounding down the correction time identifier to the nearest integer point at the high-frequency time granularity to form an aligned correction time identifier. Correction points falling under the same aligned correction time identifier are then selected as the last one in the latest arrival order to avoid duplicate corrections at the same time. The high-frequency time granularity is a preset configuration and consistent with S2. Subsequently, the correction point set and the publishable forecast set are matched point-by-point according to the typhoon identifier and forecast time identifier. The matching method involves searching the publishable forecast set for each correction point for interpolated forecast points with the same typhoon identifier and the same forecast time identifier. When a match is found, a correction record containing the correction element field is generated and marked as a non-empty correction record; otherwise, a record containing only the typhoon identifier is generated. Empty correction records with typhoon identifier, forecast time identifier, and release time identifier are marked as empty correction records to maintain consistency between the time coverage of the correction layer records and the releaseable forecast set. After generating the correction record set, the correction record set is written to the correction layer storage area according to the release time identifier. The write entry is to generate a correction write identifier for each correction record, which is formed by concatenating the typhoon identifier, forecast time identifier, and release time identifier. The correction layer storage area is then searched for the existence of the corresponding correction layer record according to the correction write identifier. If it does not exist, it is appended to the correction layer and a write success flag is output. If it exists, the existing correction layer record is kept unchanged and a rejection flag is output. Multiple correction layers can be appended under the same typhoon identifier and the same forecast time identifier under different release time identifiers. Records are kept and overwriting between each correction layer record is prohibited; the output is the correction layer write result and the correction layer write result is written to the correction layer write log for subsequent review to check whether there is duplicate correction submission or duplicate writing; exception handling includes marking the correction dataset as missing time field and stopping the generation of this correction point set when the correction dataset is missing correction time identifier, thus preventing the time alignment from being performed, in order to avoid generating unindexable correction records. At the same time, when the matching missing ratio between the correction point set and the publishable forecast set exceeds the ratio threshold, the release time identifier is marked as insufficient correction coverage and the insufficient correction coverage mark is written to the correction layer write log. The ratio threshold is set to 50% using a preset configuration to indicate that the correction source may be delayed or missing. Alignment is performed on interpolation layer records and correction layer records according to forecast time identifiers to form a pair of records to be decided. Then, visible value decision and historical rollback restriction determination are performed to ensure that the set of visible forecast points is generated according to uniform rules within the same release time identifier range and that the decided historical data remains stable under subsequent release time identifiers. The inputs are the interpolation layer records corresponding to the release time identifiers in the interpolation layer storage area, the correction layer records corresponding to the release time identifiers in the correction layer storage area, and the decided visible values ​​corresponding to the historical release time identifiers. The decided visible values ​​are those from the cloud side that were released earlier than the initial release time identifier. The visible values ​​in the set of visible forecast points that have been output and fixed under the historical release time identifier; the processing actions include reading the interpolation layer record set with the same release time identifier from the interpolation layer storage area and reading the correction layer record set with the same release time identifier from the correction layer storage area, and then performing alignment according to the forecast time identifier to form a pair of records to be decided. The formation caliber is to use the forecast time identifier as an index to make a key connection between the interpolation layer record set and the correction layer record set, and to fill in the missing records on the side with empty records to maintain the full coverage of the pair of records to be decided on the time axis; For each pair of records to be decided, a visibility value decision is made. The decision criteria are as follows: if the correction layer record is not empty, the forecast value corresponding to the correction layer record is output as the visibility value and written into the visible forecast point set; if the correction layer record is empty, the forecast value corresponding to the interpolation layer record is output as the visibility value and written into the visible forecast point set. After obtaining the visible forecast point set, a historical rollback restriction decision is made. The decision criteria are as follows: among the decided visible values ​​corresponding to the historical release time identifier, a decided visible value with the same typhoon identifier and the same forecast time identifier as the current visible forecast point set is searched. When found, the decided visible value is kept unchanged, and the current visible value is marked as ineffective. Only for forecast time identifiers where no decided visible value is found, the current visible value is retained to form a non-rollback visible forecast point set. The scope of historical release time markers includes all release time markers earlier than the release time marker, and they are retrieved sequentially in ascending time order to ensure that the earliest solidified result takes priority; the output is a set of visible forecast points that are not rolled back and is written to the visible value storage area as the input of the adjudicated visible value for subsequent release time markers; the exception handling includes marking the release time marker as having no interpolation input and stopping the output of the visible value for this time when the interpolation layer record set is empty, resulting in the inability to form a pair of records to be adjudicated, in order to avoid inconsistencies in the timeline caused by relying solely on the output of the correction layer; at the same time, when the adjudicated visible value storage area corresponding to the historical release time marker is inaccessible, the historical rollback restriction judgment is switched to read-only failure mode and a rollback restriction unavailable flag is output to indicate that the subsequent release results may have historical changes; The process of outputting the visible forecast point set in S5 also includes constructing an adjudication problem with an objective function and a set of constraints, calculating the confidence level under this adjudication problem, and performing confidence updates to quantify the credibility of the interpolation layer records and the correction layer records into comparable confidence levels and to ensure that the selection of visible values ​​converges to a verifiable output under the constraint of the upper limit of available information. The inputs are the interpolation layer records and correction layer records corresponding to the same typhoon identifier and the same forecast time identifier, as well as the evidence information and release time identifier carried in the interpolation interpretation records. The evidence information includes the input forecast point identifier, release time identifier, forecast time identifier, and lag time parameters. The processing actions include extracting forecast values ​​from the interpolation layer records and correction layer records, and extracting the input forecast point identifier and lag time parameters from the evidence information associated with the interpolation layer records. The lag time parameter is used to extract the release time identifier and construct the adjudication problem accordingly. The objective function of the adjudication problem consists of the forecast value consistency cost and the evidence integrity cost. The forecast value consistency cost is calculated as the sum of the absolute differences between the forecast values ​​recorded in the interpolation layer and the forecast values ​​recorded in the correction layer in each forecast element dimension. The evidence integrity cost is calculated as the sum of the number of missing evidence information fields and the number of times the input forecast point identifier can be located. The constraint set includes at least the upper limit constraint of available information corresponding to the release time identifier. The upper limit constraint of available information is calculated as follows: the forecast time identifier of the valid forecast point corresponding to the input forecast point identifier referenced in the evidence information of the interpolation layer shall not be later than the release time identifier and shall not exceed the release time identifier plus the lag time parameter. Under the adjudication problem, the confidence scores of the interpolation layer records and the correction layer records are calculated separately, and confidence updates are performed. The initial value of the interpolation layer record confidence score is a preset configuration and can be set to one. The confidence update caliber for the interpolation layer record confidence score is based on attenuation correction using the lag timeliness parameter in the evidence information and penalty correction based on the locatability of the input forecast point identifier. The initial value of the correction layer record confidence score is a preset configuration and can be set to one. The confidence update caliber for the correction layer record confidence score is based on convergence correction using the time alignment residual and the actual consistency residual of the correction layer record. The calculation caliber for the time alignment residual is the absolute difference between the correction time identifier and the forecast time identifier. The calculation method for the consistency residual of the actual data is the deviation between the predicted value of the correction layer record and the actual dataset after alignment with the same prediction time identifier. The calculation method of the deviation is consistent with the S1 accuracy index. The output is the interpolation confidence and the correction confidence. The interpolation confidence and the correction confidence are written into the adjudication appendix field corresponding to the record to be adjudicated for the next step of consistency verification and constraint solving. Anomaly handling includes that when the correction layer record is missing the correction time identifier, which makes the time alignment residual uncalcible, the confidence of the correction layer record is directly decayed to the preset lower limit and a missing time field is written. The preset lower limit is set to 0.1 using a preset configuration to ensure that the record can still participate in the adjudication but its priority is reduced. The system performs multi-view consistency checks on interpolation and correction confidence levels, and performs constraint solving to select visible values ​​when conflicts exist. This ensures that the selection of visible values ​​simultaneously satisfies continuity constraints and minimizes the objective function. A verifiable decision result is output when the stopping condition is met. The inputs are the pair of records to be decided, the interpolation confidence level, the correction confidence level, the decided visible values ​​corresponding to adjacent forecast time markers, and the decision objective function and constraint set. The processing actions include performing multi-view consistency checks on the interpolation and correction confidence levels and outputting conflict flags. The multi-view consistency check includes at least a continuity constraint check. The calculation method for continuity constraint verification is to calculate the continuity violation amount by comparing the forecast value recorded in the interpolation layer and the forecast value recorded in the correction layer with the adjudicated visible value corresponding to the adjacent forecast time marker. The calculation method for the continuity violation amount is the sum of the spherical distance increment between the current forecast value and the adjacent adjudicated visible value on the path element and the absolute difference increment on the intensity element. The continuity violation amount is then compared with the preset continuity threshold to obtain the continuity constraint satisfaction result. The preset continuity threshold is set using a preset configuration method and can be configured according to the typhoon intensity level to adapt to the allowable change range of different intensity stages. When both the interpolation layer records and the correction layer records satisfy the continuity constraint and the cost of consistency between their forecast values ​​does not exceed the preset consistency threshold, a conflict flag is output as "No"; otherwise, a conflict flag is output as "Yes". The preset consistency threshold is set using a preset configuration as the maximum allowable difference across all forecast element dimensions. For record pairs with a "No" conflict flag, visible values ​​are output and written into the visible forecast point set according to the non-empty priority rule of correction layer records. For record pairs with a "Yes" conflict flag, constraint solving is performed to select visible values. The constraint solving caliber is to minimize the objective function among the candidate visible value set that satisfies the constraint set, and to use the record corresponding to the larger of the interpolation confidence and the correction confidence as the initial candidate. During the iteration process, if the decrease in the objective function is less than the stopping threshold or the number of iterations reaches the maximum number of iterations, a stopping condition is triggered and the current optimal visible value is output. The stopping threshold and the maximum number of iterations are both set using a preset configuration. The stopping threshold can be set to 0.01 and the maximum number of iterations can be set to 10 to ensure that the computational latency of the cloud side is controllable in high-frequency release scenarios. The output is a set of visible forecast points, which is written to the visible value output buffer for historical rollback constraint determination. Anomaly handling includes directly reverting to the correction layer record non-empty priority rule output visible value and writing a constraint missing flag to indicate an abnormal adjudication link configuration. Through the above implementation process, the correction dataset can be aligned with high-frequency time granularity on the cloud side and matched point-by-point with the publishable forecast set to generate a correction record set, which is then appended to the correction layer storage area. This ensures that correction layer records and interpolation layer records coexist within the release time boundary and are prohibited from being overwritten. Subsequently, based on the alignment of interpolation layer records and correction layer records to form a pair of records to be decided, a set of visible forecast points is output, and a non-rollback restriction is applied to the decided history. This maintains the stability and traceability of visible results under conditions of continuous correction arrival and high-frequency release. At the same time, by constructing the decision objective function and constraint set, performing confidence updates, multi-view consistency verification, and constraint solving, the visible value decision converges to a verifiable output under the upper limit of available information and continuity constraints, reducing the risk of erroneous decisions in conflict scenarios. In practical applications: after receiving the correction dataset of the target typhoon, the cloud side first aligns the correction time identifier with a 15-minute time granularity to generate a correction point set, and then matches it point-by-point with the publishable forecast set according to the typhoon identifier and forecast time identifier. A correction record set is generated, and then appended to the correction layer storage area and coexisted with the interpolation layer storage area according to the correction write identifier formed by concatenating the typhoon identifier, forecast time identifier, and release time identifier. Subsequently, the interpolation layer record and correction layer record corresponding to the same release time identifier are read to form a pair of records to be decided. Under the correction non-empty priority rule, the initial visible forecast point set is output and written to the visible value output buffer. If there is a consistency conflict between the interpolation layer record and the correction layer record, an objective function containing the forecast value consistency cost and the evidence integrity cost is further constructed, and the upper limit constraint of available information is applied to calculate the interpolation confidence and correction confidence. The continuous default quantity verification and constraint solution are performed to select the visible value. Finally, the output visible forecast point set is compared with the decided visible values ​​under the historical release time identifier, and the decided history is kept unchanged to form a non-rollback visible forecast point set and written to the visible value storage area. Thus, under the correction delay and high-frequency release pressure of real business, a high-frequency typhoon forecasting process with end-cloud collaboration can still be implemented.

[0022] Furthermore, the present invention also includes a high-frequency typhoon forecasting device, the device comprising a multi-source screening module, a high-frequency supplementation module, an evidence solidification and uploading module, an interpolation layer addition and retention module, and a decision-making version non-return module: The multi-source filtering module is used to acquire the multi-source forecast dataset and the real-time dataset of the target typhoon. It performs field unification and format conversion on the multi-source forecast dataset and removes data sources that do not meet the requirements of continuity and accuracy according to the filtering rules. It outputs the basic forecast sequence and the set of release time identifiers. The high-frequency interpolation module is used to read the basic forecast sequence and the release time identifier set on the edge computing side, perform time alignment on the basic forecast sequence according to the preset high time frequency granularity, generate interpolated forecast points from adjacent valid forecast points for the uncovered time according to the lag interpolation rule, and output the interpolated forecast point set; The evidence solidification and uploading module is used to generate evidence information point by point from the interpolated forecast point set. The evidence information includes the referenced input forecast point identifier, release time identifier, forecast time identifier and lag time parameter. After organizing the interpolated forecast point set and evidence information into an interpolation interpretation record set, it is uploaded to the cloud to complete end-cloud collaboration. The interpolation layer appending and recording module is used to receive the interpolation interpretation record set on the cloud side and write it into the interpolation layer according to the release time identifier to generate a publishable forecast set. The interpolation layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. The decision-making non-rollback module is used to obtain the correction dataset, perform time alignment at a high time frequency granularity, and then match it point by point with the publishable forecast set to generate a correction record set, which is written into the correction layer to coexist with the interpolation layer. Subsequently, a visible value decision is performed on the interpolation layer records and the correction layer records to output a visible forecast point set. The visible values ​​corresponding to the historical forecast time identifiers that have been decided are not allowed to be rolled back under the subsequent release time identifiers.

[0023] A storage medium, comprising: The storage medium stores a computer program, which, when executed by a processor, implements the various steps of the high-frequency typhoon forecasting method.

[0024] A processor, comprising: A processor, and a memory connected to the processor; the memory is used to store computer programs; the processor is used to call and execute the computer programs in the memory to perform a high-frequency typhoon forecasting method.

[0025] Working principle: First, data from multiple forecast sources and real-time data are unified into the same set of fields and time calibers. Unreliable data sources are filtered out using continuity and deviation from real-time data to obtain the basic forecast sequence and the time set of each data release. Then, a unified time raster is generated on the edge computing side according to the preset high time frequency granularity. Only within the information range allowed by the latest release time, the gap time is filled by lag interpolation using adjacent valid forecast points to obtain interpolated forecast points. At the same time, traceable evidence information is generated for each interpolation point and sent to the cloud side. The cloud side appends the interpolation results to the interpolation layer according to the release time to form a publishable forecast set. When the correction data arrives, it is also appended to the correction layer. After coexisting with the interpolation layer, the publicly visible value is determined, and a no-rollback restriction is implemented for historical times that have already been released to ensure that subsequent updates do not repeatedly rewrite history. For example, when a typhoon approaches the coast, the operational requirement is to update every fifteen minutes. However, the forecasts from different agencies are not synchronized, and some data points are missing, while real-time data is continuously supplemented. After receiving the basic forecast sequence filtered by the cloud side, the edge side generates a fifteen-minute time grid with the latest release time as the boundary, interpolates the missing times using two adjacent points, and sends up the adjacent points and lag parameters used for each interpolation point. After receiving the data, the cloud side generates a publishable result according to the release time and publishes it. If the correction center subsequently sends a more accurate correction point, the cloud side will write the correction point into the correction layer and work with the interpolation layer to determine the final displayed value. However, the historical times that have already been released in the previous round remain unchanged, and only the new times that have not yet been determined are updated. This ensures continuous high-frequency updates without causing the historical trajectory to jump back and forth.

[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-frequency typhoon forecasting method, characterized in that, include: S1. Obtain the multi-source forecast dataset and real-time dataset of the target typhoon. Perform field unification and format conversion on the multi-source forecast dataset and remove data sources that do not meet the requirements of continuity and accuracy according to the filtering rules. Output the basic forecast sequence and release time identifier set. S2. Read the basic forecast sequence and the release time identifier set on the edge computing side, perform time alignment on the basic forecast sequence according to the preset high time frequency time granularity, generate interpolated forecast points from adjacent valid forecast points for the uncovered time according to the lag interpolation rule, and output the interpolated forecast point set; S3. Generate evidence information for each point in the interpolated forecast point set. The evidence information includes the referenced input forecast point identifier, release time identifier, forecast time identifier, and lag time parameter. Then, organize the interpolated forecast point set and evidence information into an interpolation interpretation record set and send it to the cloud side. S4. Receive the interpolation interpretation record set on the cloud side and write it into the interpolation layer according to the release time identifier to generate a publishable forecast set. The interpolation layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. S5. After obtaining the correction dataset and performing time alignment at the high time frequency granularity, it is matched point by point with the publishable forecast set to generate a correction record set and written into the correction layer to coexist with the interpolation layer. Then, the visible value adjudication is performed on the interpolation layer record and the correction layer record to output the visible forecast point set. The visible value corresponding to the adjudicated historical forecast time identifier is not allowed to be rolled back under the subsequent release time identifier.

2. The high-frequency typhoon forecasting method according to claim 1, characterized in that: S1 includes: S1-1. Obtain the source forecast sequences from each forecast data source, and extract the typhoon identifier, forecast time identifier, and release time field from the source forecast sequences. Output the source forecast sequence set carrying the release time field. S1-2. Sort each source forecast sequence in the source forecast sequence set according to the forecast time identifier and calculate the continuity index. The continuity index includes the time interval between adjacent forecast time identifiers and the length of the missing segment. When the continuity index exceeds the preset continuity constraint, output the removal flag and remove the corresponding source forecast sequence. Otherwise, output the continuous qualified forecast sequence and output the continuous qualified forecast sequence set. S1-3. For each source forecast sequence in the continuous qualified forecast sequence set, perform time alignment with the actual dataset according to the forecast time identifier and calculate the accuracy index. The accuracy index is whether the deviation after alignment is not greater than the preset accuracy boundary. When the accuracy index meets the preset accuracy boundary, perform field unification and format conversion on the source forecast sequence to generate candidate basic forecast sequences and write the release time field of the corresponding source forecast sequence into the release time identifier set. Otherwise, remove the corresponding source forecast sequence and output the basic forecast sequence and the release time identifier set.

3. The high-frequency typhoon forecasting method according to claim 2, characterized in that: S2 includes: S2-1. Read the basic forecast sequence and release time identifier set on the edge computing side, determine the upper limit of the available forecast time using the release time identifier set, and generate a target time raster that does not exceed the upper limit of the available forecast time according to the preset high time frequency time granularity, and output the target time raster. S2-2. Perform time alignment on the basic forecast sequence according to the target time grid. Mark the time without a corresponding forecast point in the target time grid as an uncovered time. For each uncovered time, select the previous and next valid forecast points in the basic forecast sequence that are time-adjacent as adjacent valid forecast points. Output the set of uncovered times and the pair of adjacent valid forecast points. S2-3. For each uncovered time point, perform interpolation calculation according to the lag interpolation rule. The lag interpolation rule includes determining the lag lead time parameter based on the release time identifier and restricting the interpolation calculation to use only adjacent valid forecast points that are no later than the upper limit of the available forecast time corresponding to the release time identifier, and outputting the interpolated forecast point set.

4. The high-frequency typhoon forecasting method according to claim 3, characterized in that: S3 includes: S3-1. On the edge computing side, extract the adjacent valid forecast points from the interpolation forecast point set point by point to generate the adjacent valid forecast points used, and use the sequence index identifier of the adjacent valid forecast points in the basic forecast sequence as the input forecast point identifier, and output the correspondence between the interpolation forecast points and the input forecast point identifiers. S3-2. Write the release time identifier, forecast time identifier, and lag time parameter for each interpolated forecast point, and use the typhoon identifier and forecast time identifier as record identifiers. Combine the interpolated forecast points with the corresponding input forecast point identifier, release time identifier, forecast time identifier, and lag time parameter to generate evidence information and output the evidence information set. S3-3. Bind the interpolation forecast point set and the evidence information set at the record level according to the typhoon identifier and the forecast time identifier, organize them into an interpolation interpretation record set, and send the interpolation interpretation record set to the cloud side through the end-to-cloud collaborative link, and output the sending result.

5. The high-frequency typhoon forecasting method according to claim 4, characterized in that: S4 includes: S4-1. Receive the interpolation interpretation record set sent from the edge computing side via the end-cloud collaborative link on the cloud side, group the interpolation interpretation record set according to the release time identifier and the typhoon identifier, and output the release group set; S4-2. Generate an interpolation layer write sequence for each release group in the release group set. The interpolation layer write sequence is arranged in ascending order according to the forecast time identifier and a write identifier is generated for each interpolation interpretation record. Output the interpolation layer write sequence. S4-3. Perform an append write judgment on the interpolation layer write sequence. When there is no interpolation layer record corresponding to the write identifier in the cloud-side interpolation layer, append the corresponding interpolation interpretation record to the interpolation layer and output a write success flag. When there is already an interpolation layer record corresponding to the write identifier in the cloud-side interpolation layer, keep the existing interpolation layer record unchanged and output a rejection overwrite flag. S4-4. After the interpolation layers have been appended, they are aggregated according to the release time identifier to generate a publishable forecast set. The publishable forecast set consists of interpolation layer records corresponding to the same release time identifier. The publishable forecast set is then output.

6. The high-frequency typhoon forecasting method according to claim 5, characterized in that: S5 includes: S5-1. Obtain the correction dataset of the target typhoon on the cloud side, and perform time alignment on the correction dataset at a high time frequency granularity to generate a correction point set. S5-2. Perform point-by-point matching between the correction point set and the publishable forecast set according to the typhoon identifier and the forecast time identifier. Generate correction records for points with matching relationships and generate empty correction records for points without matching relationships. Output the correction record set. S5-3. Write the correction record set into the correction layer storage area according to the release time identifier so that it can coexist with the interpolation layer storage area. The correction layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. Output the correction layer writing result. S5-4. Align the interpolation layer records corresponding to the release time identifier in the interpolation layer with the correction layer records corresponding to the release time identifier in the correction layer according to the forecast time identifier, and output the record pair to be decided. S5-5. Perform visible value adjudication for each pair of records to be adjudicated. When the correction layer record is a non-empty correction record, output the forecast value corresponding to the correction layer record as the visible value. When the correction layer record is an empty correction record, output the forecast value corresponding to the interpolation layer record as the visible value. Output the set of visible forecast points. S5-6. Perform historical rollback restriction judgment on the visible forecast point set. When there is a historical release time identifier corresponding to the adjudicated visible value that is earlier than the release time identifier on the cloud side, keep the adjudicated visible value unchanged and only output the visible value for the forecast time identifier that has not yet been adjudicated. Output the visible forecast point set without rollback.

7. The high-frequency typhoon forecasting method according to claim 6, characterized in that: The process of outputting the set of visible forecast points in S5 also includes: S5-51. Obtain the interpolation layer record and correction layer record corresponding to the same typhoon identifier and the same forecast time identifier, and extract the forecast value, evidence information and release time identifier from them. Construct a decision problem with the forecast value consistency cost and evidence integrity cost as the objective function and the upper limit of available information corresponding to the release time identifier as the constraint set. Output the decision objective function and constraint set. S5-52. Under the adjudication problem, calculate the confidence scores for the interpolation layer records and the correction layer records respectively, and perform confidence updates. The confidence update uses the input forecast point identifier and lag time parameter referenced in the evidence information to perform attenuation correction on the confidence score of the interpolation layer records, and uses the time alignment residual and the reality consistency residual of the correction layer records to perform convergence correction on the confidence score of the correction layer records. Output the interpolation confidence score and the correction confidence score. S5-53. Perform multi-view consistency verification on the interpolation confidence and correction confidence and output conflict flags. The multi-view consistency verification includes at least performing continuity constraint verification on the interpolation layer records and correction layer records with the adjudicated visible values ​​corresponding to adjacent forecast time identifiers and calculating the continuity default amount. S5-54. For records to be decided with a conflict mark of no, output visible values ​​according to the non-empty priority rule of the correction layer records. For records to be decided with a conflict mark of yes, perform constraint solving to select visible values ​​by minimizing the objective function and satisfying the constraint set as the stopping condition, and output the set of visible forecast points.

8. A high-frequency typhoon forecasting device, used to implement the high-frequency typhoon forecasting method according to any one of claims 1-7, the device comprising a multi-source screening module, a high-frequency supplementation module, an evidence solidification and uploading module, an interpolation layer addition and retention module, and a decision-making version non-return module, characterized in that: The multi-source filtering module is used to acquire the multi-source forecast dataset and the real-time dataset of the target typhoon. It performs field unification and format conversion on the multi-source forecast dataset and removes data sources that do not meet the requirements of continuity and accuracy according to the filtering rules. It outputs the basic forecast sequence and the set of release time identifiers. The high-frequency interpolation module is used to read the basic forecast sequence and the release time identifier set on the edge computing side, perform time alignment on the basic forecast sequence according to the preset high time frequency granularity, generate interpolated forecast points from adjacent valid forecast points for the uncovered time according to the lag interpolation rule, and output the interpolated forecast point set; The evidence solidification and uploading module is used to generate evidence information point by point from the interpolated forecast point set. The evidence information includes the referenced input forecast point identifier, release time identifier, forecast time identifier and lag time parameter. After organizing the interpolated forecast point set and evidence information into an interpolation interpretation record set, it is uploaded to the cloud to complete end-cloud collaboration. The interpolation layer appending and recording module is used to receive the interpolation interpretation record set on the cloud side and write it into the interpolation layer according to the release time identifier to generate a publishable forecast set. The interpolation layer records corresponding to the same typhoon identifier and the same forecast time identifier are only allowed to be appended and not overwritten. The decision-making non-rollback module is used to obtain the correction dataset, perform time alignment at a high time frequency granularity, and then match it point by point with the publishable forecast set to generate a correction record set, which is written into the correction layer to coexist with the interpolation layer. Subsequently, a visible value decision is performed on the interpolation layer records and the correction layer records to output a visible forecast point set. The visible values ​​corresponding to the historical forecast time identifiers that have been decided are not allowed to be rolled back under the subsequent release time identifiers.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, performs the method described in any one of claims 1-7.

10. A processor, characterized in that: The processor is configured to be connected to a memory; the memory is configured to store a computer program; the processor is configured to invoke and execute the computer program in the memory to perform the method described in any one of claims 1-7.

Images