A functional operator driven method for WRF process assembly

Abstract

This invention discloses a functional operator-driven method for WRF workflow assembly, specifically relating to the field of regional scale numerical weather prediction workflow driving. The method includes acquiring the JSON configuration file corresponding to the WRF task to be executed, a set of candidate functional operators, and the input data status information, configuration version information, and historical execution records associated with each candidate functional operator. A task configuration dictionary is generated according to preset field parsing rules, and semantic state features of each candidate functional operator are extracted based on the task configuration dictionary and historical execution records. The method outputs the current task state set and the candidate connection set. By performing learnable semantic adaptation judgment on the connection relationships between adjacent functional operators and identifying the source, time slice, configuration fields, and resource gaps before connection, the probability of intermediate results with satisfied formal dependencies but semantically incomplete connections continuing to propagate can be relatively reduced, thereby mitigating the problems of occasional failures and anomaly propagation in the WRF workflow.

Description

Technical Field

[0001] This invention relates to the field of regional scale numerical weather prediction workflow driving technology, and more specifically, to a functional operator-driven method for WRF workflow assembly. Background Technology

[0002] In the field of regional scale numerical weather prediction workflow-driven systems, the mainstream practice in the industry is to solve the problem of whether WRF forecast or simulation tasks can automatically complete preprocessing, data assimilation, model operation and post-processing in a predetermined order. Usually, programs such as geogrid, ungrib, metgrid, real, and wrrf are encapsulated into script calling units, and then combined with unified configuration files, file existence checks, return code judgments and parameter matching rules, and assembled in a preset order to form a one-click operation process. Taking the rolling business forecast scenario for high-performance computing platforms as an example, the system needs to continuously organize multiple rounds of WRF process assembly under the conditions that the arrival time of the driving field is not completely consistent, the observation data may be late for retransmission, the historical intermediate results can be reused at the opportune time, different tasks share the storage directory and the job queue fluctuates dynamically. It also needs to meet the hard constraint that each operator can automatically connect within a limited time and avoid the erroneous results from being passed on. However, under this constraint, existing technologies will expose the following problems: some processes still output abnormal results even when the files are complete, the path is correct, and the return code is normal, or the same script will occasionally fail in different batches of tasks. The root cause is that the existing driving method can only determine whether the operators are "formally connectable", but cannot identify whether the timeliness, source credibility, configuration version, consistency status and historical residual traces of the input data meet the "semantically connectable" condition. As a result, residual results of old tasks, low-reliability retransmission data or semantically mismatched intermediate states are mistakenly connected to subsequent operators and propagated along the process. The technical problem this application aims to solve is: how to perform learnable semantic adaptation judgment on the connection relationship between adjacent operators in the process of functional operator-driven assembly of WRF flow, so as to avoid the continued transmission of intermediate results that satisfy formal dependencies but have incomplete runtime semantics. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a functional operator-driven method for WRF process assembly. This method performs reinforcement learning based on semantic state differences to determine the connection relationship between adjacent functional operators, and generates bridging processing instructions, constructs a standardized operating environment, and writes back execution feedback to iteratively update the connection decision strategy based on the determination results, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a functional operator-driven method for WRF process assembly, comprising: S1. Obtain the JSON configuration file, candidate functional operator set, input data status information, configuration version information, and historical execution records associated with each candidate functional operator corresponding to the WRF task to be executed. Generate a task configuration dictionary according to the preset field parsing rules, and extract the semantic state features of each candidate functional operator based on the task configuration dictionary and historical execution records. Output the current task state set and candidate connection set. S2. Input the current task state set and candidate connection set into the pre-trained reinforcement learning decision model. The reinforcement learning decision model performs connection permission determination on the connection relationship of adjacent functional operators in each candidate connection set according to the action selection rule corresponding to the maximum cumulative reward, and outputs the connection decision result and the bridging processing instruction corresponding to the connection decision result. S3. For functional operators whose connection decision results are allowed to connect, execute input data linking, runtime directory instantiation, namelist configuration generation, historical intermediate result isolation verification or alternative data loading processing according to the bridging processing instructions, and output the standardized runtime environment and semantic closure mark corresponding to the target functional operator. S4. Pass the standardized operating environment and semantic closure mark into the target functional operator. The target functional operator calls the corresponding WRF executable program or its supporting pre- and post-processing program to perform the computation task. It also performs output integrity checks and result validity checks on the generated results and outputs the operator execution results, result semantic labels, and exception markers. S5. Construct reward feedback data based on operator execution results, result semantic labels, and anomaly markers, and update the policy parameters of the reinforcement learning decision model in combination with the connection decision results, so that the connection permission determination and bridging processing instructions of adjacent functional operators in subsequent WRF tasks are executed based on the updated policy parameters, and the updated connection decision policy is output.

[0005] In a preferred embodiment, S1 includes: S1-1. Input the JSON configuration file into the field mapping parsing unit, perform hierarchical splitting and key-value expansion processing on the WRF parameter field, extended execution field and path template field in the JSON configuration file according to the preset field parsing rules, generate field mapping results that are associated with each candidate functional operator, and output the task configuration dictionary. S1-2. Input the task configuration dictionary, the input data state information associated with each candidate functional operator, and the configuration version information into the semantic association construction unit. Generate the initial semantic state fragments of each candidate functional operator according to the preset correspondence between the functional operator and the input data source, configuration version fingerprint, and target output type, and output the initial semantic state set. S1-3. Input the initial semantic state set and historical execution records into the state correction unit. Perform credibility correction and connection constraint annotation on the initial semantic state fragments of each candidate functional operator according to the successful connection trajectory, abnormal propagation trajectory and intermediate result reuse trajectory in the historical execution records. Generate the semantic state features of each candidate functional operator and output the current task state set and candidate connection set.

[0006] In a preferred embodiment, S2 includes: S2-1. For each candidate connection edge in the candidate connection set, read the result semantic label, result data start and end time, result configuration field summary, and result resource usage record of the source functional operator, as well as the input requirement label, input time period requirement, input configuration field requirement, and execution resource requirement of the target functional operator. Then, compare them item by item according to label item, time item, field item, and resource item to obtain the label gap list, time gap list, field gap list, and resource gap list, and output the candidate connection difference table.

[0007] In a preferred embodiment, S2 further includes: S2-2. Input the candidate connection difference table into the reinforcement learning decision model. The reinforcement learning decision model calculates the cumulative reward value of connection permission action, bridging insertion action, degradation execution action and connection prohibition action for each candidate connection edge. The number of label gap items, time gap items, field gap items and resource gap items that can be eliminated after each action is applied to the corresponding candidate connection edge is summed to obtain the gap elimination amount. The number of additional processing steps required to execute the action and the additional resource consumption are summed to obtain the action cost amount. Output the action reward table and action cost table for each candidate connection edge. S2-3. For each candidate connection edge, first select the action with the largest cumulative return value from the corresponding action return table as the current action. Then, subtract the action cost from the gap elimination amount corresponding to the current action to obtain the net value of the current action. Write the current action into the WRF process connection sequence and recalculate the label gap list, time gap list, field gap list, and resource gap list of subsequent candidate connection edges one by one in the process order. If the total number of gap items of any subsequent candidate connection edge after recalculation is greater than the total number of gap items before writing, discard the current action and take the action with the second largest cumulative return value and repeat the execution until the target action with the largest net value of the current action is obtained without increasing the total number of gap items of subsequent candidate connection edges after writing is obtained. Output the connection decision result.

[0008] In a preferred embodiment, S2 further includes: S2-4. Based on the target action corresponding to each candidate connection edge in the connection decision results, read the label gap list, time gap list, field gap list and resource gap list of the candidate connection edge item by item, and convert missing input label items into input data chain completion operation, time misalignment items into time slice reorganization operation, field inconsistency items into configuration field rewriting operation, historical result residue items into intermediate result isolation operation, and resource insufficiency items into running directory reconstruction operation or alternative data loading operation. Generate bridging processing instructions according to the order of the candidate connection edges in the WRF process connection sequence, and output the bridging processing instructions.

[0009] In a preferred embodiment, S3 includes: S3-1. Read the target functional operator identifier, bridging processing instructions, task configuration dictionary, and input requirement list of the target functional operator from the connection decision results. For each input item in the input requirement list, determine the corresponding data source path, target link path, time slice identifier, and file naming item. Write the source path and target link path into the link mapping table and create the input data link. Concatenate the target functional operator identifier with the instance number to generate an independent running directory name and create an independent running directory. Then, write the field items corresponding to the target functional operator in the task configuration dictionary into the same position in the namelist template to obtain the target configuration file and output the candidate running environment. S3-2. Read the link mapping table, independent running directory, and target configuration file in the candidate running environment, and read the historical intermediate result list in the independent running directory. For each result item in the historical intermediate result list, extract the result generation task identifier, result source identifier, result time slice range, result configuration field item, and result semantic label. Then compare them item by item with the data source identifier, time slice identifier in the link mapping table, and the corresponding field item in the target configuration file. Remove historical intermediate results that have different result generation task identifiers, different result source identifiers, non-overlapping result time slice ranges, inconsistent result configuration field items, or result semantic labels that do not belong to the target functional operator input requirement list from the independent running directory. Write the corresponding replacement data into the independent running directory according to the bridging processing instructions, and output the corrected candidate running environment. S3-3. Read the revised candidate runtime environment and the input requirement list of the objective function operator. For each input item in the input requirement list, check whether there is a corresponding linked file or alternative data file in the independent runtime directory, whether there is a corresponding field in the target configuration file, whether the time slice identifier of the linked file or alternative data file falls within the time slice range corresponding to the input item, and whether the data source identifier of the linked file or alternative data file is consistent with the data source identifier corresponding to the input item. Write all the check results into the input closure table. When each input item in the input closure table has a corresponding file, corresponding field, consistent time slice, and consistent source, generate a standardized runtime environment and write the semantic closure mark as closed. When any input item in the input closure table is missing a corresponding file, missing a corresponding field, inconsistent time slice, or inconsistent source, write the input item into the unclosed item list and write the semantic closure mark as unclosed. Output the standardized runtime environment and semantic closure mark.

[0010] In a preferred embodiment, S4 includes: S4-1. Input the running directory, input file list, configuration file and semantic closure mark in the standardized running environment into the target functional operator. The target functional operator reads the program path, input directory, output directory and parallel parameters in the operator mapping table according to the operator identifier. Write each input file in the input file list into the input directory. Concatenate the program path, parallel parameters and configuration file path in sequence to generate the execution command. Generate the expected result file list according to the result file name list. Output the execution command and the expected result file list. S4-2. Input the execution command and the expected result file list into the execution unit, call the execution command to start the WRF program unit corresponding to the target function operator, after the program exits, traverse the output directory to generate the actual result file list, then search for the file with the same name in the actual result file list by each file name in the expected result file list, and search for the write record in the execution log by the path of the file with the same name, generate the file correspondence table, and output the file correspondence table and the actual result file list.

[0011] In a preferred embodiment, S4 further includes: S4-3. Input the file correspondence table, the list of actual result files, the configuration file, and the operator identifier into the verification unit. For each actual result file, read the time range, spatial grid identifier, variable field set, and source operator identifier, and compare them item by item with the start time, end time, grid field, output field, and operator identifier in the configuration file. When each actual result file satisfies the conditions of time range correspondence, spatial grid identifier correspondence, complete output field, and source operator identifier correspondence, generate the operator execution result and result semantic label. If any mismatch exists, generate an exception marker and output the operator execution result, result semantic label, and exception marker.

[0012] In a preferred embodiment, S5 includes: S5-1. Input the operator execution result, result semantic label, anomaly mark and connection decision result into the feedback construction unit. For each candidate connection edge, read the connection action identifier, bridging processing instruction, total number of result files, number of valid result files, number of abnormal result files and the total number of semantic items in the result semantic label that correspond to the input requirements of the objective function operator and the number of corresponding successful semantic items. Calculate the result completion rate, semantic correspondence rate and anomaly ratio respectively, and generate the connection feedback sequence. S5-2. Input the connection feedback sequence into the reward calculation unit, and obtain the immediate reward value for each candidate connection edge by adding the result completion rate and semantic correspondence rate and subtracting the abnormality ratio. Then, according to the connection order in the WRF process, accumulate the subsequent immediate reward values ​​from the back to the front to obtain the cumulative reward value of each candidate connection edge, and output the cumulative reward sequence.

[0013] In a preferred embodiment, S5 further includes: S5-3. Input the cumulative reward sequence and connection decision results into the parameter update unit. For each candidate connection edge, read the original policy parameter values ​​of all candidate actions. First, sum the original policy parameter values ​​to obtain the total parameter value. Then, multiply the cumulative reward value corresponding to the executed connection action with the total parameter value to obtain the reward allocation value. Add the reward allocation value to the original policy parameter value corresponding to the executed connection action to obtain the new policy parameter value of the action. Then, deduct the reward allocation value from the original policy parameter values ​​of the remaining unexecuted candidate actions according to their proportion in the total parameter value to obtain the new policy parameter values ​​of the remaining actions. Output the updated policy parameter set. S5-4. Input the updated set of strategy parameters into the strategy generation unit. For each candidate connection edge, read the corresponding new strategy parameter value according to the candidate action identifier. Sort the new strategy parameter values ​​from largest to smallest to form the action sequence of the candidate connection edge. Then write the action sequences of each candidate connection edge into the connection strategy table in the order of the WRF process to generate the updated connection decision strategy.

[0014] The technical effects and advantages of this invention are as follows: By performing learnable semantic adaptation judgment on the connection relationship of adjacent functional operators, and identifying the source, time slice, configuration field and resource gap before connection, the probability of intermediate results that satisfy formal dependencies but are not semantically closed can be relatively reduced and thus alleviate the problem of occasional failure and abnormal propagation in WRF process. By converting tag gaps, time gaps, field gaps, and resource gaps into bridging processing instructions, and performing input chaining, time slice reassembly, configuration rewriting, and residual isolation accordingly, the preparation process for the target operator can be kept consistent with the actual input requirements, thereby relatively improving the accuracy and executability of process connection. By verifying the task identifier, source identifier, time slice range, and configuration field items of historical intermediate results in an independent running directory, and removing mismatched results before loading replacement data, the interference caused by old task remnants or incorrect source results mixed into the current process can be relatively suppressed, thereby relatively improving the running stability in multi-round task parallel scenarios. By calling the corresponding WRF program unit by the objective function operator according to the operator mapping table, and establishing a file mapping table for the expected result file, the actual result file and the execution log, the program call relationship and the result generation relationship can have a clear correspondence basis, thereby relatively improving the traceability of result verification and the closure of the execution chain. By jointly feeding back the completion rate, semantic correspondence rate, and anomaly rate, and updating the strategy parameters of the reinforcement learning decision model accordingly, the connection permission determination and bridging method selection in subsequent tasks can gradually absorb the historical execution effect, thereby improving the adaptive decision-making ability and reuse efficiency of similar WRF process assembly tasks within a certain range. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method steps 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 The present invention provides a functional operator-driven method for WRF process assembly, comprising: S1. Obtain the JSON configuration file, candidate functional operator set, input data status information, configuration version information, and historical execution records associated with each candidate functional operator corresponding to the WRF task to be executed. Generate a task configuration dictionary according to the preset field parsing rules, and extract the semantic state features of each candidate functional operator based on the task configuration dictionary and historical execution records. Output the current task state set and candidate connection set. In this implementation, S1 is used to organize the original configuration content, input data status, and historical running information of the WRF task to be executed into a structured state result that can be directly used for subsequent connection determination. During processing, the entire JSON configuration file is not directly passed to subsequent steps. Instead, the fields with different purposes are first split into subsets of fields that can be read by each candidate functional operator. Then, each subset of fields is associated with the input source, time range, configuration version, and target output of the corresponding operator to form the initial semantic state of each candidate functional operator. Subsequently, historical successful connections, anomaly propagation, and intermediate result reuse are introduced to correct and constrain the initial semantic state, so that the output result has both current task attributes and historical running attributes, making it easy for subsequent steps to directly construct candidate connection edges and their differences. This implementation process includes the following steps: After mapping and parsing the JSON configuration file input fields, all fields are first split into WRF parameter fields, extended execution fields, and path template fields according to the purpose of the top-level fields. The WRF parameter fields are a set of fields used to generate various namelist configuration contents; the extended execution fields are a set of fields used to describe task identifiers, instance identifiers, data sources, parallel resources, and output directories; and the path template fields are a set of fields used to describe the driving field path, observation data path, and intermediate result directory path. Then, each type of field is expanded item by item, and for each field, a field name, field value, hierarchy, and applicable operator identifier are written. The applicable operator identifier is determined based on the actual function of the field. Fields related to the generation of geographic static fields are written in the geogrid pair. For each field record, fields related to driving field decoding are written to the corresponding field record in ungrib; fields related to intermediate field generation are written to the corresponding field record in metgrid; fields related to initial and boundary field generation are written to the corresponding field record in real; and fields related to integral prediction are written to the corresponding field record in wrrf. For scalar fields, the original field values ​​are directly retained. For multi-valued fields, multiple field records are written according to the order of the values. For path template fields, in addition to retaining the template string, the corresponding start time position, region position, and data source position in the template are also written synchronously. After all fields are expanded, they are merged according to the applicable operator identifier to obtain the field mapping results that correspond one-to-one with each candidate functional operator. Each mapping result is then written to the corresponding operator field subset in the task configuration dictionary. After inputting the task configuration dictionary, the input data status information associated with each candidate functional operator, and the configuration version information into the semantic association construction unit, the unit first reads the corresponding operator field subset for each candidate functional operator, and then reads the input data status information corresponding to that operator. The input data status information includes at least the data source identifier, start time, end time, file list, file naming field, and data type. Subsequently, all field records corresponding to the operator field subset are extracted from the configuration version information, sorted by field name, and then concatenated with the field name, field value, and field sequence number to generate the configuration version fingerprint of the candidate functional operator. This ensures that the same operator can form distinguishable status records under different configuration contents. Then, according to the correspondence between functional operators and input data sources, configuration version fingerprints, and target output types, an initial semantic state fragment is generated for the candidate functional operator. The initial semantic state fragment includes at least the source identifier, time slice range, target output type, configuration version fingerprint, resource usage record, and result semantic label field. The source identifier is taken from the main input source of the operator, the time slice range consists of the start time and the end time, the target output type is written according to the corresponding output type based on the operator category, the resource usage record is written from the parallel resource field in the task configuration dictionary, and the result semantic label field is initially set to empty. After performing the above processing on all candidate functional operators in sequence, an initial semantic state set is formed. After inputting the initial semantic state set and historical execution records into the state correction unit, the corresponding connection records in the historical execution records are first retrieved according to the source operator identifier and the target operator identifier. Then, the retrieved connection records are divided into successful connection trajectories, anomaly propagation trajectories, and intermediate result reuse trajectories. For successful connection trajectories, the number of successful connections under the same combination of source and target operator identifiers is counted, and the statistical results are written into the positive count field of the corresponding connection relationship. For anomaly propagation trajectories, the number of times subsequent execution anomalies occur due to source misconnection, time misconnection, field mismatch, or historical remnant mixing under the same combination is counted, and the statistical results are written into the negative count field. For intermediate result reuse trajectories, the number of times historical intermediate results are reused under the same combination and the results remain valid is counted, and the statistical results are written into the negative count field. The results are written into the reuse count field. After the statistics are completed, the positive count field, negative count field, and reuse count field are merged into the initial semantic state fragment of each candidate functional operator. Then, the connection constraint annotation is written according to the anomaly source recorded in the anomaly propagation trajectory. Among them, the connection relationship with the source inconsistency anomaly is written into the source consistency constraint, the connection relationship with the time coverage insufficiency anomaly is written into the time coverage constraint, the connection relationship with the configuration content inconsistency anomaly is written into the version consistency constraint, and the connection relationship with the historical result residue mixing anomaly is written into the residue isolation constraint. After the correction is completed, the semantic state features of each candidate functional operator are generated and written into the current task state set in the order of the WRF process. Then, the candidate connection set is constructed according to the order of adjacent candidate functional operators in the process. Through the above processing, the field content, input data status information, and historical execution records in the original JSON configuration file are organized into a unified task configuration dictionary, initial semantic state set, current task state set, and candidate connection set. This allows subsequent steps to directly read the source identifier, time slice range, configuration version fingerprint, target output type, and historical connection attributes when processing connection relationships, eliminating the need to infer field attribution or perform historical relationship judgments at runtime. Furthermore, after the positive count field, negative count field, reuse count field, and various connection constraints are written in this step, subsequent steps can directly generate tag gap items, time gap items, field gap items, and resource gap items based on this information, ensuring consistency in field and status sources between steps. In practical applications: for tasks containing geogrid and ungrid... For the rolling forecast task with five candidate functional operators (b), metgrid, real, and wrrf, the static geographic parameter field, driving field path field, physical scheme field, time control field, and parallel resource field can be extracted from the JSON configuration file and merged into the field subset of the corresponding operator. Then, the initial semantic state fragment of each operator is generated by combining the source identifier, start time, and end time of the current driving field file and observation data file. Subsequently, the historical execution records are queried. If there are multiple instances of insufficient driving field time coverage between metgrid and real, a time coverage constraint is written to the connection relationship. If historical intermediate results are successfully reused multiple times between real and wrrf, a reuse count field is written. The final set of current task state and candidate connection set can be directly used for subsequent connection determination steps.

[0018] S2. Input the current task state set and candidate connection set into the pre-trained reinforcement learning decision model. The reinforcement learning decision model performs connection permission determination on the connection relationship of adjacent functional operators in each candidate connection set according to the action selection rule corresponding to the maximum cumulative reward, and outputs the connection decision result and the bridging processing instruction corresponding to the connection decision result. In this implementation, S2 is used to determine the connectability of each candidate connection edge and generate a bridging method based on the current task state set and candidate connection set. During processing, the existence of a file is not directly used as the basis for connection. Instead, the existing output state of the source functional operator is compared item by item with the input requirement state of the target functional operator to form label gaps, time gaps, field gaps, and resource gaps. The gap results of each candidate connection edge are then compiled into a unified difference record, which is input into the reinforcement learning decision model to calculate the cumulative reward value of each candidate action. Subsequently, combining the gap elimination amount and action cost, the target action is determined for each candidate connection edge. The target action is written into the WRF process connection sequence and then the backward impact is recalculated to avoid the amplification of gaps in the later stages due to actions in the earlier stages. Finally, the difference items corresponding to the target action are converted into executable bridging processing instructions, providing clear input for the subsequent construction of a standardized operating environment. This implementation process includes the following steps: For each candidate connection edge in the candidate connection set, first read the result semantic label, result data start time, result data end time, result configuration field summary, and result resource usage record of the source functional operator; then read the input requirement label, input time period requirement, input configuration field requirement, and execution resource requirement of the target functional operator. The result semantic label includes at least the source identifier, result time range, spatial grid identifier, variable field set, and configuration version fingerprint; the input requirement label includes at least the target source identifier, target time range, target grid identifier, target variable field set, and target configuration version fingerprint. Then, compare each item by label, time, field, and resource. Label items are defined by the combination of label name and label value. If the result semantic label of the source functional operator does not contain a label with the same name and value required by the target functional operator, this item is added to the label gap list. The time item is defined by start time, end time, and end time. The time interval, time period, and time period are compared separately. When the time range of the result of the source functional operator cannot cover the input time period requirement of the target functional operator, the corresponding unmet time item is written into the time gap list. Field items are listed as field names. When the field name required by the target input is not found in the result configuration field summary or the field value corresponding to the same field name is inconsistent, the field name is written into the field gap list. Resource items are compared separately according to the number of CPU cores, memory capacity, and disk space. When the result resource usage record cannot meet the execution resource requirements of the target functional operator, the corresponding resource item is written into the resource gap list. After the above comparison is completed, the label gap list, time gap list, field gap list, and resource gap list corresponding to each candidate connection edge are written into the candidate connection difference table. Each difference record contains at least the source operator identifier, the target operator identifier, the four types of gap lists, and the sequence number of the candidate connection edge in the WRF process connection sequence. After inputting the candidate connection difference table into the reinforcement learning decision model, a state record is constructed for each candidate connection edge. The state record consists of the number of label gap items, time gap items, field gap items, resource gap items, source operator identifier, target operator identifier, and historical connection count fields for that candidate connection edge. The reinforcement learning decision model outputs the cumulative reward values ​​for each state record for connection permission actions, bridging insertion actions, degraded execution actions, and connection prohibition actions. The degraded execution action is limited to one of the following: alternative data loading, skipping the current bridging step and executing directly, or switching to a cached intermediate result set. Subsequently, the gap elimination amount and action cost are calculated for each candidate action. The gap elimination amount is obtained by reading the corresponding action. The impact table is obtained by summing the number of label gaps, time gaps, field gaps, and resource gaps that can be eliminated after applying the action to the current candidate connection edge. The action cost is calculated using a unified itemization method, where adding a new processing step is recorded as a step cost item, adding a new type of resource occupation is recorded as a resource cost item, and adding an execution duration is recorded as a time cost item. Finally, the number of step cost items, resource cost items, and time cost items are summed to obtain the action cost. After processing, an action reward table and an action cost table are generated for each candidate connection edge. The action reward table records the cumulative reward value corresponding to each candidate action, and the action cost table records the gap elimination amount and action cost corresponding to each candidate action. For each candidate connection edge, first select the action with the largest cumulative reward value from its action reward table as the current action. Then, read the gap elimination amount and action cost of the current action from the corresponding action cost table, and subtract the action cost from the gap elimination amount to obtain the net value of the current action. Subsequently, write the current action into the position corresponding to the current candidate connection edge in the WRF process connection sequence. Starting from the next candidate connection edge after the current candidate connection edge, recalculate the label gap list, time gap list, field gap list, and resource gap list of subsequent candidate connection edges in the order of the WRF process connection sequence. During the recalculation, the connection results and bridging results already generated by the current action are used as the basis for the previous output, and the item-by-item comparison is performed again in the same way as S2-1. If the total number of gap items of any subsequent candidate connection edge after recalculation is greater than the number written, the action is considered a candidate connection edge. If the total number of gaps in the candidate edge is determined, the current action is deemed to amplify the burden on subsequent connections. This current action is then removed from the WRF process connection sequence, and the action with the second-largest cumulative return value for that candidate edge is selected as the new current action. The above writing, recalculation, and comparison process is repeated. If multiple actions do not increase the total number of gaps in subsequent candidate edges after writing, the action with the largest net value among these actions is selected as the target action. If only one action satisfies the condition that writing does not increase the total number of gaps in subsequent candidate edges, this action is directly determined as the target action. Once the target action for the current candidate edge is determined, it is written into the corresponding connection action field in the connection decision result, and the next candidate edge in the WRF process connection sequence is processed until all candidate edges are processed, resulting in a complete connection decision result. Based on the target action corresponding to each candidate connection edge in the connection decision results, the label gap list, time gap list, field gap list, and resource gap list of the candidate connection edge are read item by item, and bridging processing instructions are generated according to a fixed transformation order. The fixed transformation order is to first process the isolation of intermediate results, then process the input data chaining or replacement data loading, then process the time slice reorganization, and finally process the configuration field rewriting and the running directory reconstruction. Specifically, for missing input label items in the label gap list, if the data source corresponding to the label item exists in the current task data source list, it is converted to an input data chaining operation; if it does not exist, it is converted to a replacement data loading operation. For each time misalignment item in the time gap list, the target function operator is used to process the missing input label items. The required start time, end time, and time interval generate a time slice reorganization operation. For inconsistencies in the field gap list, the corresponding field record in the task configuration dictionary is located by field name, and a configuration field rewrite operation is generated. For differences marked as historical result residues in the candidate connection difference table, an intermediate result isolation operation is generated. For resource shortages in the resource gap list, if a reusable directory structure exists in the current task directory, a running directory reconstruction operation is generated; otherwise, an alternative data loading operation is generated. After all difference items are converted, they are written into a bridging processing instruction sequence according to the order of the candidate connection edges in the WRF process connection sequence to form a bridging processing instruction, which is then output to the subsequent runtime environment construction step. Through the above processing, candidate connection edges no longer rely on manual experience or simple document existence judgments. Instead, they are based on the item-by-item differences between the output state of the source functional operator and the input requirements of the target functional operator to determine connection permission and generate bridging actions. This ensures that the connection decision results simultaneously consider the gap elimination capability of the current edge and the gap changes of subsequent edges, avoiding the problem of local optima in previous connection edges but overall process imbalance. Furthermore, since the bridging processing instructions are directly converted from the difference items in a fixed order, subsequent steps can directly execute input data chain completion, time slice reorganization, configuration field rewriting, intermediate result isolation, and runtime directory reconstruction, reducing intermediate interpretation steps and maintaining consistency in data sources, state sources, and constraint sources between preceding and subsequent steps. In practical applications: for a WRF process connection sequence that connects metagrid to real and then from real to wrrf, the time range and grid labels in the metagrid output results can be read first for the candidate connection edges between metagrid and real. The system identifies the set of variable fields and configuration version fingerprints, and then compares them item by item with the input time period requirements, grid requirements, variable field requirements, and configuration field requirements of the real system. If insufficient time coverage and inconsistencies in configuration fields are found, they are written into the time gap list and field gap list, respectively. Subsequently, the reinforcement learning decision model provides the cumulative reward values ​​for connection permission actions, bridging insertion actions, degradation execution actions, and connection prohibition actions, and calculates whether the bridging insertion action can eliminate time gap items and field gap items. At the same time, two additional processing steps, time slice reorganization and configuration field rewriting, are required to obtain the corresponding gap elimination amount and action cost. This action is then written into the WRF process connection sequence, and the subsequent gaps between the real system and the WRF system are recalculated. If the total number of subsequent gap items does not increase, the bridging insertion action is determined as the target action, and the time gap items are further converted into time slice reorganization operations, and the field inconsistency items are converted into configuration field rewriting operations, forming a bridging processing instruction that can be directly called for subsequent standardized runtime environment construction.

[0019] S3. For functional operators whose connection decision results are allowed to connect, execute input data linking, runtime directory instantiation, namelist configuration generation, historical intermediate result isolation verification or alternative data loading processing according to the bridging processing instructions, and output the standardized runtime environment and semantic closure mark corresponding to the target functional operator. In this embodiment, S3 is used to further implement the connection permission result and bridging processing instructions into runtime preparation results that can be directly called by the target functional operator, based on the connection decision result already determined. During processing, firstly, input data links, independent running directories, and target configuration files are constructed according to the input requirement list of the target functional operator to form a candidate runtime environment. Then, the historical intermediate results in the candidate runtime environment are verified item by item, and old results that do not belong to the current target functional operator, results with incorrect sources, results with incorrect time periods, and results with incorrect configurations are removed, and corresponding replacement data is added according to the bridging processing instructions. Finally, an input closure check is performed on the corrected candidate runtime environment to confirm that each input item has a corresponding file, corresponding field, time slice consistency, and source consistency. After confirming that each input item has a corresponding file, corresponding field, time slice consistency, and source consistency, a standardized runtime environment and semantic closure mark are output to ensure that the execution basis of subsequent computing tasks is complete, the source is clear, and the state is closed. This implementation process includes the following steps: After reading the target functional operator identifier, bridging processing instructions, task configuration dictionary, and input requirement list of the target functional operator from the connection decision results, the corresponding data source path, target link path, time slice identifier, and file naming item are determined for each input item in the input requirement list. Specifically, the data source path is obtained by expanding the data source identifier and time slice identifier in the path template field of the task configuration dictionary. Specifically, the start time bit in the path template is replaced with the current task start time, the data source bit is replaced with the data source identifier of the input item, and the region bit is replaced with the region identifier of the current target functional operator, thus obtaining the actual data source path for the input item. Then, the data source path and target link path corresponding to each input item are written into the link mapping table. Before creating the independent running directory, the target link path is generated item by item according to the link mapping table. Input the data link; then generate an independent running directory name. The independent running directory name is obtained by concatenating the target functional operator identifier, the current WRF task identifier, and the sequence number of the target functional operator in the current WRF process. This ensures that the running directories corresponding to different operators within the same task and the same operator across different tasks can be distinguished. Then, create an independent running directory based on this independent running directory name. On this basis, read the field items corresponding to the target functional operator in the task configuration dictionary, locate the corresponding positions in the namelist template according to the field name, write the scalar fields directly to the corresponding positions, write the multi-value fields to the corresponding multiple positions according to the value position order, and write the time field to the start time position, end time position, and time interval position respectively to generate the target configuration file. After completing the above processing, write the link mapping table, the independent running directory, and the target configuration file together into the candidate running environment. After reading the link mapping table, standalone running directory, and target configuration file in the candidate runtime environment, the historical intermediate result list in the standalone running directory is then read. This list consists of all result files in the standalone running directory except for the link files created in the current step and the configuration files written in the current step. Subsequently, for each result item in the historical intermediate result list, the result generation task identifier, result source identifier, result time slice range, result configuration field, and result semantic tag are extracted. The result generation task identifier is taken from the directory registration record corresponding to the result item; the result source identifier and result time slice range are taken from the result semantic tag; and the result configuration field is taken from the configuration record saved when the result item was generated. Then, the extracted results are compared item by item with the data source identifier, time slice identifier in the link mapping table, and the corresponding field in the target configuration file. If the result generation task identifier is not equal to the current WRF task identifier, the result item is determined to be a residual result from an old task. If the result source identifier is not equal to the data source identifier of the corresponding input item in the link mapping table, the result item is determined to be a residual result from an old task. If the time slice range of the result cannot cover the time slice range of the corresponding input item in the target functional operator input requirement list, the result item is determined to be an incorrect time slice result. If any field name in the result configuration field item is inconsistent with the value of the same field in the target configuration file, the result item is determined to be an incorrect configuration result. If the output type or source identifier in the result semantic tag does not belong to the set of input items that the target functional operator input requirement list can accept, the result item is determined to be a mismatch result. For any of the above determinations that are true, the historical intermediate result is removed from the independent running directory and written to the isolation directory. At the same time, the current WRF task identifier, the target functional operator identifier, and the file name of the removed result are registered in the isolation record. After the removal is completed, the replacement data loading is performed item by item according to the bridging processing instructions. For the replacement data required to be added in the bridging processing instructions, the actual replacement data path is located in the task configuration dictionary according to the replacement data source identifier and time slice identifier. Then, the corresponding replacement data file is written to the file name position with the same as the original result item in the independent running directory, thereby obtaining the corrected candidate running environment. After reading the revised candidate runtime environment and target functional operator input requirement list, each input item in the input requirement list is checked for the existence of a corresponding linked file or alternative data file in the independent runtime directory, the existence of a corresponding field in the target configuration file, whether the time slice identifier of the linked file or alternative data file falls within the time slice range corresponding to the input item, and whether the data source identifier of the linked file or alternative data file is consistent with the data source identifier corresponding to the input item. The results of each check are written into the input closure table. Among them, the check of corresponding files is completed by searching the file name item in the independent runtime directory one by one; the check of corresponding fields is completed by searching the field name item in the target configuration file one by one; the check of time slice consistency is completed by comparing the file registration start time and end time with the start time and end time required by the input item. When the file start time is not later than the start time of the input item and the file end time is not later than the start time of the input item, the check of time slice consistency is completed. If the time slice of the file is determined to cover the time slice range corresponding to the input item no earlier than the end time of the input item, the consistency of the source is checked by comparing the source identifier of the file registration and the source identifier of the input item. After checking all input items, if all input items in the input closure table simultaneously satisfy the conditions of having a corresponding file, having a corresponding field, having a consistent time slice, and having a consistent source, then the corrected candidate runtime environment is determined as the standardized runtime environment, and the semantic closure mark is written as closed. If any input item in the input closure table is missing a corresponding file, missing a corresponding field, has an inconsistent time slice, or has an inconsistent source, then the input item is added to the unclosed item list, and the semantic closure mark is written as unclosed. At the same time, the current corrected candidate runtime environment is retained as a pending return state. Finally, the standardized runtime environment and the semantic closure mark are output. The unclosed item list is returned along with the semantic closure mark for subsequent steps to regenerate the corresponding difference items. Through the above processing, the connection decision results no longer remain at the action selection layer, but are further implemented as actually executable input data links, independent running directories, target configuration files, isolated historical intermediate results, and alternative data loading results. This ensures that the objective functional operator has a complete, independent, and traceable operational foundation before entering the execution phase. Simultaneously, historical intermediate results are cleaned and isolated through item-by-item comparison of task identifiers, source identifiers, time slice ranges, configuration field items, and result semantic tags, reducing the risk of old task remnants, incorrect source results, incorrect time-period results, and incorrect configuration results continuing to be passed to subsequent steps. Furthermore, by registering file existence, field existence, time slice consistency, and source consistency item by item through the input closure table, the generation of semantic closure markers has a clear basis, and the list of unclosed items can be directly returned as input for subsequent difference correction, ensuring consistency between the state chain and field chain between preceding and following steps. In practical applications: when the objective functional operator is real, the metagrid can be determined first based on its input requirement list. The source paths of the result files and background field files are established, and input data links are created in the corresponding target link paths. Then, the independent running directory name is generated by concatenating the real, the current task identifier, and the real's sequence number in the process. The corresponding time control field, grid field, and physical scheme field are written to obtain the target configuration file. Subsequently, the historical intermediate results in the independent running directory are read. If a result file is found to originate from the previous round of tasks, or its time slice range only covers the first half after the current task starts reporting, or the time interval in its configuration field is inconsistent with the current target configuration file, the result file is moved to the isolated directory, and the replacement data of the same time slice is loaded according to the bridging processing instruction. After the correction is completed, each input item corresponding to the real is checked to see if it has the corresponding file, corresponding field, time slice consistency, and source consistency. If all are satisfied, a standardized running environment that can be directly run by the real is output, and the semantic closure mark is written as closed. If any item is not satisfied, a list of unclosed items is output, and the semantic closure mark is written as unclosed.

[0020] S4. Pass the standardized operating environment and semantic closure mark into the target functional operator. The target functional operator calls the corresponding WRF executable program or its supporting pre- and post-processing program to perform the computation task. It also performs output integrity checks and result validity checks on the generated results and outputs the operator execution results, result semantic labels, and exception markers. In this embodiment, S4 is used to drive the target functional operator to execute the corresponding WRF program unit based on the standardized operating environment already constructed and the semantic closure marker already given, and to perform file-level and semantic-level verification on the execution result. During processing, the target functional operator first reads the program path, input directory, output directory, parallel parameters, and result file name list uniquely corresponding to itself from the operator mapping table based on the operator identifier. Then, it organizes the input files and configuration files in the standardized operating environment into the corresponding locations to form an execution command that can be directly invoked. Subsequently, the execution unit starts the corresponding program, and after the program exits, it generates a file correspondence table based on the expected result file list and execution log to ensure that the result files correspond one-to-one with the execution process. Finally, it performs item-by-item verification on the time range, spatial grid identifier, variable field set, and source operator identifier of the actual result file. Only when the content of the result file is consistent with the configuration file and operator identifier is the operator execution result and result semantic label generated; otherwise, an anomaly marker is generated. This implementation process includes the following steps: After inputting the runtime directory, input file list, configuration file, and semantic closure flag from the standardized runtime environment into the target function operator, the semantic closure flag is checked first. If the semantic closure flag is closed, subsequent processing continues; if the semantic closure flag is not closed, the execution of the current target function operator is stopped, and the unclosed state is directly passed to the exception flag. When the semantic closure flag is closed, the target function operator reads the program path, input directory, output directory, parallel parameters, and result file name list uniquely corresponding to the operator identifier from the operator mapping table based on the operator identifier. The operator mapping table is a pre-established static mapping table, where each operator identifier corresponds to only one WRF program unit or one matching pre- and post-processing program unit. Subsequently, each input file in the input file list is processed according to its format. The file name and target path mapping is written to the input directory. The writing method includes copying or linking. When the running directory and the input directory are in the same file system, the linking method is used. When the running directory and the input directory are in different file systems, the copying method is used. After the input file is written, the program path, parallel parameters, and configuration file path are concatenated in sequence to form the execution command. The number of processes and threads in the parallel parameters are directly taken from the corresponding fields in the operator mapping table. The configuration file path is taken from the actual storage path of the configuration file in the current running directory. Then, the expected result file list is generated item by item according to the result file name list. Each file name in the expected result file list is consistent with the output rules of the WRF program unit corresponding to the current target function operator. Finally, the execution command and the expected result file list are output. After the execution command and the list of expected result files are input into the execution unit, the execution unit calls the execution command to start the WRF program unit corresponding to the objective function operator. At program startup, it records the startup time and running directory path, and simultaneously creates an execution log file for continuous writing to standard output and standard error output. After the program exits, the execution unit traverses all files in the output directory, using the filenames in the result filename list as filtering criteria, retaining only files with the same name as those in the expected result file list, as well as fragmented files generated with that same name as a prefix, forming the actual result file list. Then, using each filename in the expected result file list as a search term, it searches for files with the same name in the actual result file list item by item. If a file with the same name is found, it is... The path of a file with the same name and its corresponding expected file name are written into a file mapping table. If no file with the same name is found, the actual file path corresponding to the expected file name is left empty in the file mapping table. After the file name search is completed, the actual file path in the file mapping table is used as the search item to search for the corresponding write records in the execution log. The write record must include at least the file write path and the write time. When a write path with the same actual file path exists in the execution log, the write record is written into the corresponding item in the file mapping table. When it does not exist, the item is recorded as a missing write record. This yields the file mapping table and the list of actual result files. The file mapping table is used to represent the correspondence between the expected result files, the actual result files, and the log write records. After inputting the file mapping table, actual result file list, configuration file, and operator identifier into the verification unit, the time range, spatial grid identifier, variable field set, and source operator identifier are read from each actual result file in the file mapping table. The time range is obtained by reading the start and end time fields in the actual result file; the spatial grid identifier is obtained by reading the grid number and region number fields; the variable field set is obtained by traversing the variable name set in the result file; and the source operator identifier is obtained by reading the operator registration field written to the result file. Then, the start and end times of each actual result file are compared with the start and end times in the configuration file. If the start time of an actual result file is not earlier than the start time and the end time is not later than the end time, the time range is considered to correspond. The spatial grid identifiers of each actual result file are compared with the grid fields in the configuration file. If both the grid number and region number are consistent, the spatial grid identifiers are considered to correspond. Finally, the variable field sets of each actual result file are compared item by item with the output fields in the configuration file. When each output field in the file can be found in the variable field set with the same name, the output field is considered complete. The source operator identifier of each actual result file is compared with the current operator identifier. When they match, the source operator identifier is considered to correspond. When all actual result files meet the requirements of time range correspondence, spatial grid identifier correspondence, complete output fields, and source operator identifier correspondence, and each expected result file in the file correspondence table has an actual file path and a write record, the operator execution result and result semantic label are generated. The result semantic label includes at least the source operator identifier, result time range, spatial grid identifier, variable field set, and configuration version fingerprint. The configuration version fingerprint is generated by reading the field records in the current configuration file, sorting them by field name, and then concatenating the field name, field value, and field sequence number. If any actual result file does not meet any of the above correspondences, or if there is an expected result file in the file correspondence table where the actual file path is not found or the write record is missing, an exception mark is generated, and the non-compliance item is written to the exception item list. Finally, the operator execution result, result semantic label, and exception mark are output. Through the above processing, the execution relationship between the objective function operator and the WRF program unit is established as a verifiable program call chain, file generation chain, and result verification chain, making the source of the execution command, the source of the result file, and the correspondence between the result content and the configuration content all have clear basis. Furthermore, the item-by-item construction of the expected result file list, the actual result file list, and the file correspondence table ensures that whether the program execution truly generates the target result no longer depends on a single exit state, but is simultaneously constrained by filename matching, log writing records, and file content verification. The result semantic label, generated synchronously with the operator execution result in this step, can be directly used as input for subsequent state updates and connection corrections, maintaining consistency of semantic fields in the preceding and following steps. In practical applications: when the objective function operator is real, the real program path, real input directory, and real output directory are first read from the operator mapping table by the operator identifier corresponding to real. The parallel parameters and result file name list are generated, and then the metgrid result file and configuration file are written to the real input directory. The real execution command is then generated by concatenating them. Subsequently, the execution unit starts the real program. After the program exits, the actual result files corresponding to wrfinput and wrfbdy are selected from the real output directory, and the write records of these two files are retrieved in the execution log to form a file correspondence table. Then, the verification unit reads the time range, grid identifier, variable field set and source operator identifier in wrfinput and wrfbdy, and compares them item by item with the start time, end time, grid field, output field and real operator identifier in the configuration file. When both meet the correspondence, the real operator execution result and result semantic label are generated. If any file is not generated, no write record appears in the log, or variable field is missing, an anomaly mark is generated and written to the anomaly item list.

[0021] S5. Construct reward feedback data based on operator execution results, result semantic labels, and anomaly markers, and update the policy parameters of the reinforcement learning decision model in combination with the connection decision results, so that the connection permission determination and bridging processing instructions of adjacent functional operators in subsequent WRF tasks are executed based on the updated policy parameters, and the updated connection decision policy is output. In this implementation, S5 is used to write back the result file information, semantic correspondence information, and anomaly information obtained during the actual execution phase of the current WRF task to the reinforcement learning decision model. This allows the connection decision to move beyond static action selection and instead adjust the action order of subsequent candidate connection edges based on the actual execution effect of the completed task. During processing, the execution results and semantic results corresponding to each candidate connection edge are first organized into directly calculable feedback records. Then, based on the result completion rate, semantic correspondence rate, and anomaly percentage, immediate reward values ​​and cumulative reward values ​​are generated. Subsequently, the cumulative reward values ​​are allocated to the policy parameters of executed and unexecuted actions according to the connection decision results, resulting in an updated set of policy parameters. Finally, based on the updated set of policy parameters, an action sequence is regenerated for each candidate connection edge, forming a connection decision strategy that can be directly invoked by subsequent WRF tasks. This implementation process includes the following steps: After inputting the operator execution results, result semantic labels, anomaly markers, and connection decision results into the feedback construction unit, connection feedback records are first established edge by edge according to the order of candidate connection edges in the WRF process. For each candidate connection edge, the connection action identifier and bridging processing instruction corresponding to the edge are read, and then the total number of result files, the number of valid result files, and the number of abnormal result files obtained after the execution of the target functional operator connected to the edge are read. The total number of result files is taken from the expected result file list of the target functional operator, the number of valid result files is taken from the number of actual result files that have passed S4 verification, and the number of abnormal result files is taken from the number of result files written to the anomaly item list. Subsequently, the total number of semantic items corresponding to the input requirements of the target functional operator and the number of corresponding successful semantic items are read from the result semantic labels. The total number of semantic items is determined by the source identifier and time in the target functional operator input requirement list. The scope item, spatial grid item, variable field item, and configuration version item are statistically obtained item by item. The number of successful semantic items is obtained by comparing the same items in the result semantic tags with the input requirement list item by item and counting the number of consistent items. On this basis, the result completion rate, semantic correspondence rate, and anomaly ratio are calculated respectively. The result completion rate is calculated by dividing the number of valid result files by the total number of result files, the semantic correspondence rate is calculated by dividing the number of successful semantic items by the total number of semantic items, and the anomaly ratio is calculated by dividing the number of abnormal result files by the total number of result files. If the total number of result files is zero, the result completion rate and anomaly ratio are both set to zero. If the total number of semantic items is zero, the semantic correspondence rate is set to zero. After completing each calculation, the connection action identifier, bridging processing instruction, result completion rate, semantic correspondence rate, and anomaly ratio are written into the connection feedback record of the corresponding candidate connection edge. Finally, the connection feedback sequence is generated according to the WRF process order. After inputting the connection feedback sequence into the reward calculation unit, the immediate reward value is calculated edge by edge in the order of the connection feedback sequence. For each candidate connection edge, the immediate reward value of the candidate connection edge is obtained by adding the result completion rate to the semantic correspondence rate and subtracting the anomaly ratio, and the immediate reward value is written into the corresponding connection feedback record. Then, the cumulative reward value is calculated edge by edge from back to front according to the connection order in the WRF process. Specifically, the immediate reward value of the last candidate connection edge in the connection feedback sequence is directly written as the cumulative reward value of the edge. Then, the previous candidate connection edge is taken as the current edge, and the immediate reward value of the current edge is added to the cumulative reward value of the next candidate connection edge to obtain the cumulative reward value of the current edge. The same method is used to process the edges forward until the first candidate connection edge in the connection feedback sequence is processed. After completion, the cumulative reward values ​​corresponding to each candidate connection edge are written into the cumulative reward sequence in the original order. Each item in the cumulative reward sequence includes at least the candidate connection edge identifier, the immediate reward value, and the cumulative reward value, so that the previous candidate connection edge reflects the execution effect of the current edge and the execution effect of the subsequent edge when updating the strategy parameters. After inputting the cumulative reward sequence and connection decision results into the parameter update unit, the original policy parameter values ​​of all candidate actions for each candidate connection edge are read. All candidate actions include at least connection permission actions, bridging insertion actions, degradation execution actions, and connection prohibition actions. Then, the original policy parameter values ​​of all current candidate connection edges are summed to obtain the total parameter value. Next, the executed connection action identifier of the candidate connection edge is read from the connection decision results, and the cumulative reward value corresponding to the candidate connection edge is read from the cumulative reward sequence. Then, the cumulative reward value is multiplied by the total parameter value to obtain the reward allocation value, and this reward allocation value is added to the original policy parameter value corresponding to the executed connection action to obtain the new policy parameter value for that action. Finally, for the remaining unexecuted candidate actions, the original policy parameter values ​​of each unexecuted candidate action are read first, and then the original policy parameter values ​​of each unexecuted candidate action are divided... Divide the reward allocation value by the sum of the original strategy parameter values ​​of all unexecuted candidate actions to obtain the deduction ratio of each unexecuted candidate action. Then multiply the reward allocation value by the deduction ratio of each unexecuted candidate action to obtain the deduction value of the corresponding action. Subtract the corresponding deduction value from the original strategy parameter value of the action to obtain the new strategy parameter value of the action. If the new strategy parameter value of an unexecuted candidate action is less than zero, write the new strategy parameter value as zero and redistribute the remaining amount of the unexecuted action to the other unexecuted candidate actions with new strategy parameter values ​​greater than zero. Deduct the remaining amount according to the original strategy parameter value ratio until all remaining amounts are distributed or the new strategy parameter values ​​of the other unexecuted candidate actions all drop to zero. After calculating the new strategy parameter values ​​of all candidate actions, write the new strategy parameter values ​​of all candidate actions of the candidate connection edge into the updated strategy parameter set according to the action identifier. After inputting the updated set of policy parameters into the policy generation unit, the new policy parameter values ​​corresponding to each candidate connection edge are read according to the candidate action identifier, and then sorted in descending order of the new policy parameter values ​​to form the action sequence of that candidate connection edge. When two candidate actions have the same new policy parameter value, the connection actions already executed in the current WRF task are retained first, and the remaining actions are arranged in order of action identifier. Subsequently, the action sequences of all candidate connection edges are written into the connection policy table in the order of the WRF process. Each record in the connection policy table includes at least the candidate connection edge identifier, the action sequence, and the new policy parameter values ​​corresponding to each action. After the connection policy table is written, the updated connection decision policy is generated. The updated connection decision strategy is stored in the strategy storage area that can be read by subsequent WRF tasks. Before the next WRF task enters S2, the action sequence in the connection strategy table is read according to the candidate connection edge identifier, and the cumulative reward value of the corresponding action is called in the order of the action sequence. When the first action in the connection strategy table is consistent with the first action output by the reinforcement learning decision model, the first action is directly taken as the priority action. When the two are inconsistent, the action order of the candidate connection edge in the connection strategy table is readjusted according to the cumulative reward value output by the reinforcement learning decision model. This allows the updated connection decision strategy to retain the priority of the previously executed effective actions and to continue to accept the correction from the feedback of the new task. Through the above processing, the result file state, semantic correspondence state, and abnormal state formed during the actual execution phase of the current WRF task are transformed into a unified connection feedback sequence, cumulative reward sequence, updated policy parameter set, and connection policy table. This ensures that the policy parameter update of the reinforcement learning decision model has a clear data source, calculation order, and allocation rules. Furthermore, after the cumulative reward value is accumulated edge by edge from back to front, the parameter update of the preceding candidate connection edge can reflect the chain effect of the subsequent connection edge. The new policy parameter values ​​of the candidate actions that have not been executed are obtained through percentage deduction and residual redistribution, ensuring that the parameter changes between each candidate action remain closed. After the updated connection decision policy is saved in the form of a connection policy table, it can directly participate in the action priority determination in the next WRF task, ensuring that the policy continuity relationship between the preceding and following tasks is clear and traceable. In practical applications: when the metagrid to real candidate connection edge of a WRF task performs a bridging insertion action, the objective function operator generates two expected result files, which... Both result files passed validation, resulting in a completion rate of 1. The semantic labels of the results were compared item by item with the five items in the real input requirements: source identifier, time range, spatial grid, variable field, and configuration version. Four items matched, resulting in a semantic correspondence rate of four divided by five. If the number of abnormal result files was zero, the abnormality rate was zero, and the immediate reward value for the candidate connection edge was calculated as one plus four divided by five minus zero. This immediate reward value was then accumulated with the immediate reward values ​​of subsequent real-to-WRF candidate connection edges from back to front, yielding a cumulative reward value. Subsequently, the original policy parameter values ​​of the four candidate actions for the candidate connection edge were read. The reward allocation value was increased for the executed bridging insertion action, and the reward allocation value was deducted from the remaining three unexecuted actions according to their proportion of the original policy parameter values, resulting in four new policy parameter values. Finally, the four actions were sorted from largest to smallest according to the new policy parameter values ​​and written into the connection policy table. Before the next round of similar WRF tasks enters the connection determination stage, the reinforcement learning decision model was preferentially invoked to select actions based on this sorting result.

[0022] Working principle: First, each stage of the WRF process is encapsulated into independent but sequentially connectable functional operators. Then, edge-by-edge judgment is performed to determine whether adjacent operators have the conditions to continue passing results. Specifically, based on the JSON configuration file, input data status information, and historical execution records, the task status and candidate connection relationships of each candidate functional operator are generated. Then, for each candidate connection edge, the output status of the source functional operator is compared with the input requirements of the target functional operator to identify gaps in labels, time, configuration fields, and resources. The reinforcement learning decision model then selects from actions such as direct connection, bridging insertion, degraded execution, and connection prohibition. After the action is determined, the corresponding bridging is performed. The processing instructions complete input data chaining, time slice reassembly, configuration field rewriting, historical intermediate result isolation, and runtime directory reconstruction to form a standardized operating environment. Then, the objective function operator calls the corresponding WRF program unit to perform calculations and verifies the generation, content integrity, and semantic consistency of the result file. Finally, the completion status, semantic correspondence, and anomaly status of this round of execution are fed back to the reinforcement learning decision model to update the connection decision strategy in subsequent tasks. Thus, the output status of the previous stage directly determines the connection method of the next stage, and the actual execution result of the next stage in turn corrects the action order of subsequent tasks, thereby forming a closed-loop driving mechanism for WRF process assembly. In practical applications, taking a rolling forecast task on a high-performance computing platform as an example, the system needs to sequentially organize program units such as geogrid, ungrib, metgrid, real, and wrrf to complete the entire forecast process. However, during the operation, issues such as incomplete arrival of driving fields, delayed arrival of observation data, residual files from old tasks mixed into the current directory, or changes in resource allocation often occur. In this scenario, this solution does not simply execute the script sequentially according to a fixed script. Instead, it first determines, for example, whether the connection edge from metgrid to real truly satisfies the conditions for continuing the transmission. If it is found that the time range output by metgrid cannot cover the input time period requirements of real, or that the relevant fields are inconsistent with the current configuration, or that old intermediate data from the previous round of tasks still remain in the running directory, the solution will proceed accordingly. As a result, the system first generates corresponding bridging instructions to isolate the old results, fill in the missing data, rewrite the configuration to the state required by the current task, and recheck whether the input is closed. Only when the source, time slice, and fields are consistent are real processes allowed to continue. After real processes are completed, the system further checks whether the generated result files are complete, whether the time range and grid information are consistent with the current configuration, and writes the actual effect of this connection and execution back to the decision policy. If an exception occurs during execution, the exception is written back as negative feedback. In this way, the system can prioritize the use of historically more stable and suitable connection and bridging methods for the current data and resource conditions in subsequent similar forecast tasks, thereby improving the stability, reusability, and actual operating efficiency of the WRF process assembly.

[0023] 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 functional operator-driven method for WRF process assembly, characterized in that, include: S1. Obtain the JSON configuration file, candidate functional operator set, input data status information, configuration version information, and historical execution records associated with each candidate functional operator corresponding to the WRF task to be executed. Generate a task configuration dictionary according to the preset field parsing rules, and extract the semantic state features of each candidate functional operator based on the task configuration dictionary and historical execution records. Output the current task state set and candidate connection set. S2. Input the current task state set and candidate connection set into the pre-trained reinforcement learning decision model. The reinforcement learning decision model performs connection permission determination on the connection relationship of adjacent functional operators in each candidate connection set according to the action selection rule corresponding to the maximum cumulative reward, and outputs the connection decision result and the bridging processing instruction corresponding to the connection decision result. S3. For functional operators whose connection decision results are allowed to connect, execute input data linking, runtime directory instantiation, namelist configuration generation, historical intermediate result isolation verification or alternative data loading processing according to the bridging processing instructions, and output the standardized runtime environment and semantic closure mark corresponding to the target functional operator. S4. Pass the standardized operating environment and semantic closure mark into the target functional operator. The target functional operator calls the corresponding WRF executable program or its supporting pre- and post-processing program to perform the computation task. It also performs output integrity checks and result validity checks on the generated results and outputs the operator execution results, result semantic labels, and exception markers. S5. Construct reward feedback data based on operator execution results, result semantic labels, and anomaly markers, and update the policy parameters of the reinforcement learning decision model in combination with the connection decision results, so that the connection permission determination and bridging processing instructions of adjacent functional operators in subsequent WRF tasks are executed based on the updated policy parameters, and the updated connection decision policy is output.

2. The functional operator-driven method for WRF process assembly according to claim 1, characterized in that: S1 includes: S1-1. Input the JSON configuration file into the field mapping parsing unit, perform hierarchical splitting and key-value expansion processing on the WRF parameter field, extended execution field and path template field in the JSON configuration file according to the preset field parsing rules, generate field mapping results that are associated with each candidate functional operator, and output the task configuration dictionary. S1-2. Input the task configuration dictionary, the input data state information associated with each candidate functional operator, and the configuration version information into the semantic association construction unit. Generate the initial semantic state fragments of each candidate functional operator according to the preset correspondence between the functional operator and the input data source, configuration version fingerprint, and target output type, and output the initial semantic state set. S1-3. Input the initial semantic state set and historical execution records into the state correction unit. Perform credibility correction and connection constraint annotation on the initial semantic state fragments of each candidate functional operator according to the successful connection trajectory, abnormal propagation trajectory and intermediate result reuse trajectory in the historical execution records. Generate the semantic state features of each candidate functional operator and output the current task state set and candidate connection set.

3. The functional operator-driven method for WRF process assembly according to claim 2, characterized in that: S2 includes: S2-1. For each candidate connection edge in the candidate connection set, read the result semantic label, result data start and end time, result configuration field summary, and result resource usage record of the source functional operator, as well as the input requirement label, input time period requirement, input configuration field requirement, and execution resource requirement of the target functional operator. Then, compare them item by item according to label item, time item, field item, and resource item to obtain the label gap list, time gap list, field gap list, and resource gap list, and output the candidate connection difference table.

4. The functional operator-driven method for WRF process assembly according to claim 3, characterized in that: S2 further includes: S2-2. Input the candidate connection difference table into the reinforcement learning decision model. The reinforcement learning decision model calculates the cumulative reward value of connection permission action, bridging insertion action, degradation execution action and connection prohibition action for each candidate connection edge. The number of label gap items, time gap items, field gap items and resource gap items that can be eliminated after each action is applied to the corresponding candidate connection edge is summed to obtain the gap elimination amount. The number of additional processing steps required to execute the action and the additional resource consumption are summed to obtain the action cost amount. Output the action reward table and action cost table for each candidate connection edge. S2-3. For each candidate connection edge, first select the action with the largest cumulative return value from the corresponding action return table as the current action. Then, subtract the action cost from the gap elimination amount corresponding to the current action to obtain the net value of the current action. Write the current action into the WRF process connection sequence and recalculate the label gap list, time gap list, field gap list, and resource gap list of subsequent candidate connection edges one by one in the process order. If the total number of gap items of any subsequent candidate connection edge after recalculation is greater than the total number of gap items before writing, discard the current action and take the action with the second largest cumulative return value and repeat the execution until the target action with the largest net value of the current action is obtained without increasing the total number of gap items of subsequent candidate connection edges after writing is obtained. Output the connection decision result.

5. The functional operator-driven method for WRF process assembly according to claim 4, characterized in that: S2 further includes: S2-4. Based on the target action corresponding to each candidate connection edge in the connection decision results, read the label gap list, time gap list, field gap list and resource gap list of the candidate connection edge item by item, and convert missing input label items into input data chain completion operation, time misalignment items into time slice reorganization operation, field inconsistency items into configuration field rewriting operation, historical result residue items into intermediate result isolation operation, and resource insufficiency items into running directory reconstruction operation or alternative data loading operation. Generate bridging processing instructions according to the order of the candidate connection edges in the WRF process connection sequence, and output the bridging processing instructions.

6. The functional operator-driven method for WRF process assembly according to claim 5, characterized in that: S3 includes: S3-1. Read the target functional operator identifier, bridging processing instructions, task configuration dictionary, and input requirement list of the target functional operator from the connection decision results. For each input item in the input requirement list, determine the corresponding data source path, target link path, time slice identifier, and file naming item. Write the source path and target link path into the link mapping table and create the input data link. Concatenate the target functional operator identifier with the instance number to generate an independent running directory name and create an independent running directory. Then, write the field items corresponding to the target functional operator in the task configuration dictionary into the same position in the namelist template to obtain the target configuration file and output the candidate running environment. S3-2. Read the link mapping table, independent running directory, and target configuration file in the candidate running environment, and read the historical intermediate result list in the independent running directory. For each result item in the historical intermediate result list, extract the result generation task identifier, result source identifier, result time slice range, result configuration field item, and result semantic label. Then compare them item by item with the data source identifier, time slice identifier in the link mapping table, and the corresponding field item in the target configuration file. Remove historical intermediate results that have different result generation task identifiers, different result source identifiers, non-overlapping result time slice ranges, inconsistent result configuration field items, or result semantic labels that do not belong to the target functional operator input requirement list from the independent running directory. Write the corresponding replacement data into the independent running directory according to the bridging processing instructions, and output the corrected candidate running environment. S3-3. Read the revised candidate runtime environment and the input requirement list of the objective function operator. For each input item in the input requirement list, check whether there is a corresponding linked file or alternative data file in the independent runtime directory, whether there is a corresponding field in the target configuration file, whether the time slice identifier of the linked file or alternative data file falls within the time slice range corresponding to the input item, and whether the data source identifier of the linked file or alternative data file is consistent with the data source identifier corresponding to the input item. Write all the check results into the input closure table. When each input item in the input closure table has a corresponding file, corresponding field, consistent time slice, and consistent source, generate a standardized runtime environment and write the semantic closure mark as closed. When any input item in the input closure table is missing a corresponding file, missing a corresponding field, inconsistent time slice, or inconsistent source, write the input item into the unclosed item list and write the semantic closure mark as unclosed. Output the standardized runtime environment and semantic closure mark.

7. A functional operator-driven method for WRF process assembly according to claim 6, characterized in that: S4 includes: S4-1. Input the running directory, input file list, configuration file and semantic closure mark in the standardized running environment into the target functional operator. The target functional operator reads the program path, input directory, output directory and parallel parameters in the operator mapping table according to the operator identifier. Write each input file in the input file list into the input directory. Concatenate the program path, parallel parameters and configuration file path in sequence to generate the execution command. Generate the expected result file list according to the result file name list. Output the execution command and the expected result file list. S4-2. Input the execution command and the expected result file list into the execution unit, call the execution command to start the WRF program unit corresponding to the target function operator, after the program exits, traverse the output directory to generate the actual result file list, then search for the file with the same name in the actual result file list by each file name in the expected result file list, and search for the write record in the execution log by the path of the file with the same name, generate the file correspondence table, and output the file correspondence table and the actual result file list.

8. A functional operator-driven method for WRF process assembly according to claim 7, characterized in that: S4 further includes: S4-3. Input the file correspondence table, the list of actual result files, the configuration file, and the operator identifier into the verification unit. For each actual result file, read the time range, spatial grid identifier, variable field set, and source operator identifier, and compare them item by item with the start time, end time, grid field, output field, and operator identifier in the configuration file. When each actual result file satisfies the conditions of time range correspondence, spatial grid identifier correspondence, complete output field, and source operator identifier correspondence, generate the operator execution result and result semantic label. If any mismatch exists, generate an exception marker and output the operator execution result, result semantic label, and exception marker.

9. A functional operator-driven method for WRF process assembly according to claim 8, characterized in that: S5 includes: S5-1. Input the operator execution result, result semantic label, anomaly mark and connection decision result into the feedback construction unit. For each candidate connection edge, read the connection action identifier, bridging processing instruction, total number of result files, number of valid result files, number of abnormal result files and the total number of semantic items in the result semantic label that correspond to the input requirements of the objective function operator and the number of corresponding successful semantic items. Calculate the result completion rate, semantic correspondence rate and anomaly ratio respectively, and generate the connection feedback sequence. S5-2. Input the connection feedback sequence into the reward calculation unit, and obtain the immediate reward value for each candidate connection edge by adding the result completion rate and semantic correspondence rate and subtracting the abnormality ratio. Then, according to the connection order in the WRF process, accumulate the subsequent immediate reward values ​​from the back to the front to obtain the cumulative reward value of each candidate connection edge, and output the cumulative reward sequence.

10. A functional operator-driven method for WRF process assembly according to claim 9, characterized in that: The S5 also includes: S5-3. Input the cumulative reward sequence and connection decision results into the parameter update unit. For each candidate connection edge, read the original policy parameter values ​​of all candidate actions. First, sum the original policy parameter values ​​to obtain the total parameter value. Then, multiply the cumulative reward value corresponding to the executed connection action with the total parameter value to obtain the reward allocation value. Add the reward allocation value to the original policy parameter value corresponding to the executed connection action to obtain the new policy parameter value of the action. Then, deduct the reward allocation value from the original policy parameter values ​​of the remaining unexecuted candidate actions according to their proportion in the total parameter value to obtain the new policy parameter values ​​of the remaining actions. Output the updated policy parameter set. S5-4. Input the updated set of strategy parameters into the strategy generation unit. For each candidate connection edge, read the corresponding new strategy parameter value according to the candidate action identifier. Sort the new strategy parameter values ​​from largest to smallest to form the action sequence of the candidate connection edge. Then write the action sequences of each candidate connection edge into the connection strategy table in the order of the WRF process to generate the updated connection decision strategy.

Images