Agent-based minute-level precipitation monitoring and influence analysis method

Abstract

The present application relates to the technical field of geographic information database updating, in particular to a kind of minute-level precipitation monitoring and influence analysis method based on agent.The present application records the event minute of multi-source precipitation data, completion progress and evidence version according to source partition, identifies precipitation event and forms versioned event piece with event blood relationship and event topological difference;On specified graph version and time delay profile version, time delay constraint influence subgraph is formed according to path propagation time delay and operator time support;Evidential revision state key and topological change state key are summarized, and according to necessary source partition completion progress and evidence, graph and calculation configuration version consistency form submission, tentative or revocation influence state;Accordingly, generate graph layer difference and directional supplement request.The present application can avoid mixing different versions of minute influence layer, reduce invalid full graph recalculation and no-target supplement, realize consistent update of minute-level precipitation influence state.

Description

Technical Field

[0001] This invention relates to the field of geographic information database update technology, specifically to a method for minute-level precipitation monitoring and impact analysis based on intelligent agents. Background Technology

[0002] Minute-level precipitation monitoring is typically used for urban flooding, water system operation, road waterlogging, and impact analysis of key facilities. Existing technologies generally access rain gauge, weather radar, and forecast grid data through message interfaces, perform time alignment and spatial mapping according to fixed time windows, and then overlay precipitation grids with watershed, river segment, road, or facility layers. Some systems also combine water system topology, hydrological models, or process engines to infer downstream impacts and update map labels or monitoring reports to achieve precipitation process identification and geographic impact results display.

[0003] During minute-level operations, data from the same observation minute arrives asynchronously due to site communication, radar volume scan stitching, and quality control. Data already used in calculations may also be subsequently corrected. Existing methods typically rely on currently received data, completion markers for fixed batches, or the entire impact map to perform calculations. This makes it difficult to simultaneously determine the actual source minute data required for a specific affected object at a specific target minute, the arrival status of each source, and the version used in the calculation. When data continues to arrive or is corrected, it is easy to encounter intermediate consequences such as mixing different versions of results, failing to accurately define the objects and minute ranges that should be recalculated, and excessively large ranges for supplementing missing data. Therefore, it is necessary to solve the problems of deterministic calculation of minute-level impact results under asynchronous multi-source input conditions, stable version writing, and partial updates. Summary of the Invention

[0004] In view of the above-mentioned problems in the prior art, the present invention provides the following technical solution: This invention provides a method for minute-level precipitation monitoring and impact analysis based on intelligent agents, including: S1. Collect the event minutes, completion progress and evidence version of the multi-source precipitation records by the intelligent agent according to the source partition, and form minute precipitation evidence frames; S2. The monitoring agent identifies precipitation events and changes in evidence or event boundaries based on the minute precipitation evidence frames, forming a versioned precipitation event slice with event lineage and event topological difference. S3. The orchestration agent maps the versioned precipitation event slices to the geographic impact dependency graph on the specified graph version and time delay profile version. Based on the path propagation delay and operator time support, it forms a time delay constraint impact subgraph including the source minutes required for the target minute, the necessary source partitions, and the topology change state key. S4. The orchestration agent summarizes the evidence correction state key and the topology change state key according to the time delay constraint influence subgraph to form a state update set. The influence analysis agent marks the influence state associated with the addition, remapping or evidence correction instruction as submitted or provisional, and marks the influence state associated with the revocation instruction as revoked, according to the time support, the completion progress of the required source partition and the consistency of the evidence, graph and computing configuration versions, to form a consistent minute influence state set. S5. The publishing agent forms a layer difference based on the consistent minute impact state set, and generates a targeted supplementary sampling request based on the transaction identifier of the layer difference according to the missing evidence set in the consistent minute impact state set, thus forming the minute precipitation impact monitoring result.

[0005] Optionally, the continuous sequence number status and source partition water level are maintained for the source partition key composed of the source identifier and the partition identifier, respectively; If all records due at or before the corresponding event minute in a source partition have been received, or if an explicit missing marker with spatial location and event minute is written for the missing records, the waterline of the source partition will cross the event minute. As the numerical value, quality status, or missing status of the source record is corrected, the evidence version is incremented and the replaced version is retained. The source partitions are aggregated according to spatial units and event minutes to form the minute precipitation evidence frame.

[0006] Optionally, spatial connectivity marking is performed on spatial units that are of valid quality and reach the precipitation identification threshold, and new event components are formed based on the continuity relationship of adjacent event minutes; A two-part association relationship is constructed based on the spatial overlap between the new event component and the previous version event component. Event component matching is performed based on the two-part association relationship. When an event splits, the successor event component with the largest spatial overlap inherits the event identifier of the previous version event component. When multiple successor event components have the same maximum spatial overlap, they are sorted according to the smallest spatial unit identifier of each successor event component. The successor event component with the first sorted value inherits the event identifier of the previous version event component. When an event merges, the new event component inherits the event identifier of the previous version event component with the largest spatial overlap. When multiple previous version event components have the same maximum overlap, the new event component inherits the event identifier with the smallest lexicographical order, thus forming an event lineage record. By comparing the evidence citations, spatial units, event minute ranges, and event lineage records of the new and old event components, the event topology difference, including the set of newly added event slices, the set of revoked event slices, and the set of remapped event slices, is formed to obtain the versioned precipitation event slice.

[0007] Optionally, the propagation delay support of each side in the path from the source spatial unit covered by the event to the affected object is combined with the time support of the operators along the path by interval synthesis or discrete convolution to obtain the path time support, and a versioned path time kernel is obtained in the discrete convolution method. Based on the path time support, the source minutes required for the target minutes are derived from the target minutes, and the required source partitions are determined according to the spatial mapping relationship and operator configuration; According to the deterministic aggregation rule corresponding to the path time support or the versioned path time kernel, consume all the necessary evidence corresponding to the source minutes required for the target minute and the necessary source partition, and combine it with the topology change state key to form the delay constraint influence subgraph.

[0008] Optionally, the versioned precipitation event slices are mapped to the specified map version to form candidate map data including the current path, current source partition, current time support, and current affected object mapping; The candidate graph data is compared with the previous version of the delay constraint-affected subgraph, and the topology change state key is determined according to the mutual exclusion relationship. Among them, the key that does not have an existing affected state but has the current affected object mapping is determined as the new state key, the key that has an existing affected state but the current affected object mapping disappears is determined as the cancel state key, and the key that has an existing affected state, the current affected object mapping is retained and the event lineage, graph version, path, time support or affected object mapping changes is determined as the remapping state key. The newly added state key, the undo state key, and the remapping state key are written into the candidate graph data to form the delay constraint influence subgraph.

[0009] Optionally, an existing state key that supports the signature referencing the replaced evidence version and whose correction event minute belongs to the source minute required for the target minute will be determined as the evidence correction state key; Read the topology change state key from the time delay constraint influence subgraph, summarize and deduplicate the evidence correction state key and the read topology change state key to form the state update set; For the state update set, evidence correction, addition, revocation, and remapping instructions are retained. For keys that have both revocation and other instructions, the revocation instruction is used as the standard and a deletion tombstone is generated. For keys that do not have a revocation instruction but have an addition instruction, initial calculation is performed. For keys that do not have a revocation instruction, do not have an addition instruction, but have evidence correction or remapping instructions, recalculation is performed according to the topological order of the time delay constraint-affected subgraph. For keys that have undergone initial calculation or recalculation, a supporting signature for the actual consumption version of the record is generated. Based on the completion progress of the required source partition, the validity of the evidence, the consistency of the version corresponding to the supported signature, and the deletion of the tombstone, the consistency minute impact state set is formed.

[0010] Optionally, the fields supporting the signature include the affected object identifier, target minute, event version, actual consumption evidence version, fusion rule version, graph version, path delay profile version, time support configuration version, operator version, parameter version, and state writing benchmark read during computation. Before writing the state, the orchestration agent rereads the current values ​​of the evidence version, event version, fusion rule version, graph version, path delay profile version, time support configuration version, operator version, and parameter version and checks them item by item against the corresponding fields in the supporting signature. It also rereads the current value of the state version and checks it against the state writing benchmark. If both checks are consistent, the affected state to be written is located using the affected object identifier and the target minute, and a comparison and exchange is performed using the state writing benchmark as the expected state version. If any check is inconsistent or the comparison and exchange fails, the state is re-queued. If the current value of the graph version, path delay profile version, time support configuration version, or operator version is inconsistent with the corresponding field in the supporting signature, the delay constraint affected subgraph is regenerated.

[0011] Optionally, the impact analysis agent, based on time support, the completion progress of required source partitions, and the consistency of evidence, graph, and computation configuration versions, marks the impact states associated with add, remap, or evidence correction instructions as committed or provisional, and marks the impact states associated with revocation instructions as revoked, forming a consistent minute-by-minute impact state set, including: Select the target affected object, and when the revocation instruction is read from the time delay constraint influence subgraph, mark the influence status of the target affected object as revocation, and generate a deletion tombstone to instruct the publishing agent to delete the original layer object; If the revocation instruction is not read and the necessary evidence is incomplete, the influence status of the target object is marked as tentative, and the status value is set to empty, the valid status value is set to invalid, the status level is set to undetermined, and the candidate value is only used for interface preview. If the revocation instruction is not read, and the required evidence is complete and consistent with the version, the impact status of the target object is marked as committed; If the revocation instruction is not read, and the required evidence is complete but there is a version inconsistency, the current calculation result is not marked as submitted. Instead, depending on the type of version inconsistency, the impact status of the target object is marked as provisional and then re-queued, or the time delay constraint impact subgraph is regenerated. The impact states marked as committed, provisional, or revoked are aggregated to form the consistent minute impact state set.

[0012] Optionally, a layer difference dispatch record is generated based on the layer difference, and a targeted supplementary dispatch record is generated based on the targeted supplementary dispatch request; Establish a local database transaction, and write the layer differential dispatch record, the targeted supplementary sampling dispatch record, and the result transaction record into the local database transaction, so that the three types of records contain the same transaction number and target minute, and make the layer differential dispatch record and the targeted supplementary sampling dispatch record each contain their respective idempotency term numbers; Submit the local database transaction and output the minute precipitation impact monitoring results, which include the layer difference, the targeted supplementary sampling request, and the three types of record identifiers.

[0013] Beneficial effects The technical solution provided by this invention has the following beneficial effects: 1. By maintaining the completion progress by source partition, and performing interval synthesis or discrete convolution of path propagation delay and operator time support on specified graph versions and delay profile versions, all necessary source minutes and necessary source partitions can be determined for the target affected object and target minute; combined with deterministic aggregation rules and the distinction between submission and provisional states, the candidate values ​​corresponding to missing evidence will not be written into the formal layer, and the same evidence snapshots will form reproducible calculation inputs according to fixed time support, avoiding arbitrary selection of minutes within the time window or replacing the completion status of all partitions with the completion status of one partition.

[0014] 2. By aggregating, deduplicating, and retaining various indicators of the evidence correction status key, the system performs initial calculations on the new key, recalculates on the correction or remapping key, and generates deletion tombstones for the undo key. This allows state changes caused by late corrections, threshold crossings, and event splitting and merging to enter the same update queue. This avoids missing newly created affected states or leaving old layer objects after the mapping disappears when only existing states are retrieved, and limits numerical calculations to the actual affected state range.

[0015] 3. By using evidence, events, graphs, latency profiles, time support configurations, operators, parameter versions, and status write benchmarks that support signature binding to actual consumption, and rereading the current version and performing a comparison before swapping before writing, it is possible to prevent old tasks from overwriting new states with mixed data or computational configurations. Furthermore, by binding layer difference, targeted supplementary acquisition, and result transactions in local database transactions, and setting independent idempotent item numbers for each sending record, network retries only replay the corresponding sending item, avoiding layer updates and supplementary acquisition requests from becoming detached from each other or taking effect repeatedly. Attached Figure Description

[0016] Figure 1 A schematic diagram illustrating the steps of a minute-level precipitation monitoring and impact analysis method based on intelligent agents provided by this invention; Figure 2 This invention provides an overall flowchart of a minute-level precipitation monitoring and impact analysis method based on intelligent agents; Figure 3 This is a schematic diagram of the multi-boundary intelligent agent collaborative processing architecture provided by the present invention; Figure 4 A schematic diagram illustrating the steps for generating a state update set and forming a consistent state, provided by this invention. Figure 5 This is a schematic diagram of the timing interaction between late correction driving and state update provided by the present invention; Figure 6 This is a time-delayed propagation evolution diagram of precipitation impact along the path provided by the present invention. Detailed Implementation

[0017] 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 a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] This invention can be applied to minute-level precipitation impact monitoring in urban small watersheds, drainage zones, roads along river networks, and key facilities. The system initiates processing upon the arrival of rain gauge messages, radar quantitative precipitation estimation grid points or other hydrological rainfall records, corrections to existing records, changes in geographic impact maps, updates to time-delay profiles, or triggering of whole-minute scheduling. Participants include source interfaces providing precipitation records, computing nodes performing data acquisition and analysis tasks, and GIS layer services receiving layer differences. The final results are used to update the official or provisional layer for the target minute and, when necessary evidence is lacking, send targeted supplementary data acquisition requests to the corresponding source zone.

[0019] In this embodiment, the targeted supplementary acquisition request is a data acquisition request initiated to the corresponding source interface for the source partition, spatial location and event minute that have been determined in the missing evidence set; the data acquisition may include re-pulling, retransmission, supplementary transmission or historical record acquisition, and is used to obtain the source record that the corresponding event minute has been generated but has not yet entered the current evidence set, without requiring physical observation to be re-executed for the past event minutes.

[0020] An implementable hardware and software environment includes a minute-by-minute rain gauge interface, a radar grid interface, a message bus, a spatiotemporal database, a geographic impact map database, an operator and time-delay profile library, an impact state library, a transaction sender box, and a GIS layer interface. Data acquisition agents, monitoring agents, orchestration agents, impact analysis agents, and publishing agents run as container services or rule engine processes on intranet servers or edge-cloud collaborative computing nodes. Each agent exchanges data through structured messages and uses a service identity that only allows reading and writing to its assigned data tables and message topics. The server processor executes deterministic data parsing, graph retrieval, version checking, numerical calculations, and conditional writing programs. The memory stores program instructions, a run queue, version snapshots, and failure records. The natural language generated by the model does not participate in the judgment of numerical values ​​or submission conditions.

[0021] All interfaces use a unified whole-minute business time. The original Coordinated Universal Time (UTC) and Beijing Time conversion results are saved simultaneously; arrival time is used only for auditing. Precipitation intensity is uniformly expressed in millimeters per minute, area in square meters, and flow rate in cubic meters per second. Records that cannot be converted retain their original units and are marked as having invalid dimensions, and are not included in formal numerical calculations. Messages are deduplicated using an idempotent key composed of source identifier, partition identifier, source sequence number, and event minutes. Duplicate deliveries with unchanged content do not increment the evidence version. Agent task states include at least: pending, running, waiting for evidence, waiting for recalculation, ready to commit, committed, and retryable after failure. Each state transition saves the task identifier, input version summary, start time, end time, and failure reason, enabling the computing node to continue processing from the last persistent state after restarting.

[0022] Example 1 like Figure 1 As shown, the aforementioned method includes the following steps in sequence: S1. Collect the event minutes, completion progress and evidence version of the multi-source precipitation records by the intelligent agent according to the source partition, and form minute precipitation evidence frames; S2. The monitoring agent identifies precipitation events and changes in evidence or event boundaries based on the minute precipitation evidence frames, forming a versioned precipitation event slice with event lineage and event topological difference. S3. The orchestration agent maps the versioned precipitation event slices to the geographic impact dependency graph on the specified graph version and time delay profile version. Based on the path propagation delay and operator time support, it forms a time delay constraint impact subgraph including the source minutes required for the target minute, the necessary source partitions, and the topology change state key. S4. The orchestration agent summarizes the evidence correction state key and the topology change state key according to the time delay constraint influence subgraph to form a state update set. The influence analysis agent marks the influence state associated with the addition, remapping or evidence correction instruction as submitted or provisional, and marks the influence state associated with the revocation instruction as revoked, according to the time support, the completion progress of the required source partition and the consistency of the evidence, graph and computing configuration versions, to form a consistent minute influence state set. S5. The publishing agent forms a layer difference based on the consistent minute impact state set, and generates a targeted supplementary sampling request based on the transaction identifier of the layer difference according to the missing evidence set in the consistent minute impact state set, thus forming the minute precipitation impact monitoring result.

[0023] like Figure 2 As shown, the aforementioned five steps sequentially transmit minute-level precipitation evidence frames, versioned precipitation event slices, time-delay constraint impact subgraphs, and consistent minute-level impact state sets. The publishing end checks for missing evidence sets in the state; if the set is not empty, the source partition and event minute range are written into the targeted supplementary collection item and fed back to the collection interface; if the set is empty, no supplementary collection item is created. This object chain ensures that the identity, completion progress, and version of the source record are preserved up to the layer difference, rather than being compressed into a single precipitation value with an untraceable source in intermediate steps.

[0024] like Figure 3 As shown, the boundary between precipitation sources and the acquisition interface includes the minute-by-minute message interface of rain gauges, the grid interface for radar quantitative precipitation estimation, and the acquisition message interface. Five types of intelligent agents are deployed within the boundary of the intranet server or edge-cloud collaborative computing node, and configured with a spatiotemporal database, a geographic impact map database, an operator and time-delay profile library, an impact state library, and a transaction sending box. Multi-source precipitation records enter the acquisition intelligent agent unidirectionally from the source interface; minute evidence, event slices, impact sub-maps, and impact states are sequentially transmitted along the message topic. The publishing intelligent agent sends layer differences to the GIS layer service boundary via the transaction sending box and sends targeted supplementary sampling requests back to the acquisition message interface. The data collection agent has write permissions to the spatiotemporal database, while subsequent agents only have read permissions to the corresponding evidence tables. The orchestration agent has version read permissions to the geographic impact map database, operator and time-delay profile database, and impact state database, and has write permissions to impact state conditions after version verification. The impact analysis agent generates candidate impact states, calculation results, and supporting signatures, and returns these to the orchestration agent, without directly writing formal impact states to the impact state database. The publishing agent has write permissions to the transaction mailbox. No service identity has database permissions exceeding the aforementioned scope.

[0025] For S1, specifically, the data acquisition agent reads multi-source precipitation records from the rain gauge message interface and the radar grid interface. Each record contains at least the source identifier, partition identifier, spatial unit identifier, event time, arrival time, source sequence number, precipitation value, unit, quality status, and correction flag. The source identifier and partition identifier are concatenated in a fixed field order to form the source partition key, and the spatial unit identifier corresponds to a unified geographic grid. The data acquisition agent first verifies the required fields and units, then rounds the event time down to the nearest minute. Records that pass verification are written to the spatiotemporal database according to the spatial unit identifier and event minutes. If the units are inconsistent but a conversion factor exists, they are first converted to millimeters per minute. If a field is missing or the unit cannot be converted, an invalid reason is written, and the record's identity is retained.

[0026] The partition range of a source interface is the same as the range to which it can independently declare a closure. Rain gauges can partition by station, and radars can partition by tile, sector, or single mosaic task; the sequence number and closure flag of one partition do not replace other partitions. The data acquisition agent maintains a set of received source sequence numbers, the next expected sequence number, the partition water level, and a list of unreached locations for each source partition key. When a duplicate idempotent key is received, the precipitation value, quality status, and missing status are compared field by field; if all are the same, the received status is returned; if any field changes, the original record is marked as replaced and a new evidence version is created.

[0027] In some implementations, the data acquisition agent records the event minutes, completion progress, and evidence versions of multi-source precipitation records by source partition, forming minute-by-minute precipitation evidence frames, including: Maintain the sequence number continuity and the source partition water level for the source partition key composed of the source identifier and the partition identifier, respectively; If all records due at or before the corresponding event minute in a source partition have been received, or if an explicit missing marker with spatial location and event minute is written for the missing records, the waterline of the source partition will cross the event minute. As the numerical value, quality status, or missing status of the source record is corrected, the evidence version is incremented and the replaced version is retained. The source partitions are aggregated according to spatial units and event minutes to form the minute precipitation evidence frame.

[0028] Specifically, the data acquisition agent scans records within the same source partition in ascending order of source sequence number. If there are no gaps starting from the next expected sequence number, the event minute corresponding to the end of the continuous interval is written as a candidate waterline. If there are gaps in the sequence number, but the source interface has issued a closure flag for that minute, the missing list is checked item by item, and an explicit missing flag containing the source partition key, spatial unit identifier, event minute, reason for absence, and evidence version is written for each missing item. Only after this is completed is the waterline allowed to cross that minute. If the closure flag has not arrived or the missing item cannot be located, the waterline remains at its original value, and the task status is set to "waiting for evidence." An explicit missing flag only indicates a known gap at that location; it does not set the evidence completeness status to valid.

[0029] When precipitation values, quality status, or missing status change, the data acquisition agent increments the version integer by one under the same evidence identifier, writes the old version number into the replaced version field, and retains the old record for subsequent signature retrieval support. Subsequently, records are aggregated in ascending order by spatial unit identifier, event minute, source partition key, and evidence version to generate minute-level precipitation evidence frames. Each evidence frame at least stores the frame identifier, analysis minute, spatial unit identifier, each source partition record, evidence identifier and version, replaced version, quality status, explicit missing set, water level vector, spatial mapping rule version, fusion rule version, and quality rule version, and is sent to the monitoring agent via a minute-level evidence message topic. No message is published if database write fails; if message publication fails, the written frame and the pending delivery flag are retained, and the same version is resent with the frame identifier upon restart.

[0030] For S2, specifically, the monitoring agent uses high-quality valid records from the minute-based precipitation evidence frames as input for connectivity analysis, while retaining the evidence identities of high-quality invalid records and those with explicit missing markers. The precipitation identification threshold, four-neighbor or eight-neighbor adjacency mode, and allowed adjacent minute discontinuities are all read from the versioned event matching configuration; precipitation values ​​equal to the threshold are assigned to branches reaching the threshold, while those below the threshold are not included in the current connected component. The monitoring agent traverses valid units in ascending order of event minute and spatial unit identifier, writes adjacent units that have reached the threshold into the same new event component, and connects components with adjacent minutes not exceeding the registered discontinuity length into a continuous precipitation process.

[0031] In some implementations, the monitoring agent identifies precipitation events and changes in evidence or event boundaries based on the minute-by-minute precipitation evidence frames, forming a versioned precipitation event slice with event lineage and event topological differences, including: Spatial connectivity is marked for spatial units that are of valid quality and reach the precipitation identification threshold, and new event components are formed based on the continuity of adjacent event minutes; A two-part association relationship is constructed based on the spatial overlap between the new event component and the previous version event component. Event component matching is performed based on the two-part association relationship. When an event splits, the successor event component with the largest spatial overlap inherits the event identifier of the previous version event component. When multiple successor event components have the same maximum spatial overlap, they are sorted according to the smallest spatial unit identifier of each successor event component. The successor event component with the first sorted value inherits the event identifier of the previous version event component. When an event merges, the new event component inherits the event identifier of the previous version event component with the largest spatial overlap. When multiple previous version event components have the same maximum spatial overlap, the new event component inherits the event identifier with the smallest lexicographical order, thus forming an event lineage record. By comparing the evidence citations, spatial units, event minute ranges, and event lineage records of the new and old event components, the event topology difference, including the set of newly added event slices, the set of revoked event slices, and the set of remapped event slices, is formed to obtain the versioned precipitation event slice.

[0032] Specifically, the monitoring agent establishes a spatial unit set for each new event component and calculates its intersection with each active event component of the previous version, using the number of spatial units in the intersection as the spatial overlap. Candidate association edges with zero overlap and not meeting the adjacent minute condition are first removed. Based on the retained candidate association edges, a bipartite association relationship is constructed between the previous version event component and the new event component. Then, the candidate association edges are processed in descending order of spatial overlap to determine the event identifier inheritance relationship and event lineage. One-to-one matching uses the original event identifier; when one predecessor corresponds to multiple successors, the successor with the largest overlap inherits the original identifier. If the maximum values ​​are tied, the smallest spatial unit identifier of each successor is compared and the one with the highest overlap is selected. Other successors acquire new event identifiers and record the same predecessor; when multiple predecessors correspond to one successor, the predecessor with the largest overlap is selected first. If the maximum values ​​are tied, the one with the smallest lexicographical order of the event identifier is selected, and all predecessor identifiers are written into the predecessor set. New components without candidate predecessors acquire new event identifiers, and old components without candidate successors are written to the termination state.

[0033] After matching is completed, the monitoring agent compares the evidence reference set, spatial unit set, event minute start and end values, and predecessor and successor lineage edges of the old and new components item by item. Fragments that do not exist in the old version but exist in the new version are written into the newly added event fragment set; fragments that exist in the old version but do not exist in the new version are written into the revoked event fragment set; fragments that exist in both but have any changed comparison field are written into the remapped event fragment set; identical fragments are not written into the difference set. The output versioned precipitation event fragments contain at least the event identifier, event fragment version, predecessor set, successor set, spatial unit set, event minute range, precipitation sequence, evidence reference set, partition water level vector, missing minute set, three types of event fragment differences, and matching rule version, and are sent to the orchestration agent. When the matching task times out or the configuration version changes during computation, the old event fragments are not overwritten; instead, the failed version summary is saved and re-queued.

[0034] For S3, specifically, the orchestration agent reads versioned precipitation event slices, target minutes, geographic impact dependency graphs, and operator registry entries. Graph nodes include precipitation grids, catchment units, drainage river sections, roads, low-lying areas, or key facilities. Graph edges represent spatial intersections, water flow convergence, river propagation, drainage connectivity, or registered impact transmission relationships. Each edge stores the non-negative propagation delay interval, delay profile identifier, delay profile version, and applicable operating conditions. Each node operator stores the operator identifier, operator version, parameter version, required source partitioning rules, fusion rules, time support configuration, and deterministic aggregator. Local networks containing loops are shrunk into composite nodes before going live. The output and time support are provided by the registered solver within the composite node, and the external graph remains directed acyclic. If loops are not shrunk or internal solvers are missing, the corresponding paths are marked as invalid and prevented from forming formal states.

[0035] In this embodiment, the path delay profile version is a version vector composed of the delay profile versions corresponding to each edge of the target propagation path in the edge order of the target propagation path; the time support configuration version is a version vector composed of the time support configuration versions corresponding to each node operator participating in the time support calculation on the target propagation path in topological order. When the target propagation path corresponds to only one delay profile version or one time support configuration version, the corresponding version vector is composed of a single version identifier.

[0036] In some implementations, mapping the versioned precipitation event slices to a geographic impact dependency graph on a specified map version and time-delay profile version, and forming a time-delay constrained impact subgraph including the source minutes required for the target minute, the necessary source partitions, and the topology change state key based on path propagation delay and operator time support, includes: The propagation delay support of each side in the path from the source spatial unit covered by the event to the affected object is combined with the time support of the operators along the path by interval synthesis or discrete convolution to obtain the path time support, and the versioned path time kernel is obtained in the discrete convolution method. Based on the path time support, the source minutes required for the target minutes are derived from the target minutes, and the required source partitions are determined according to the spatial mapping relationship and operator configuration; According to the deterministic aggregation rule corresponding to the path time support or the versioned path time kernel, consume all the necessary evidence corresponding to the source minutes required for the target minute and the necessary source partition, and combine it with the topology change state key to form the delay constraint influence subgraph.

[0037] Specifically, at the start of the task, the orchestration agent reads and fixes the graph version, path delay profile version, time support configuration version, fusion rule version, operator version, and parameter version. For each source spatial unit covered by an event, it traverses the paths that can reach the target object in the topological order of the graph. In the interval method, it sequentially adds the lower bound of the delay of each edge and the lower bound of the operator memory on the path, and sequentially adds the upper bound of the delay and the upper bound of the memory, which are used as the lower and upper bounds of the path time support, respectively. In the discrete method, it performs convolution on the edge delay weight sequence and the node memory weight sequence in the path order, and then divides each discrete weight by the sum of the weights. If the sum is zero, negative weights are present, or the upper and lower bounds of the delay are reversed, the path is added to the invalid path set and is not included in the formal calculation.

[0038] The orchestration agent offsets the target minute forward by the upper and lower bounds supported by the path time, obtaining the start and end minutes of the closed interval. Integer minutes with weights greater than zero in the discrete kernel are written into the source minute set required for the target minute in ascending time order. If this set is earlier than the start minute of the event slice, zero-value or non-zero-value evidence is still read from the spatiotemporal database according to spatial unit, source partition, and event minute; zero-value records cannot be replaced by the fact that the minute did not enter the event slice. The orchestration agent then searches for candidate partitions item by item according to the mapping table from source spatial unit to source interface partition, and filters the required source partitions according to the required flags registered by the node operators. Optional partitions only participate in the augmentation calculations allowed by the registration fusion rules and do not participate in the submission integrity judgment.

[0039] For multi-source evidence of the same spatial unit and event minutes, invalid values ​​are first removed according to the quality validity indicator. Then, a fusion value is generated according to the source priority, fixed weight, or deterministic model version registered in the fusion rule version. Only then are all required source minutes weighted according to the path time kernel, or the registered aggregator bound to the interval time support is called to consume the complete minute vector. When there is no valid required evidence, the missing position is retained and not replaced with a zero value. When there are multiple valid sources and the fusion weight sum is not one, it is normalized according to the weight sum. The resulting set of source minutes, required source partitions, fusion input references, and time aggregation input vectors are written into the candidate graph data for subsequent state calculation and correction filtering.

[0040] Further, the step of mapping the versioned precipitation event slices to a geographic impact dependency graph on a specified graph version and time-delay profile version, and forming a time-delay constraint impact subgraph including the source minutes required for the target minute, the necessary source partitions, and the topology change state key based on path propagation delay and operator time support, includes: The versioned precipitation event slices are mapped to the specified map version to form candidate map data including the current path, current source partition, current time support, and current affected object mapping; The candidate graph data is compared with the previous version of the delay constraint-affected subgraph, and the topology change state key is determined according to the mutual exclusion relationship. Among them, the key that does not have an existing affected state but has the current affected object mapping is determined as the new state key, the key that has an existing affected state but the current affected object mapping disappears is determined as the cancel state key, and the key that has an existing affected state, the current affected object mapping is retained and the event lineage, graph version, path, time support or affected object mapping changes is determined as the remapping state key. The newly added state key, the undo state key, and the remapping state key are written into the candidate graph data to form the time delay constraint influence subgraph.

[0041] Specifically, the orchestration agent uses key-by-key comparison units composed of the affected object identifier, target minute, and path identifier. It first records the current path, current source partition set, current source minute set, current time support, current affected object mapping, and event lineage summary in the candidate graph data. Then, it reads the existing state and previous version fields from the affected state library and old subgraph index within the same comparison unit. If the existing state does not exist but the current mapping does, an addition indicator is written; if the existing state exists but the current mapping does not, a cancellation indicator is written; if both exist and at least one of the event lineage summary, graph version, path edge sequence, source minute set, time support, or affected object mapping is different, a remapping indicator is written; if neither exists or all comparison fields are the same, no topology change indicator is written. The judgment order first compares the existence of the existing state and the current mapping, then compares the differences in the fields of the retained mapping. Therefore, addition, cancellation, and remapping are mutually exclusive within the same comparison unit.

[0042] The orchestration agent writes the three sets of state keys, along with the node set, edge set, topological order, event identifier and version, graph version, path identifier, target minute, required source minute, required source partition, fusion rule version, time kernel or registration aggregator identifier, delay profile version, time support configuration version, operator version, and parameter version, into the delay constraint influence subgraph and sends it to the influence analysis agent. When a graph read fails, the old subgraph is retained and the graph version is recorded as unreadable. When the graph version changes during construction, unpublished candidate data is discarded, and the path is retraced with the new version to avoid writing nodes or edges from two different graph versions into the same subgraph.

[0043] In this embodiment, the status key used for status update filtering is a path-level status key. This path-level status key consists of an affected object identifier, a target minute, and a path identifier. It is used to identify path-level comparison units in the delay constraint influence subgraph and to form new status keys, revoked status keys, remapping status keys, and evidence correction status keys. The final affected status written into the affected status database is an object-minute-level status. The affected object identifier and target minute are used to locate the corresponding affected status record. The path-level status key is used to determine the initial calculation, recalculation, or revocation range involved in the object-minute-level status. The status writing benchmark is used to control the conditional writing of the object-minute-level status.

[0044] For S4, specifically, as Figure 4 As shown, the orchestration agent first reads the topology change state keys from the time-delay constraint influence subgraph and retrieves existing states from the influence state library that support the signature referencing the replaced evidence version. The two types of keys are sorted and summarized by the influence object identifier, target minute, and path identifier. When the same key appears repeatedly, the source indication is merged, but only one update task is retained. The orchestration agent first checks the revocation indication; tasks with revocation indications directly enter the tombstone deletion branch without calling the influence numerical operator. The remaining tasks are then distinguished between new initial calculations and evidence corrections or remapping recalculations, and are handed over to the influence analysis agent according to the subgraph topology order.

[0045] In some implementations, the orchestration agent, based on the delay constraint influence subgraph, aggregates the evidence correction state keys and the topology change state keys to form a state update set. The influence analysis agent, based on time support, the completion progress of required source partitions, and the consistency of evidence, graph, and computation configuration versions, marks the influence states associated with addition, remapping, or evidence correction instructions as committed or provisional, and marks the influence states associated with revocation instructions as revoked, forming a consistent minute-by-minute influence state set, including: The existing state key that supports the signature referencing the replaced evidence version and whose correction event minute belongs to the source minute required for the target minute is determined as the evidence correction state key; Read the topology change state key from the time delay constraint influence subgraph, summarize and deduplicate the evidence correction state key and the read topology change state key to form the state update set; For the state update set, evidence correction, addition, revocation, and remapping instructions are retained. For keys that have both revocation and other instructions, the revocation instruction is used as the standard and a deletion tombstone is generated. For keys that do not have a revocation instruction but have an addition instruction, initial calculation is performed. For keys that do not have a revocation instruction, do not have an addition instruction, but have evidence correction or remapping instructions, recalculation is performed according to the topological order of the time delay constraint-affected subgraph. For keys that have undergone initial calculation or recalculation, a supporting signature for the actual consumption version of the record is generated. Based on the completion progress of the required source partition, the validity of the evidence, the consistency of the version corresponding to the supported signature, and the deletion of the tombstone, the consistency minute impact state set is formed.

[0046] Specifically, the orchestration agent reads the replaced version and correction event minutes of each correction record, then uses the evidence version inverted index to find the supporting signature referencing the replaced version; only the state key whose correction event minute is within the corresponding required source minute set is retained and written as the evidence correction state key. This set is combined with the addition, revocation, and remapping state keys in the delay constraint influence subgraph, and deduplication is performed by comparison unit. When a revocation instruction coexists with other instructions, only the revocation action is retained, while other sources are retained in the task record for auditing; when there is no revocation instruction but there is an addition instruction, the calculation starts from the initial state of the target node; when there is no revocation instruction, no addition instruction, but there is evidence correction or remapping instruction, the unaffected upstream stable state is read and the affected nodes are recalculated in topological order. When an operator fails, the failed node, input version, and error code are written, the old formal state is not updated, and the task is marked as retrievable.

[0047] The impact analysis agent reads the latest minute of the required source minutes for each non-reversal path and compares the current water level for each source partition. If the water level is earlier than the latest minute, the corresponding minute range is written into the missing evidence set; if the water level is equal to or later than the latest minute, closure is only considered to have been reached. Subsequently, it checks the explicit missing markers, quality valid markers, and spatial mapping results for all required minutes. If explicit missing, quality invalid, or unmappable conditions exist, they are written into the missing set, quality invalid set, or invalid path set, respectively. Only when all required partitions have reached closure, all required evidence exists, and quality is valid, are the node operators called according to the fusion input order and time aggregation input vector saved in S3. Each operator outputs the node value, node level, completion status, and a list of actual consumed evidence. Downstream nodes only read the upstream outputs that have been successfully completed within the same task.

[0048] Optionally, the fields supporting the signature include the affected object identifier, target minute, event version, actual consumption evidence version, fusion rule version, graph version, path delay profile version, time support configuration version, operator version, parameter version, and state writing benchmark read during computation. Before writing the state, the orchestration agent rereads the current values ​​of the evidence version, event version, fusion rule version, graph version, path delay profile version, time support configuration version, operator version, and parameter version and checks them item by item against the corresponding fields in the supporting signature. It also rereads the current value of the state version and checks it against the state writing benchmark. If both checks are consistent, the affected state to be written is located using the affected object identifier and the target minute, and a comparison and exchange is performed using the state writing benchmark as the expected state version. If any check is inconsistent or the comparison and exchange fails, the state is re-queued. If the current value of the graph version, path delay profile version, time support configuration version, or operator version is inconsistent with the corresponding field in the supporting signature, the delay constraint affected subgraph is regenerated.

[0049] Specifically, when the impact analysis agent reads an old state, it saves the state version as the state writing baseline; if the initial state does not exist, it uses an empty version identifier registered in the database. After calculation, the agent assembles a support signature in the fixed field order of the affected object identifier, target minute, event identifier and event slice version, actual consumption evidence identifier and version, fusion rule version, graph version, path identifier, path delay profile version, time support configuration version, operator version, parameter version, and state writing baseline. The actual consumption evidence is arranged in ascending order by source partition key, spatial unit identifier, and event minute, so that the same input produces the same version tuple; the signature can be saved as the version tuple, or a cryptographic hash can be calculated on the tuple while retaining auditable details.

[0050] Before writing, the orchestration agent rereads the current value from the spatiotemporal database, event index, rule base, graph database, latency profile database, operator database, and impact state database, and checks each field with the same name. When all configuration fields are the same and the current state version equals the state write baseline, the impact state database performs an update with the affected object identifier, target minute, and state write baseline conditions; the exchange is successful when the number of affected rows is one, and it indicates that the concurrent state has changed when the number of affected rows is zero. When there is a conflict between evidence, event, fusion rule, or parameter versions, the current calculated value is discarded and the conflicting fields are recorded; when the current state version is still equal to the state write baseline, a null value is written to the provisional state under the same conditions and then re-queued; when the current state version has changed, the new concurrent state is not overwritten and the queue is directly re-queued. When there is a conflict between the graph version, path latency profile version, time support configuration version, or operator version, the current subgraph is discarded and the new snapshot is read from S3. When the database connection is interrupted, the task and support signature are kept in a pending state, and unconfirmed writes are not considered successful; after the connection is restored, the current state version is read first, and then a decision is made on whether to retry the condition update or recalculate.

[0051] Furthermore, the impact analysis agent, based on time support, the completion progress of required source partitions, and the consistency of evidence, graph, and computation configuration versions, marks the impact states associated with add, remap, or evidence correction instructions as committed or provisional, and marks the impact states associated with revocation instructions as revoked, forming a consistent minute-by-minute impact state set, including: Select the target affected object, and when the revocation instruction is read from the time delay constraint influence subgraph, mark the influence status of the target affected object as revocation, and generate a deletion tombstone to instruct the publishing agent to delete the original layer object; If the revocation instruction is not read and the necessary evidence is incomplete, the influence status of the target object is marked as tentative, and the status value is set to empty, the valid status value is set to invalid, the status level is set to undetermined, and the candidate value is only used for interface preview. If the revocation instruction is not read, and the required evidence is complete and consistent with the version, the impact status of the target object is marked as committed; If the revocation instruction is not read, and the required evidence is complete but there is a version inconsistency, the current calculation result is not marked as submitted. Instead, depending on the type of version inconsistency, the impact status of the target object is marked as provisional and then re-queued, or the time delay constraint impact subgraph is regenerated. The impact states marked as committed, provisional, or withdrawn are aggregated to form the consistent minute impact state set.

[0052] Specifically, in the undo branch, the new graph version, event slice version, and current state version are rechecked. When these three are equal to the undo task snapshot and the state write baseline, a deletion tombstone with a committed, undo lifecycle, and empty state value is written. If the check fails, the old official state is retained and version conflicts are recorded to prevent erroneous deletion. In the non-undo branch, if the required partition watermark, required evidence existence, quality status, and all supporting signature fields pass the check, the operator result is written to the state value, the valid status value is set to valid, and the commit status is written to committed. If the required evidence is incomplete, the official state value is set to empty, the valid status value is set to invalid, the state level is written to undetermined, and the commit status is written to tentative. If there is a conflict between evidence, event, fusion rule, or parameter versions and the current state version is still equal to the state write baseline, a tentative state is written according to the same empty value structure, and the conflict field is recorded. If the current state version also has a conflict, no write is performed, the concurrent new state is retained, and the state is requeued. When there is a conflict between the map version, path delay profile version, time support configuration version, or operator version, a submission status based on the current delay constraint-affected submap is not generated. Instead, the current delay constraint-affected submap is discarded according to the aforementioned version verification, and the process returns to S3 to regenerate the aforementioned delay constraint-affected submap. Candidate values ​​are stored in the preview field. When the GIS interface reads the provisional status, only a gray or dashed line style is used, and this field is not allowed to trigger formal levels, alarm messages, or downstream controls.

[0053] The consistency-minute impact state set is aggregated in ascending order by the affected object identifier and target minute. Each state contains at least the state value, valid flag, candidate value, state level, commit state, lifecycle state, supported signature, state write baseline, source partition version vector, graph version, latency profile version vector, time support configuration version vector, operator version, missing evidence set, quality invalid set, version conflict set, state version, and previous version identifier. A definitive state version is returned after writing to a commit state, provisional state, or revoking a tombstone. When consecutive conditional write failures reach the registered retries, the state is moved to a manually visible failure queue, but the old official state and the current input version summary are still retained.

[0054] like Figure 5As shown, late-arrival corrections are performed across agents in sequential order. First, the impact analysis agent returns a provisional state, calculation results, and supporting signature to the orchestration agent when necessary evidence is missing. Second, the orchestration agent verifies the version and writes it into the provisional impact state. Third, the publishing agent reads the missing evidence set and sends a targeted supplementary collection request containing the source partition and event minute range to the acquisition agent. Fourth, the acquisition agent receives the late-arrival evidence, changes the explicit missing state to valid evidence, and increments the evidence version. Fifth, the monitoring agent updates the event lineage, event topology difference, and event slice version and sends it to the orchestration agent. Sixth, the orchestration agent updates the delay-constrained impact subgraph, filters the state update set, and sends it to the impact analysis agent. Seventh, the impact analysis agent recalculates the support at the same time and returns the supporting signature and calculation results composed of the actual consumed versions. Eighth, the orchestration agent rereads each version, verifies each item, performs a comparison, and then exchanges the results. Ninth, the publishing agent receives the new state set and updates the provisional layer to the official layer. If any response times out, only the same task identifier is resent; version verification is not skipped.

[0055] For S5, specifically, the publishing agent reads the consistent minute-level impact state set and the previous layer version, comparing them item by item according to the impact object identifier and target minute. For committed states with active lifecycles, formal new or updated items are generated; for provisional states, provisional style items without formal values ​​or levels are generated; and for committed tombstones that have been revoked, deletion items are generated. Only missing evidence sets are converted into re-collection items; invalid quality, invalid graph paths, version conflicts, or operator failures are respectively entered into the quality control, subgraph reconstruction, or calculation retry queues, and no targetless re-collection requests are sent to meteorological sources.

[0056] In some implementations, the publishing agent forms a layer difference based on the consistent minute-level impact state set, and generates a targeted supplementary sampling request using the transaction identifier of the layer difference based on the set of missing evidence in the consistent minute-level impact state set, thus forming a minute-level precipitation impact monitoring result, including: Generate layer difference dispatch records based on the layer difference, and generate targeted supplementary dispatch records based on the targeted supplementary dispatch request; Establish a local database transaction, and write the layer differential dispatch record, the targeted supplementary sampling dispatch record, and the result transaction record into the local database transaction, so that the three types of records contain the same transaction number and target minute, and make the layer differential dispatch record and the targeted supplementary sampling dispatch record each contain their respective idempotency term numbers; Submit the local database transaction and output the minute precipitation impact monitoring results, which include the layer difference, the targeted supplementary sampling request, and the three types of record identifiers.

[0057] Specifically, the publishing agent encodes each formal addition, formal update, provisional style, or deletion action as a layer difference item, and merges each set of missing evidence by adjacent event minutes into source partitions, start minutes, end minutes, and priority-defined supplementary collection items. The publishing agent requests a transaction number, reads the current layer version at the target minute as the desired version, and then writes layer difference dispatch records and targeted supplementary collection dispatch records by item in the same local database transaction. Each record is assigned an idempotent item number, and a result transaction record storing the transaction number, target minute, desired layer version, and the attribution relationship of all idempotent item numbers is written. If the desired layer version condition update fails, the local transaction is rolled back, the layer state is reread, and a new difference is generated; when all writes are successful, the transaction is committed, and all three types of records are simultaneously visible.

[0058] After a transaction is committed, the layer sender sends the data to the GIS interface according to the transaction number and idempotency number of the layer sending record, while the supplementary collection sender sends the data to the collection message interface according to the same type of identifier in the supplementary collection sending record. The receiver first checks whether the binary identifier has been confirmed. If confirmed, the original confirmation result is returned without being repeated; if not confirmed, the corresponding differential or supplementary collection action is executed and a confirmation status is written. If a single interface fails, only the original idempotency is resent; a new status version is not generated for retry. After the transaction sending boxes of both receiving directions write back confirmation, the delivery status of the resulting transaction is updated to "delivered." If the server crashes after a local transaction is committed, it restarts, scans unconfirmed sending records, and continues delivery. If a local transaction is not committed, there is no visible sending record. The minute precipitation impact monitoring results include the target minute, layer version, transaction number, set of updated objects, set of deleted objects, commit status summary, supported signature references, set of supplementary collection requests, three types of record identifiers, confirmation status, and delivery status.

[0059] From an overall technical perspective, when multiple records arrive asynchronously, the completion status of a single data source or a single partition cannot indicate that the necessary evidence for the target object is complete. This method first fixes the completion progress to a source partition that can be independently closed, and then uses path delay and operator time to reversely limit the actual source minutes and source partitions required for the target minute, thus converging the integrity judgment from a full-source, full-time-window state to a set of causal evidence at the object minute level. Explicit missing data can advance to closure, but still invalidates the evidence completeness marker. Therefore, the system can generate provisional states in a timely manner without writing candidate values ​​corresponding to gaps into the formal layer.

[0060] When a record or correction changes a value, threshold is crossed, or event splits or merges, existing supporting signatures can only find the existing state referencing the old evidence, and cannot cover mappings from nothing to something or from something to nothing. This method merges the set of evidence corrections with the set of additions, reversals, and remappings generated by comparing the candidate graph, and prioritizes reversals, initializes the calculation of addition keys, and recalculates correction or remapping keys in topological order, thereby incorporating numerical changes and lifecycle changes into the same local update queue. The resulting supporting signature is bound to the actual consumed evidence and configuration snapshot, and pre-write rereading and conditional swapping prevent old tasks from overwriting new states.

[0061] Layer subtraction, targeted data acquisition, and result transactions share transaction numbers within the source-side local transaction, while each delivery item has an independent idempotent item number. This allows the corresponding item to be replayed even if the layer interface and the acquisition interface fail separately. As a result, minute-level output, subsequent evidence acquisition, and formal layer updates remain traceable, while avoiding mistaking cross-system communication for instantaneous atomic commits.

[0062] like Figure 6 As shown, this embodiment uses a propagation path consisting of a precipitation grid G1, a catchment unit C1, a drainage river section R1, and a low-lying road O1. The registered propagation delay intervals from G1 to C1, C1 to R1, and R1 to O1 are 0 to 1 minute, 4 to 6 minutes, and 2 to 3 minutes, respectively, with a cumulative path support of 6 to 10 minutes. Therefore, the target minute of O1, 14:40, corresponds inversely to 14:30 to 14:34 of G1. In the discrete convolution mode, the normalized path time kernel weights of the five source minutes are 0.05, 0.15, 0.25, 0.30, and 0.25, respectively, with a total weight of 1. The light and dark colors in the figure represent the kernel weights, the diagonal lines represent the cumulative time support along the path, and the thick box marks the target position of O1 at 14:40. This view shows that the target minute only consumes the five source minutes and does not temporarily select a single minute within any time window.

[0063] In a set of recalculated operational data, the uniform grid side length is 500 meters, the C1 area is 2,000,000 square meters, and the runoff coefficient is 0.55; the map version is gv31, the time-delay profile version is dp9, the time support configuration version is tp5, the fusion rule version is fr4, the operator version is ov5, and the parameter version is pv3. The complete fused precipitation of G1 from 14:30 to 14:34 is 0, 1.0, 1.2, 1.5, and 1.3 mm / min, respectively. Multiplying and summing these values ​​according to the aforementioned five weights yields 1.225 mm / min. Multiplying this value by the area of ​​C1 and the runoff coefficient, and then dividing by the conversion factor of 60,000 (millimeters per minute to cubic meters per second), we get 22.4583 cubic meters per second. Multiplying this by the registration transfer factor of R1 to O1 (0.82) and adding the local supply of 0.4 cubic meters per second, we get 18.8158 cubic meters per second, which is higher than the registration concern threshold of 18 cubic meters per second.

[0064] When only an explicit missing marker exists in the 14:34 record of radar partition tile-03, the water level of rain gauge S17 partition is 14:40, and the water level of the radar partition is 14:34. Both exceed their respective latest required minutes, but the radar evidence completeness status is invalid. Therefore, the impact status is written as provisional, the status value is empty, the valid marker is invalid, and the level is undetermined. The missing evidence set precisely points to 14:34 in tile-03. After the actual radar record arrives, the evidence version increments from v207 to v208; the monitoring agent maintains the event identifier E7 and updates the event slice version, the orchestration agent filters out the old supporting signature referencing v207 and the O1 state where 14:34 belongs to the required source minute set, and the impact analysis agent recalculates with the complete vector. After all versions and states are written to the baseline and verified to be consistent, the conditional exchange updates the O1 state to committed, and the publishing agent updates the provisional style to the formal attention style.

[0065] Example 2 In another implementation, path-time support uses interval synthesis without storing discrete weights. The orchestration agent still accumulates the lower and upper bounds of the time support for each edge and node according to the topological order of the graph, reverses to obtain the complete source minute interval of the target minute, and reads the deterministic aggregator bound to that interval and applicable conditions from the operator registry. For example, the aggregator can be registered to take the maximum value of all valid fused values ​​within the interval, or to send the complete vector arranged in ascending time order into the registered version of the hydrological model; the aggregator identifier, model version, input minute set, and actual evidence version are all entered into the support signature. When there are missing values ​​within the interval, the minute is not deleted, nor is it replaced by an adjacent minute, but a provisional state and corresponding supplementary entries are generated. Thus, although the interval method and the discrete convolution method use different time support representations, they both maintain the common mechanism of reverse-engineering all necessary evidence from the target minute, performing deterministic aggregation according to a fixed version, and checking the version before writing.

[0066] The completion progress of the source partition can also be represented by a continuous minute cursor, a continuous source sequence upper bound, or a partition closure batch, as long as each representation is limited to a source partition that can be closed independently and can assign the missing records to spatial locations and event minutes. Signature support can store sorted version tuples or their password hashes; comparison and exchange can be implemented using database condition updates, optimistic locking, or transaction version columns. Agents can be hosted by container services, rule engines, or software frameworks with tool invocation capabilities, but evidence filtering, graph traversal, numerical calculation, version verification, state writing, and transaction commit are still executed by deterministic procedures. GIS output can be a two-dimensional thematic layer or a three-dimensional scene layer, both consuming the same layer difference according to the affected object identifier, target minute, transaction number, and idempotent term number.

[0067] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for minute-level precipitation monitoring and impact analysis based on intelligent agents, characterized in that, include: S1. Collect the event minutes, completion progress and evidence version of the multi-source precipitation records by the intelligent agent according to the source partition, and form minute precipitation evidence frames; S2. The monitoring agent identifies precipitation events and changes in evidence or event boundaries based on the minute precipitation evidence frames, forming a versioned precipitation event slice with event lineage and event topological difference. S3. The orchestration agent maps the versioned precipitation event slices to the geographic impact dependency graph on the specified graph version and time delay profile version. Based on the path propagation delay and operator time support, it forms a time delay constraint impact subgraph including the source minutes required for the target minute, the necessary source partitions, and the topology change state key. S4. The orchestration agent summarizes the evidence correction state key and the topology change state key according to the time delay constraint influence subgraph to form a state update set. The influence analysis agent marks the influence state associated with the addition, remapping or evidence correction instruction as submitted or provisional, and marks the influence state associated with the revocation instruction as revoked, according to the time support, the completion progress of the required source partition and the consistency of the evidence, graph and computing configuration versions, to form a consistent minute influence state set. S5. The publishing agent forms a layer difference based on the consistent minute impact state set, and generates a targeted supplementary sampling request based on the transaction identifier of the layer difference according to the missing evidence set in the consistent minute impact state set, thus forming the minute precipitation impact monitoring result.

2. The method according to claim 1, characterized in that, The data acquisition agent records the event minutes, completion progress, and evidence versions of multi-source precipitation records according to source partitions, forming minute-by-minute precipitation evidence frames, including: Maintain the sequence number continuity and the source partition water level for the source partition key composed of the source identifier and the partition identifier, respectively; If all records due at or before the corresponding event minute in a source partition have been received, or if an explicit missing marker with spatial location and event minute is written for the missing records, the waterline of the source partition will cross the event minute. As the numerical value, quality status, or missing status of the source record is corrected, the evidence version is incremented and the replaced version is retained. The source partitions are aggregated according to spatial units and event minutes to form the minute precipitation evidence frame.

3. The method according to claim 1, characterized in that, The monitoring agent identifies precipitation events and changes in evidence or event boundaries based on the minute-by-minute precipitation evidence frames, forming versioned precipitation event slices with event lineage and event topological differences, including: Spatial connectivity is marked for spatial units that are of valid quality and reach the precipitation identification threshold, and new event components are formed based on the continuity of adjacent event minutes; A two-part association relationship is constructed based on the spatial overlap between the new event component and the previous version event component. Event component matching is performed based on the two-part association relationship. When an event splits, the successor event component with the largest spatial overlap inherits the event identifier of the previous version event component. When multiple successor event components have the same maximum spatial overlap, they are sorted according to the smallest spatial unit identifier of each successor event component. The successor event component with the first sorted value inherits the event identifier of the previous version event component. When an event merges, the new event component inherits the event identifier of the previous version event component with the largest spatial overlap. When multiple previous version event components have the same maximum spatial overlap, the new event component inherits the event identifier with the smallest lexicographical order, thus forming an event lineage record. By comparing the evidence citations, spatial units, event minute ranges, and event lineage records of the new and old event components, the event topology difference, including the set of newly added event slices, the set of revoked event slices, and the set of remapped event slices, is formed to obtain the versioned precipitation event slice.

4. The method according to claim 1, characterized in that, The process involves mapping the versioned precipitation event slices to a geographic impact dependency map on a specified map version and time-delay profile version. Based on path propagation delay and operator time support, a time-delay constrained impact sub-map is formed, including the source minutes required for the target minute, the necessary source partitions, and the topology change state key. The propagation delay support of each side in the path from the source spatial unit covered by the event to the affected object is combined with the time support of the operators along the path by interval synthesis or discrete convolution to obtain the path time support, and the versioned path time kernel is obtained in the discrete convolution method. Based on the path time support, the source minutes required for the target minutes are derived from the target minutes, and the required source partitions are determined according to the spatial mapping relationship and operator configuration; According to the deterministic aggregation rule corresponding to the path time support or the versioned path time kernel, consume all the necessary evidence corresponding to the source minutes required for the target minute and the necessary source partition, and combine it with the topology change state key to form the delay constraint influence subgraph.

5. The method according to claim 1, characterized in that, The process involves mapping the versioned precipitation event slices to a geographic impact dependency map on a specified map version and time-delay profile version. Based on path propagation delay and operator time support, a time-delay constrained impact sub-map is formed, including the source minutes required for the target minute, the necessary source partitions, and the topology change state key. The versioned precipitation event slices are mapped to the specified map version to form candidate map data including the current path, current source partition, current time support, and current affected object mapping; The candidate graph data is compared with the previous version of the delay constraint-affected subgraph, and the topology change state key is determined according to the mutual exclusion relationship. Among them, the key that does not have an existing affected state but has the current affected object mapping is determined as the new state key, the key that has an existing affected state but the current affected object mapping disappears is determined as the cancel state key, and the key that has an existing affected state, the current affected object mapping is retained and the event lineage, graph version, path, time support or affected object mapping changes is determined as the remapping state key. The newly added state key, the undo state key, and the remapping state key are written into the candidate graph data to form the delay constraint influence subgraph.

6. The method according to claim 1, characterized in that, The orchestration agent, based on the time-delay constraint influence subgraph, summarizes the evidence correction state keys and the topology change state keys to form a state update set. The influence analysis agent, based on time support, the completion progress of required source partitions, and the consistency of evidence, graph, and computation configuration versions, marks the influence states associated with new, remapping, or evidence correction instructions as committed or provisional, and marks the influence states associated with revocation instructions as revoked, forming a consistent minute-by-minute influence state set, including: The existing state key that supports the signature referencing the replaced evidence version and whose correction event minute belongs to the source minute required for the target minute is determined as the evidence correction state key; Read the topology change state key from the time delay constraint influence subgraph, summarize and deduplicate the evidence correction state key and the read topology change state key to form the state update set; For the state update set, evidence correction, addition, revocation, and remapping instructions are retained. For keys that have both revocation and other instructions, the revocation instruction is used as the standard and a deletion tombstone is generated. For keys that do not have a revocation instruction but have an addition instruction, initial calculation is performed. For keys that do not have a revocation instruction, do not have an addition instruction, but have evidence correction or remapping instructions, recalculation is performed according to the topological order of the time delay constraint-affected subgraph. For keys that have undergone initial calculation or recalculation, a supporting signature for the actual consumption version of the record is generated. Based on the completion progress of the required source partition, the validity of the evidence, the consistency of the version corresponding to the supported signature, and the deletion of the tombstone, the consistency minute impact state set is formed.

7. The method according to claim 6, characterized in that, The fields supporting the signature include the affected object identifier, target minute, event version, actual consumption evidence version, fusion rule version, graph version, path delay profile version, time support configuration version, operator version, parameter version, and the state writing benchmark read during computation. Before writing the state, the orchestration agent rereads the current values ​​of the evidence version, event version, fusion rule version, graph version, path delay profile version, time support configuration version, operator version, and parameter version and checks them item by item against the corresponding fields in the supporting signature. It also rereads the current value of the state version and checks it against the state writing benchmark. If both checks are consistent, the affected state to be written is located using the affected object identifier and the target minute, and the state writing benchmark is used as the expected state version for comparison and exchange. If any check is inconsistent or the comparison and exchange fails, the state is re-queued. If the current value of the graph version, path delay profile version, time support configuration version, or operator version is inconsistent with the corresponding field in the supporting signature, the delay constraint affected subgraph is regenerated.

8. The method according to claim 1, characterized in that, The impact analysis agent, based on time support, the completion progress of required source partitions, and the consistency of evidence, graph, and computation configuration versions, marks the impact states associated with add, remap, or evidence correction instructions as committed or provisional, and marks the impact states associated with revocation instructions as revoked, forming a consistent minute-by-minute impact state set, including: Select the target affected object, and when the revocation instruction is read from the time delay constraint influence subgraph, mark the influence status of the target affected object as revocation, and generate a deletion tombstone to instruct the publishing agent to delete the original layer object; If the revocation instruction is not read and the necessary evidence is incomplete, the influence status of the target object is marked as tentative, and the status value is set to empty, the valid status value is set to invalid, the status level is set to undetermined, and the candidate value is only used for interface preview. If the revocation instruction is not read, and the required evidence is complete and consistent with the version, the impact status of the target object is marked as committed; If the revocation instruction is not read, and the required evidence is complete but there is a version inconsistency, the current calculation result is not marked as submitted. Instead, depending on the type of version inconsistency, the impact status of the target object is marked as provisional and then re-queued, or the time delay constraint impact subgraph is regenerated. The impact states marked as committed, provisional, or withdrawn are aggregated to form the consistent minute impact state set.

9. The method according to claim 1, characterized in that, The publishing agent forms a layer difference based on the consistent minute-level impact state set, and generates a targeted supplementary sampling request using the transaction identifier of the layer difference based on the missing evidence set in the consistent minute-level impact state set, thus forming a minute-level precipitation impact monitoring result, including: Generate layer difference dispatch records based on the layer difference, and generate targeted supplementary dispatch records based on the targeted supplementary dispatch request; Establish a local database transaction, and write the layer differential dispatch record, the targeted supplementary sampling dispatch record, and the result transaction record into the local database transaction, so that the three types of records contain the same transaction number and target minute, and make the layer differential dispatch record and the targeted supplementary sampling dispatch record each contain their respective idempotency term numbers; Submit the local database transaction and output the minute precipitation impact monitoring results, which include the layer difference, the targeted supplementary sampling request, and the three types of record identifiers.

Images