An agent-based flood-prevention meteorological daily report generation method and system

Abstract

The application is suitable for the technical field of meteorological daily report, and provides a flood prevention meteorological daily report generation method and system based on an agent, which comprises the following steps: loading strategy parameters, a dynamic threshold table, an engineering object set, a daily report style template and a timing listening configuration, establishing a task context, and the daily report style template comprising chapter slot positions; determining whether the task type of this time is timing report or real-time correction, and if it is timing report, generating a complete directed acyclic task graph; executing each node according to the task graph, wherein a visual research and judgment agent performs positioning, classification, quantification and narration on four steps of research and judgment of meteorological real-time graphs and forecast graphs; and an arrangement agent dispatching a report agent fills each data object in a fact package according to the chapter slot positions, and generates a flood prevention meteorological daily report document. Through the full-automatic task graph scheduling of the arrangement agent, the traditional serial manual process is changed into a parallel automatic pipeline, and the minute-level report requirement of flood prevention emergency response is met.

Description

Technical Field

[0001] This invention relates to the field of meteorological daily reports, specifically to a method and system for generating flood control meteorological daily reports based on intelligent agents. Background Technology

[0002] Daily flood control reports are the core basis for engineering flood control consultations and dispatch decisions. They have a strict format, typically including fixed sections such as rainfall and water levels, engineering conditions and potential hazards, response plans, weather trends, and multi-term forecast maps. Currently, the compilation of these reports mainly relies on manual operation, which has the following significant technical defects and industry pain points: In the traditional process, experts need to manually review satellite cloud images, radar echoes, numerical forecast grid points, and reservoir hydrological subsystem data before writing and formatting the reports according to the format. This manual mode cannot meet the timeliness requirements of emergency response. Different on-duty experts have different interpretations of the same rainfall forecast map, qualitative descriptions of rainfall levels, and statistical standards for reservoirs exceeding flood limits across multiple engineering lines, resulting in inconsistent report quality and a lack of standardized objective criteria. Therefore, there is a need to provide a method and system for generating daily flood control meteorological reports based on intelligent agents, aiming to solve the above problems. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for generating flood control meteorological daily reports based on intelligent agents, so as to solve the problems existing in the above-mentioned background technology.

[0004] This invention is implemented as follows: a method for generating daily flood control weather reports based on intelligent agents, the method comprising the following steps:

[0005] Loading strategy parameters, dynamic threshold table, and project object collection. Daily report format template and timed monitoring configuration, establish task context, the daily report format template includes chapter slots S0~S7;

[0006] The task type is determined as either scheduled reporting or real-time correction. If it is scheduled reporting, a complete directed acyclic task graph is generated, which includes data acquisition, structured evidence extraction, graph-data consistency check, fusion and time-series trend analysis, risk and hydrological statistics, visual assessment, reporting and archiving. If it is real-time correction, an incremental subgraph containing only the affected nodes is generated based on the triggering reason.

[0007] Each node is executed according to the task diagram. The visual judgment agent performs four steps of judgment on the meteorological map and forecast map: location, classification, quantification and description. The quantitative result must meet the gate condition Gatei=pass if and only if Δi≤θgate. Gatei represents the gate judgment result of region i, Δi is the absolute value of the difference between the image inversion quantity level and the structured evidence level, and θgate is the maximum allowed level difference.

[0008] The orchestration agent schedules the reporting agent to fill the fact packet according to the chapter slots S0~S7. For each data object, generate and archive the daily flood control meteorological report document; in real-time correction mode, also generate chapter-level difference records including changes in trough identifiers, risk level changes, and changes in the number of troughs exceeding the flood limit.

[0009] As a further aspect of the present invention, the orchestration agent employs a finite state machine to orchestrate nodes. The state set includes initialization, planning, observation, thinking, action, verification, retry, review, degradation, success, and failure. During node execution, the following steps are performed sequentially: observing and reading upstream results, thinking and comparing against conditions to generate a judgment, and action and submitting the result or transitioning to exception handling. Errors can be recovered by entering a retry state and waiting in an exponential backoff manner before re-execution. The waiting time for the nth retry is Δtwait = min(2...). n (60) seconds; if quantitative gating fails or the maximum difference between the three risk levels is ≥2, it enters the review state; if there is a fatal error or the retry limit is exceeded, it enters the degraded state, and a rule template is used to generate a report and the degraded mode is marked.

[0010] As a further aspect of this invention, the four-step judgment performed by the visual judgment agent are as follows: The localization step outputs the key region number, the range polygon, and the coordinates of the center point; the classification step outputs weather-type labels, taking values ​​from stratiform, convection, front, typhoon, clear sky, and mixed; the quantitative step outputs the upper and lower bounds of the physical quantities of each region and the representative value qmid; the descriptive step generates a professional description of no more than 200 characters, which must reference the located region number; the structured evidence extraction involves converting the image from RGB to HSV and generating a binary mask according to a color scale table. K-means clustering is performed on pixels with a mask of 1 to extract the edge contours of each cluster and calculate the area ratio and compactness. Finally, image evidence is obtained by inverting the color scale levels. .

[0011] As a further aspect of the present invention, the chapter slots S0 to S7 correspond to the heading, rainfall, water conditions, engineering and emergency situations, response plans, weather trends, forecast diagrams, and the closing paragraph, respectively.

[0012] As a further aspect of the present invention, in the risk and hydrological statistics, multi-source risk decision-making is performed as follows: visual risk Rvis, numerical risk Rnum, and time-series risk Rts are calculated using a dynamic threshold Tdy=Tbase·αsea·αper·βrisk·γcli·δcli, where Tbase is the base threshold, αsea is the seasonal factor, αper is the time-period factor, βrisk is the hazard area factor, γcli is the climate anomaly correction factor, and δcli is the climate anomaly index; the final risk level R*=max(Rvis, Rnum, Rts); if R*-min(Rvis, Rnum, Rts)≥2, a consistency review is triggered, and the multi-source discrepancies must be stated in the report; if δcli>1.5, R* is increased by one level but does not exceed the red level.

[0013] As a further aspect of the present invention, the triggering condition for real-time correction is that any of the following events are met: any element observation value X exceeds the dynamic threshold Tdy; the climate anomaly index δcli > 1.5; the relative change rate of the area ratio of any region > 0.3; the number of reservoirs exceeding the flood limit on any engineering line changes; the time series trend predicts that the risk level will be upgraded in the next 3 hours; the same region is only fully re-executed once within 30 minutes; the corresponding incremental sub-graph is selected according to the triggering type: when the rainfall intensity or the landing area changes abruptly, the data acquisition, structured evidence extraction, risk and hydrological statistics, visual judgment, and reporting nodes are re-executed, and the S1, S5, and S6 slots are updated first; when the number of reservoirs exceeding the flood limit changes, the data acquisition, risk and hydrological statistics, and reporting nodes are re-executed, and the S2 and S4 slots are updated first; when the forecast field is updated, the data acquisition, fusion and time series trend analysis, visual judgment, and reporting nodes are re-executed, and the S5 and S6 slots are updated first.

[0014] As a further embodiment of the present invention, the nodes of the complete directed acyclic task graph are as follows: T1 Acquire meteorological images, numerical grid fields, automatic station sequences, reservoir water conditions and engineering status; T2 Extract structured evidence from the actual situation map and forecast map; T3 Acquire numerical products and water and engineering statistics; T4 Perform multimodal fusion and time series trend analysis of radar, satellite, numerical forecast and time series features to obtain the fused feature Tfused and trend labels for the next 3 hours, 6 hours and 12 hours, with the trend label value being enhanced, maintained or weakened; T5 Calculate the visual, numerical and time series risks according to dynamic thresholds and take the highest level, while counting the number of Noover and the number of Noover exceeding the design limit for each engineering line; T6 The visual judgment agent performs four-step judgment on the actual situation map and forecast map and passes the gating; T7 The reporting agent fills the slots to generate the main text of each chapter; T8 Connect the 72-hour, 48-hour and 24-hour rainfall forecast maps to S6 and perform trend-attachment analysis. Figure 1 Consistency check; T9 document export; T 10Archive and record the execution path.

[0015] Another objective of this invention is to provide a flood control meteorological daily report generation system based on intelligent agents. The system includes a front-end interaction layer, an intelligent agent scheduling layer, and a meteorological and engineering data layer. The front-end interaction layer is responsible for task distribution and result display. The intelligent agent scheduling layer is configured with an orchestration intelligent agent, a visual judgment intelligent agent, a report generation intelligent agent, and a professional toolset. The orchestration intelligent agent is used to perform task graph generation and node orchestration scheduling; the visual judgment intelligent agent is used to perform four-step judgment and gating; and the report generation intelligent agent is used to perform slot filling and document generation. The professional toolset includes an image structured evidence extraction tool, an image-data consistency check tool, a multimodal fusion tool, a time-series trend extrapolation tool, a dynamic threshold and risk classification tool, a hydrological and engineering status statistics tool, a response deployment rule matching tool, and a forecast map attachment tool. The meteorological and engineering data layer provides grid field, station sequence, reservoir water level, and engineering status data. Scheduled tasks and real-time tasks run in separate queues, with real-time tasks prioritizing computing power.

[0016] Compared with the prior art, the beneficial effects of the present invention are:

[0017] By orchestrating intelligent agents and fully automated task graph scheduling, the traditional serial manual process is transformed into a parallel automated pipeline, meeting the minute-level reporting requirements of high-level flood control emergency responses. A four-step decoupling method—location, classification, quantification, and narration—is used, introducing an independent image structured evidence extraction algorithm as an objective benchmark, with a hard threshold for difference. If the quantitative result does not match the objective color scale of the image, a forced reread or review is required, ensuring a high first-pass rate for quantitative judgment from the algorithmic level. Formalized chapter slot constraints ensure strict alignment between the machine-generated daily reports and business templates. Rolling updates further ensure that every modification is traceable. The finite state machine and exponential backoff retry mechanism of the orchestrating intelligent agents endow the system with self-recovery capabilities in the event of data source anomalies. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for generating daily flood control weather reports based on intelligent agents.

[0019] Figure 2 This is a structural block diagram of a flood control meteorological daily report generation system based on intelligent agents. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0022] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for generating daily flood control weather reports based on intelligent agents. The method includes the following steps:

[0023] S100, Load strategy parameters, dynamic threshold table, and project object collection. The daily report format template and timed monitoring configuration are set up to establish a task context. The daily report format template includes chapter slots S0 to S7.

[0024] In this embodiment of the invention, the loaded strategy parameters include modal prior weights, risk level mapping rules, etc.; a dynamic threshold table is used for subsequent multi-source risk decision-making; and an engineering object set... Clearly defining the specific engineering lines under jurisdiction (such as the central line and the eastern line) and the list of reservoirs along these lines forms the basis for line-specific statistics. The daily report format template defines an unchangeable structure of eight chapter slots, S0 to S7, ensuring that the output format strictly aligns with business specifications. These chapter slots S0 to S7 correspond to the heading, rainfall, water levels, engineering and hazard information, response plans, weather trends, forecast maps, and the closing paragraph, respectively.

[0025] S200 determines whether the task type is scheduled reporting or real-time correction. If it is scheduled reporting, a complete directed acyclic task graph is generated, including data acquisition, structured evidence extraction, graph-data consistency check, fusion and time-series trend analysis, risk and hydrological statistics, visual assessment, reporting and archiving. If it is real-time correction, an incremental subgraph containing only the affected nodes is generated based on the triggering reason.

[0026] In this embodiment of the invention, the orchestration agent acts as the overall commander and first performs task routing. A timed mode generates a full directed acyclic graph (DAG) containing all nodes T1-T from data retrieval to archiving. 10 The real-time mode, on the other hand, trims out an "incremental subgraph" containing only the affected nodes based on the triggering cause, thereby achieving a lightweight and fast response.

[0027] S300 executes each node according to the task map schedule. The visual judgment agent performs four-step judgment on the meteorological real-time map and forecast map: location, classification, quantification and description. The quantitative result must meet the gate condition Gatei=pass if and only if Δi≤θgate. Gatei represents the gate judgment result of region i, Δi is the absolute value of the difference between the image inversion quantity level and the structured evidence level, and θgate is the maximum allowed level difference.

[0028] In this embodiment of the invention, the orchestration agent proceeds according to the DAG (Directed Acyclic Graph). The visual judgment agent does not directly generate a conclusion, but is forcibly broken down into four steps: localization, classification, quantification, and narration. Its core anti-hallucination mechanism lies in quantitative gating: the rainfall value output by the model is compared with the level retrieved from the structured evidence in the image (Δi≤θgate). If the condition is not met, a limited number of image rereads are triggered, thereby suppressing the speculative output of the large visual model from the root.

[0029] S400, the orchestration agent schedules the reporting agent to fill the fact packet according to the chapter slots S0~S7. For each data object, generate and archive the daily flood control meteorological report document; in real-time correction mode, also generate chapter-level difference records including changes in trough identifiers, risk level changes, and changes in the number of troughs exceeding the flood limit.

[0030] In this embodiment of the invention, the reporting agent packages the structured facts generated upstream. The data is then filled into each slot using a mapping function. In real-time mode, the system does not simply overwrite old reports, but generates structured difference records containing "changes in slot identifiers," "changes in risk levels," and "changes in the number of seats exceeding flood limits," thus enabling traceability of modification history.

[0031] In a preferred embodiment of the present invention, the orchestration agent employs a finite state machine to orchestrate nodes. The state set includes initialization, planning, observation, thinking, action, verification, retry, review, degradation, success, and failure. During node execution, the following steps are performed sequentially: observing and reading upstream results, thinking and comparing against conditions to generate a judgment, and submitting action to either pass the result or transition to exception handling. This refined decomposition allows for precise location of failure points. For recoverable errors (such as network jitter), recoverable errors enter a retry state and wait exponentially before re-execution. The waiting time for the nth retry is Δtwait = min(2...). n (60) seconds are used to avoid unnecessary and intensive retries that exacerbate system congestion. When quantitative gating fails or the maximum difference between the three risk levels is ≥2, the system enters a review state, allowing manual or secondary verification intervention. Fatal errors or exceeding the retry limit result in a degraded state, which uses rule templates to generate reports and marks the degraded mode, ensuring high availability of the system under extreme conditions.

[0032] As a preferred embodiment of the present invention, the four-step judgment performed by the visual judgment agent are as follows: the localization step outputs the key area number, the range polygon, and the coordinates of the center point; the classification step outputs weather-type labels, taking values ​​from stratiform, convection, front, typhoon, clear sky, and mixed; the quantitative step outputs the upper and lower bounds of the physical quantities of each area and the representative value qmid, where the representative value is the median of the upper and lower bounds; the descriptive step generates a professional description of no more than 200 words, which must cite the located area number; the structured evidence extraction is performed by converting the image from RGB to HSV and generating a binary mask according to the color scale table, performing K-means clustering on the pixels with a mask of 1, extracting the edge contours of each cluster and calculating the area ratio and compactness, and then inverting according to the color scale level to obtain the image evidence. Define the gear difference Δi = |bin(qmid) - bin( `bin()` is a function that maps physical quantities to light, medium, and heavy weight color scales. The clever use of scales instead of absolute values ​​for comparison is to tolerate minor numerical errors. Setting `θgate=1` means that the rainfall level retrieved by the model from the image must match the actual color scale level of the image. If the difference is more than one scale (e.g., reading moderate rain as heavy rain), it is considered a hallucination, triggering a reread. If the reread still fails after two attempts, it is better to mark it for further review than to output erroneous data.

[0033] In a preferred embodiment of the present invention, the multi-source risk decision-making in the risk and hydrological statistics is performed as follows: Visual risk Rvis, numerical risk Rnum, and temporal risk Rts are calculated using a dynamic threshold Tdy = Tbase·αsea·αper·βrisk·γcli·δcli, respectively. The threshold is no longer fixed but dynamically adjusted according to season, time period, hazard level, and climate anomaly, reflecting refined flood control management experience. Tbase is the base threshold, αsea is the seasonal factor, αper is the time period factor, βrisk is the hazard area factor, γcli is the climate anomaly correction factor, and δcli is the climate anomaly index. The final risk level R* = max(Rvis, Rnum, Rts). If R* - min(Rvis, Rnum, Rts) ≥ 2, it indicates a severe conflict between data sources, triggering a consistency review, and the multi-source discrepancy must be stated in the report. If δcli > 1.5, R* is increased by one level but not exceeding the red level. This indicates abnormal current weather, thus capturing unconventional risks under extreme weather conditions.

[0034] As a preferred embodiment of the present invention, the triggering condition for real-time correction is that any of the following events are met: the observed value X of any element exceeds the dynamic threshold Tdy; the climate anomaly index δcli > 1.5; the relative change rate of the area proportion of any region > 0.3; the number of reservoirs exceeding the flood limit on any engineering line changes; the time series trend predicts that the risk level will be upgraded in the next 3 hours, realizing trend-based predictive correction; the same area is only fully re-executed once within 30 minutes, effectively preventing system storms caused by data fluctuations near the critical value. The corresponding incremental sub-graph is selected according to the trigger type: when rainfall intensity or precipitation area changes abruptly, data acquisition, structured evidence extraction, risk and hydrological statistics, visual judgment, and reporting nodes are re-executed, with priority given to updating slots S1, S5, and S6; when the number of reservoirs exceeding the flood limit changes, data acquisition, risk and hydrological statistics, and reporting nodes are re-executed, with priority given to updating slots S2 and S4; when the forecast field is updated, data acquisition, fusion and time series trend analysis, visual judgment, and reporting nodes are re-executed, with priority given to updating slots S5 and S6.

[0035] In a preferred embodiment of the present invention, the nodes of the complete directed acyclic task graph are as follows: T1 acquires meteorological images, numerical grid fields, automatic station sequences, and reservoir water and engineering status; T2 extracts structured evidence from the actual map and forecast map; T3 acquires numerical products and water and engineering statistics; T4 performs multimodal fusion and time series trend analysis of radar, satellite, numerical forecast and time series features to obtain the fused feature Tfused and trend labels for the next 3 hours, 6 hours and 12 hours. The trend label values ​​are enhanced, maintained or weakened. The T4 node integrates radar (short-term), satellite (large-scale), numerical forecast (future) and self-observation history (time series) to provide qualitative trend judgment for S5 (weather trend). T5 calculates the risk levels based on visual, numerical, and temporal thresholds, taking the highest level, and simultaneously counts the number of seats exceeding the flood limit and the number exceeding the design limit for each engineering line; T6 The visual analysis agent performs a four-step analysis on the real-time map and forecast map, and passes the gating; T7 The reporting agent generates the main text of each chapter by filling in the slots; T8 Connects the 72-hour, 48-hour, and 24-hour rainfall forecast maps to S6 and executes the trend-attached analysis. Figure 1 Consistency check; T9 document export; T 10 Archive and record execution traces to support audit backtracking.

[0036] In a preferred embodiment of the present invention, the reporting agent performs reporting and degradation according to the following steps: filling in the S0 header entry; generating the S1 rainfall section from rainfall facts and visual descriptions; and verifying whether the maximum rainfall Pmax for the time period is written into the main text. (Based on the engineering object set...) Generate the S2 hydrological information section line by line. Generate the S3 engineering situation and hazard section based on the engineering status mapping. Generate the S4 response and deployment section based on rule base matching. Generate the S5 weather trend section based on trend tags and key points of sub-line forecasts. Attach the 72-hour, 48-hour, and 24-hour forecast maps to S6. If heavy rain or torrential rain is described in S5 but there is no evidence of the same magnitude in the corresponding time-leading forecast map in S6, mark it as pending review. Fill in the S7 document ending. Export the document and archive it; if the visual analysis service is unavailable, use structured evidence, station extreme values, and hydrological statistics combined with rule templates to generate a report and clearly mark the downgrade mode. If the S6 attached map is missing, declare it in the text.

[0037] like Figure 2 As shown in the figure, this embodiment of the invention also provides a flood control meteorological daily report generation system based on intelligent agents. The system includes a front-end interaction layer, an intelligent agent scheduling layer, and a meteorological and engineering data layer. The front-end interaction layer is responsible for task distribution and result display. The intelligent agent scheduling layer is configured with an orchestration intelligent agent, a visual judgment intelligent agent, a report generation intelligent agent, and a professional toolset. The orchestration intelligent agent is used to perform task graph generation and node orchestration scheduling, the visual judgment intelligent agent is used to perform four-step judgment and gating, and the report generation intelligent agent is used to perform slot filling and document generation. The professional toolset includes an image structured evidence extraction tool, a map-data consistency check tool, a multimodal fusion tool, a time-series trend inference tool, a dynamic threshold and risk classification tool, a hydrological and engineering status statistics tool, a response deployment rule matching tool, and a forecast map attachment tool. The meteorological and engineering data layer provides grid field, station sequence, reservoir water level, and engineering status data. Scheduled tasks and real-time tasks run in separate queues, with real-time tasks taking priority in consuming computing power.

[0038] The above description only details the preferred embodiments 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 scope of protection of the present invention.

[0039] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0040] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A method for generating daily flood control meteorological reports based on intelligent agents, characterized in that, The method includes the following steps: Loading strategy parameters, dynamic threshold table, and project object collection. Daily report format template and timed monitoring configuration, establish task context, the daily report format template includes chapter slots S0~S7; The task type is determined as either scheduled reporting or real-time correction. If it is scheduled reporting, a complete directed acyclic task graph is generated, which includes data acquisition, structured evidence extraction, graph-data consistency check, fusion and time-series trend analysis, risk and hydrological statistics, visual assessment, reporting and archiving. If it is real-time correction, an incremental subgraph containing only the affected nodes is generated based on the triggering reason. Each node is executed according to the task diagram. The visual judgment agent performs four steps of judgment on the meteorological map and forecast map: location, classification, quantification and description. The quantitative result must meet the gate condition Gatei=pass if and only if Δi≤θgate. Gatei represents the gate judgment result of region i, Δi is the absolute value of the difference between the image inversion quantity level and the structured evidence level, and θgate is the maximum allowed level difference. The orchestration agent schedules the reporting agent to fill the fact packet according to the chapter slots S0~S7. For each data object, generate and archive the daily flood control meteorological report document; in real-time correction mode, also generate chapter-level difference records including changes in trough identifiers, risk level changes, and changes in the number of troughs exceeding the flood limit.

2. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 1, characterized in that, The orchestration agent uses a finite state machine to orchestrate nodes. The state set includes initialization, planning, observation, thinking, action, verification, retry, review, degradation, success, and failure. When a node is executed, it sequentially performs the following actions: reading upstream results from observation, generating a judgment by comparing with the conditions from thinking, and submitting the action to either pass the result or enter the exception handling stage. Errors can be recovered and the system can enter a retry state, waiting in an exponential backoff manner before re-execution. The waiting time for the nth retry is Δtwait = min(2). n (60) seconds; if quantitative gating fails or the maximum difference between the three risk levels is ≥2, it enters the review state; if there is a fatal error or the retry limit is exceeded, it enters the degraded state, and a rule template is used to generate a report and the degraded mode is marked.

3. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 1, characterized in that, The four-step judgment process performed by the visual judgment agent is as follows: the localization step outputs the key area number, the range polygon and the coordinates of the center point; the classification step outputs weather-type labels, with values ​​ranging from stratiform, convection, front, typhoon, clear sky and mixed; the quantitative step outputs the upper and lower bounds of the physical quantities of each area and the representative value qmid; and the descriptive step generates a professional description of no more than 200 words, which must reference the localized area number. The structured evidence extraction process involves converting the image from RGB to HSV, generating a binary mask according to a color scale, performing K-means clustering on pixels with a mask of 1, extracting the edge contours of each cluster, calculating the area ratio and compactness, and then retrieving the image evidence according to the color scale level. .

4. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 1, characterized in that, The chapter slots S0 to S7 correspond to the heading, rainfall, water conditions, engineering and emergency situations, response plans, weather trends, forecast charts, and the closing paragraph, respectively.

5. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 4, characterized in that, In the aforementioned risk and hydrological statistics, multi-source risk decision-making is performed as follows: Visual risk Rvis, numerical risk Rnum, and time-series risk Rts are calculated using a dynamic threshold Tdy = Tbase·αsea·αper·βrisk·γcli·δcli, where Tbase is the base threshold, αsea is the seasonal factor, αper is the time-period factor, βrisk is the hazard area factor, γcli is the climate anomaly correction factor, and δcli is the climate anomaly index; the final risk level R* = max(Rvis, Rnum, Rts); if R* - min(Rvis, Rnum, Rts) ≥ 2, a consistency review is triggered, and the multi-source discrepancies must be stated in the report; if δcli > 1.5, R* is increased by one level but does not exceed the red level.

6. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 4, characterized in that, The real-time correction is triggered by any of the following events: any observed value X of any element exceeds the dynamic threshold Tdy; the climate anomaly index δcli > 1.5; the relative change rate of the area ratio of any region > 0.3; the number of reservoirs exceeding the flood limit on any engineering line changes; the risk level is predicted to be upgraded in the next 3 hours according to the time series trend; the same area is only fully re-executed once within 30 minutes; the corresponding incremental sub-map is selected according to the trigger type: when the rainfall intensity or the landing area changes abruptly, the data acquisition, structured evidence extraction, risk and hydrological statistics, visual judgment, and reporting nodes are re-executed, and the S1, S5, and S6 slots are updated first; when the number of reservoirs exceeding the flood limit changes, the data acquisition, risk and hydrological statistics, and reporting nodes are re-executed, and the S2 and S4 slots are updated first; when the forecast field is updated, the data acquisition, fusion and time series trend analysis, visual judgment, and reporting nodes are re-executed, and the S5 and S6 slots are updated first.

7. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 4, characterized in that, The nodes of the complete directed acyclic task graph are as follows: T1 acquires meteorological images, numerical grid fields, automatic station sequences, and reservoir water and engineering status; T2 extracts structured evidence from the actual situation map and forecast map. T3 acquires numerical products and hydrological and engineering statistics; T4 performs multimodal fusion and time series trend analysis of radar, satellite, numerical forecast and time series features to obtain the fused feature Tfused and trend labels for the next 3 hours, 6 hours and 12 hours. The trend label values ​​are enhanced, maintained or weakened. T5 calculates the risks based on visual, numerical, and temporal risk levels using dynamic thresholds and assigns the highest level. It also calculates the number of seats exceeding flood limits (Nover) and exceeding design limits for each engineering line. T6 The visual analysis agent performs a four-step analysis of the actual situation map and forecast map, and passes the gating process. T7 The reporting agent fills in slots to generate the main text for each chapter. T8 Connects the 72-hour, 48-hour, and 24-hour rainfall forecast maps to S6 and performs a trend-map consistency check. T9 Exports the document. 10 Archive and record the execution path.

8. The method for generating daily flood control meteorological reports based on intelligent agents according to claim 4, characterized in that, The reporting agent performs reporting and downgrading according to the following steps: Fill in the S0 header entry; The S1 rainfall chapter is generated from rainfall facts and visual descriptions, and the maximum rainfall Pmax for the time period is verified to be written into the main text. By project object set Generate the S2 water condition chapter line by line; The S3 section on work-related emergencies is generated by mapping the work status. The S4 response deployment section is generated by matching the rule base; The S5 weather trend section is generated from trend tags and key points of the sub-line forecast. Connect the 72-hour, 48-hour, and 24-hour forecast maps to S6. If S5 shows heavy rain or rainstorms but the corresponding time-leading forecast map in S6 does not have evidence of color spots of the same magnitude, then mark it as pending review. Fill in the S7 end of the document; Export and archive the document; if the visual assessment service is unavailable, use a rule template that combines structured evidence, station extreme values ​​and hydrological statistics to generate a report and clearly mark the downgrade mode. If the S6 attached figure is missing, declare it in the text.

9. A system for generating daily flood control meteorological reports based on intelligent agents, used to implement the method for generating daily flood control meteorological reports based on intelligent agents as described in any one of claims 1 to 8, characterized in that, The system comprises a front-end interaction layer, an agent scheduling layer, and a meteorological and engineering data layer. The front-end interaction layer is responsible for task distribution and result display. The agent scheduling layer is configured with an orchestration agent, a visual judgment agent, a reporting agent, and a professional toolset. The orchestration agent is used to perform task graph generation and node orchestration scheduling, the visual judgment agent is used to perform four-step judgment and gating, and the reporting agent is used to perform slot filling and document generation. The professional toolset includes image structured evidence extraction tools, image-data consistency checking tools, multimodal fusion tools, time-series trend extrapolation tools, dynamic threshold and risk classification tools, hydrological and engineering statistics tools, response deployment rule matching tools, and forecast map attachment tools. The meteorological and engineering data layer provides grid field, station sequence, reservoir water level, and engineering status data. Scheduled tasks and real-time tasks run in separate queues, with real-time tasks taking priority in consuming computing power.

Images