Urban engineering disaster prevention and reduction oriented vertical domain agent driven management and control system

Abstract

The application discloses a vertical domain intelligent agent driven management and control system for urban engineering disaster prevention and reduction, relates to the field of intelligent disaster prevention decision-making, and after multi-source heterogeneous data is subjected to space-time alignment and quality checking, a unified coupling state object is constructed through entity feature extraction, node dynamic risk assessment and cascade relationship fusion. On this basis, the probability sampling mechanism of a large language model is used to drive the generation of multiple candidate schemes, and the candidate scheme set is generated in combination with vertical domain knowledge base rules, historical case retrieval and long short-term memory. Then, the candidate schemes are screened through complexity identification and shunt deduction verification, and are subjected to cross-interception checking and weighted scoring and sequencing based on multi-dimensional resilience indexes. Finally, the decision trajectory and the checking conclusion are stored in a structured manner through a long short-term memory linkage writing mechanism, continuous rolling decision and cross-event experience reuse are supported, and integrated resilience management and control for disaster prevention, disaster disposal and disaster recovery is realized.

Description

Technical Field

[0001] This invention belongs to the field of disaster prevention and mitigation decision-making, specifically to the field of intelligent disaster prevention decision-making, and more specifically, to a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation. Background Technology

[0002] In the complex system of cities, the impact of rainstorms and floods has obvious characteristics of cross-system coupling, dynamic evolution and cascading failure. The interaction between ground transportation and underground space makes it difficult to accurately characterize the disaster situation through the analysis of a single subsystem, which puts forward an urgent need for integrated disaster prevention, mitigation and control capabilities for the whole life cycle.

[0003] However, existing research and applications in urban engineering disaster prevention and mitigation mostly focus on single aspects such as flood risk assessment, ground traffic scheduling, underground space evacuation organization, or post-disaster recovery optimization. They generally employ rule-based optimization or multi-objective optimization methods, making it difficult to uniformly represent the cascading effects between systems and to simultaneously consider multi-dimensional resilience objectives such as safety, efficiency, social equity, and resource consumption. Meanwhile, as large language models have demonstrated strong capabilities in complex reasoning and task planning, some solutions have begun to explore their introduction into urban disaster prevention auxiliary decision-making. However, when directly applied to the management and control of urban flood-coupled systems, they still face problems such as insufficient professional capabilities, severe reasoning illusions, and unstable output in complex environments. While existing solutions have made some improvements through multi-source monitoring data access, geographic information visualization, and resilience assessment, they still lack a unified state expression mechanism for the system-level evolution of ground traffic and underground space. They cannot achieve continuous extrapolation and rolling correction under constantly changing disaster conditions, and they lack structured state constraints, external simulation verification, and historical experience reuse mechanisms, making it difficult to meet the actual needs of integrated management and control of pre-disaster prevention, in-disaster response, and post-disaster recovery under urban flood disasters.

[0004] Therefore, it is necessary to construct a vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation, so as to realize the unified representation and continuous rolling decision-making of cross-system cascading impacts, thereby effectively improving the overall resilience management and control capabilities of urban complex infrastructure systems in flood disaster scenarios. Summary of the Invention

[0005] This invention application provides a vertical domain intelligent agent-driven control system for disaster prevention and mitigation in urban engineering.

[0006] The technical solution of this invention is as follows: A vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation, comprising: a heterogeneous data coupling module, used to perform coupling state modeling on multi-source heterogeneous data, including rainfall and water accumulation monitoring, road traffic operation, underground space facility topology, pedestrian and vehicle flow, and emergency resource allocation, through spatiotemporal alignment and data quality verification, to obtain coupled state objects; and a candidate solution generation module, used to generate multiple candidate solutions for the current disaster situation based on the coupled state objects, combined with the rules and case retrieval results of the vertical domain knowledge base and long short-term memory, using the probability sampling mechanism of a large language model, to obtain control solutions including lockdown, evacuation, diversion, and resource deployment. The system comprises: a candidate solution set for control actions; a solution screening module for complexity identification and breakdown simulation verification of the candidate solution set to obtain a screened solution set; a resilience score result generation module for cross-interception verification of the screened solution set in sequence based on rule consistency, state consistency, security, and goal achievement, and weighted scoring and ranking based on multi-dimensional resilience indicators to determine the resilience score result; and a solution verification and writing module for strategy determination and memory update based on the resilience score result, outputting the solution with the highest score and meeting the constraints as the optimal control strategy and retaining the candidate solution set, while writing the current round of decision trajectory and verification conclusion into long short-term memory after structured summarization.

[0007] Compared with existing technologies, the beneficial effects of this invention are as follows: Addressing the core problems of complex multi-system coupling, diverse decision-making objectives, and the difficulty of achieving unified cross-system representation and continuous rolling control in urban flood disaster scenarios, this invention proposes a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation. Specifically, based on structured coupled state modeling, multi-source heterogeneous data is spatiotemporally aligned and quality-verified. Through entity feature extraction, node dynamic risk assessment, and cascading relationship fusion, a unified coupled state object is constructed, overcoming the shortcomings of existing technologies where subsystem analysis is isolated and unable to represent the impact of cross-system cascading failures. Furthermore, a probabilistic sampling mechanism of a large language model drives the generation of multiple candidate solutions. Combined with vertical domain knowledge base rules, historical case retrieval, and long short-term memory, a candidate solution set is generated, solving the problems of insufficient professional capabilities and inference illusions inherent in general large language models. Subsequently, candidate solutions are screened through complexity identification and diversionary deduction verification, then cross-interception verification, and weighted scoring and ranking based on multi-dimensional resilience indicators to ensure the output solution is optimal in terms of safety, efficiency, fairness, and resource consumption. Ultimately, through a long short-term memory linkage writing mechanism, the decision trajectory and verification conclusions are structured and stored, supporting continuous rolling decision-making and cross-event experience reuse, and realizing integrated resilience management of pre-disaster prevention, in-disaster response and post-disaster recovery. Attached Figure Description

[0008] Figure 1This is a block diagram of a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of data flow in a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application.

[0010] Figure 3 This is a block diagram of a heterogeneous data coupling module in a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application.

[0011] Figure 4 This is a schematic diagram of the knowledge enhancement link of the vertical domain knowledge base in a vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation according to an embodiment of this application.

[0012] Figure 5 This is a schematic diagram of the memory enhancement link of long short-term memory in a vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation according to an embodiment of this application.

[0013] Figure 6 This is a schematic diagram of the expansion and trajectory classification of a multi-branch candidate action tree based on the current coupling state in a vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation according to an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0015] In view of the shortcomings of the prior art, this application proposes a vertical domain intelligent agent-driven control system 100 for urban engineering disaster prevention and mitigation. Figure 1 and Figure 2 As shown, Figure 1 This is a block diagram of a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application.

[0016] Specifically, the heterogeneous data coupling module 110 is used to perform coupling state modeling on multi-source heterogeneous data, including rainfall and water accumulation monitoring, road traffic operation, underground space facility topology, pedestrian and vehicle flow, and emergency resource allocation, through spatiotemporal alignment and data quality verification, in order to obtain a coupled state object. Figure 3This is a block diagram of a heterogeneous data coupling module in a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application. Figure 3 As shown, during urban flooding disasters, rainfall and waterlogging monitoring data, road traffic operation data, underground space facility topology data, pedestrian and vehicle flow data, and emergency resource allocation data are collected from different terminals such as meteorological sensors, traffic detectors, underground facility management platforms, passenger flow monitoring equipment, and emergency command terminals. These data types differ significantly in collection frequency, coordinate reference system, field naming rules, and data format. Furthermore, there is a lack of unified entity identifiers and cross-domain relationship descriptions among the data sources, making it impossible to directly characterize the cascading impact between surface traffic and underground space. Simultaneously, the water depth, flow velocity, inundation duration, and pedestrian density faced by each spatial node continuously change during disaster evolution, making it difficult to fully reflect the comprehensive risk level of a node by relying on only one type of data. Therefore, a heterogeneous data coupling module is used to provide a unified, complete, and time-varying updatable state input basis for subsequent candidate solution generation and resilience scoring verification.

[0017] In one exemplary embodiment of this application, the heterogeneous data coupling module 110 includes: a heterogeneous data structuring unit 111, used to perform acquisition frequency alignment and spatial dimension unified encoding on multi-source heterogeneous data to obtain a structured input object; an entity extraction risk assessment unit 112, used to perform entity feature extraction and node dynamic risk assessment on the structured input object to obtain an entity relationship set and a node risk set; and a coupling state graph construction unit 113, used to perform cascade relationship fusion and coupling state graph construction on the entity relationship set and the node risk set to obtain a coupled state object.

[0018] The heterogeneous data coupling module 110 receives multi-source heterogeneous data covering at least the following categories: rainfall process data and water accumulation monitoring data, including hourly rainfall from rain gauges, real-time water depth values ​​collected by road surface water depth sensors, and water levels at the boundaries of drainage networks; road traffic operation data, including road capacity, intersection signal status, average vehicle speed, queue length, and road closure status; underground space facility topology data, including the connectivity between subway station concourses and platforms, entrance and exit locations and opening / closing status, stair and escalator directions, and remaining capacity; pedestrian and vehicle flow data, including regional population density, number of stranded individuals, proportion of vulnerable groups, and vehicle queue growth rate; and emergency resource allocation data, including the current location, available quantity, occupancy status, and estimated arrival time of resources such as flood barriers, mobile drainage pumps, station staff, and ground traffic controllers. The aforementioned data are generated by different data collection terminals, such as automatic rain gauges from meteorological departments, roadside detectors from traffic management departments, station monitoring terminals from rail transit operators, crowd flow heat monitoring platforms from public security departments, and resource dispatch terminals from emergency management departments. Their collection frequencies range from seconds to minutes, and their geographic coordinate reference systems and data field naming rules are also different.

[0019] The heterogeneous data structuring unit first performs time-series interpolation on the multi-source heterogeneous data with varying collection frequencies to align the timestamp dimension. Specifically, rainfall and waterlogging monitoring data are reported every minute, road traffic operation data is aggregated every 5 minutes, underground space facility status data is updated every 10 minutes, pedestrian density data is refreshed every 2 minutes, and emergency resource allocation data is updated irregularly based on the dispatch command issuance node. To unify the data with different frequencies onto the same time base, the heterogeneous data structuring unit uses the highest collection frequency, i.e., 1 minute, as the unified time step. For data with collection frequencies lower than this step step, linear interpolation or previous value preservation interpolation is performed, thus obtaining the alignment value of each data source at each unified time step. For example, if road traffic operation data has collected values ​​at minute 0 and minute 5, then the missing moments from minute 1 to minute 4 are filled using linear interpolation; since underground space facility status data is a discrete switch state quantity, the previous value retention method is used to carry over the most recently collected status value to the next update moment.

[0020] After aligning the temporal dimension, the heterogeneous data structuring unit performs coordinate system projection transformations on geographic information from different sources to achieve spatial dimension alignment. Since meteorological rain gauge coordinates use latitude and longitude geographic coordinates, traffic detector location information is based on local urban planar projection coordinates, and underground facility topology location information uses building-internal local coordinates, these three types of coordinate reference systems cannot be directly spatially correlated. The heterogeneous data structuring unit projects all geographic coordinates onto a unified planar coordinate reference system, enabling spatial entities from different sources to perform location matching and neighborhood association within the same coordinate space.

[0021] After spatiotemporal alignment, the heterogeneous data structuring unit uses a Uniform Identifier (UID) mapping table to standardize and unify the fields of the cleaned and aligned data. The UID mapping table is a pre-built multi-column lookup table whose structure includes a data source identifier column, an original field name column, a UID column, and a data type and unit column. Domain engineers establish a one-to-one correspondence between the original field names and the UIDs for each data source's field definition documents. For example, a field named "real-time water depth" in a meteorological data source, a field named "road surface water" in a transportation data source, and a field named "station water level" in an underground space data source are all mapped to the UID "water depth" in the UID mapping table, with the unit unified to meters. Through line-by-line matching and field renaming operations in this mapping table, fields with the same semantics but different names from different data sources are merged under the same identifier, eliminating semantic ambiguity caused by field heterogeneity.

[0022] After field unification, the heterogeneous data structuring unit performs outlier removal on each field value, marking and removing data exceeding reasonable physical thresholds as invalid. For example, if the road surface water depth value exceeds 3.0 meters or is negative, it is determined to be an abnormal sensor reading and removed. All data that passes quality verification is assembled and encapsulated according to four levels: basic metadata, entity status, relationships, and rule constraints, generating a structured input object with a unified access format. The basic metadata level records the current timestamp, spatial coverage, event stage identifier, and data confidence level; the entity status level records the current attribute values ​​of objects such as road nodes, underground space entrances and exits, stairs, turnstiles, passages, functional zones, and emergency resources; the relationship level records the connectivity, dependency, and coupling effects between the above entities; and the rule constraint level records the water accumulation control threshold, traffic capacity restrictions, and evacuation priority rules.

[0023] In one exemplary embodiment of this application, the entity extraction risk assessment unit 112 includes: an entity relationship set construction subunit 1121, used to extract the attributes and topological connections of nodes from the structured input object to obtain an entity relationship set; and a risk quantification subunit 1122, used to quantify the risk of water depth, flow velocity, inundation duration, and pedestrian density of each node in the structured input object to obtain a node risk set. The entity relationship set construction subunit traverses all entity entries in the structured input object, extracting each road node, underground space entrance / exit, staircase, turnstile, passage zone, and emergency resource point as an independent entity record. Each record contains attribute fields such as entity identifier, entity type, spatial location, current accessibility, and health status. Simultaneously, the entity relationship set construction subunit extracts edge information between entities from the relationship hierarchy, including road segment connectivity between road nodes, vertical connection between underground space entrance / exit and station hall zones, and cross-domain coupling between road nodes and underground space entrance / exit. The aforementioned entity records and edge information together constitute the entity relationship set. The risk quantification subunit identifies target monitoring nodes within the potential disaster area based on hydrodynamic and pedestrian / vehicle flow data contained in the structured input object. For each target monitoring node, the risk quantification subunit extracts four real-time disaster parameters: water depth, flow velocity, inundation duration, and pedestrian density. By introducing normalized weighting coefficients, it calculates the node's comprehensive disaster risk value over time using a weighted summation method. The calculation formula is as follows: in, The identifier representing the target monitoring node. Indicates the current moment. Represents a node At any moment Normalized water depth values This represents the normalized flow rate value. This indicates the duration of flooding after normalization. This represents the normalized pedestrian density value. , , , These are normalized weighting coefficients that sum to a value of 1. The values ​​of these weighting coefficients are determined by domain engineers based on the statistical correlation between various disaster parameters and actual losses in historical flood disaster events. In a specific setting, they can be set to... =0.35、 = 0.20、 =0.20、 =0.25, with water depth having the highest weight, reflecting the dominant influence of water depth on the interruption of surface road traffic and the risk of backflow into underground spaces; pedestrian density has the second highest weight, reflecting the evacuation pressure faced by high-density population areas during disasters. Taking the northwest entrance node of a certain urban rail transit hub station as an example, at a certain moment, the measured water depth at this node is 0.32 meters, the flow velocity is 0.45 meters per second, the duration of flooding is 18 minutes, and the pedestrian density is 3.8 people per square meter. After normalization, these values ​​are mapped to: =0.64、 =0.30、 =0.36、 =0.76, substituting into the above formula, we get... =0.35×0.64+0.20×0.30+0.20×0.36+0.25×0.76=0.546. The risk quantification subunit performs the above calculation on each of the target monitoring nodes, organizes the risk values ​​of all nodes into a risk measurement matrix, and then outputs the node risk set. When a node's risk value is... When the value exceeds the preset high-risk threshold, such as 0.50, the node is marked as a high-risk node.

[0024] The coupled state graph construction unit receives an entity relationship set and a node risk set. First, using the road nodes and their road segment connectivity in the entity relationship set as a foundation, the unit instantiates a ground traffic state sub-object. This sub-object includes attributes such as the road node set, road edge set, intersection control status, average vehicle speed, queue length, and road closure status. Simultaneously, using the underground space entities and their connectivity in the entity relationship set as a foundation, the unit instantiates an underground space state sub-object. This sub-object includes attributes such as the topology, accessibility, and remaining capacity of subway station concourse zones, platforms, stairs, escalators, turnstiles, entrances / exits, underground passages, and underground commercial areas. Subsequently, the coupled state graph construction unit injects the risk metric matrix from the node risk set into the node attributes corresponding to the two physical network foundations mentioned above. For each node with a risk value record in the node risk set, its... The values ​​and corresponding original parameters such as water depth, flow velocity, inundation duration, and pedestrian density are written into the attribute fields of the node, thereby generating flood risk state sub-objects and pedestrian / vehicle flow state sub-objects. The flood risk state sub-object records the water depth, flow velocity, inundation duration, risk level, and risk propagation direction of each spatial unit; the pedestrian / vehicle flow state sub-object records the population density, proportion of vulnerable groups, passenger congestion, and path utilization rate of each area. After completing the construction of the above sub-objects, the coupled state graph construction unit further extracts the cross-domain connectivity relationships identified in the entity relationship set and establishes a cross-domain coupling edge set. The cross-domain coupling edge set is used to characterize the cascading influence relationships between ground road nodes and underground space entrances / exits, between underground parking lots and underground passages, and between ground public transport hubs and subway connection areas. Taking the aforementioned hub station as an example, a cross-domain coupling edge is established between the northwest entrance / exit node and its corresponding external low-lying ground road node. The attributes of this edge record the cascading direction of surface water flowing back into underground space and the time delay parameter of influence propagation. The coupled state graph construction unit encapsulates and graphs the ground traffic state sub-objects, underground space state sub-objects, flood risk state sub-objects, pedestrian and vehicle flow state sub-objects, emergency resource state sub-objects, and cross-domain coupling edge sets. The attributes of the emergency resource state sub-objects are directly inherited from the entity state hierarchy of the structured input object. The final output coupled state object uses a unified identifier to connect all sub-objects and edge sets, allowing the attributes, risk values, and cascading relationships with adjacent nodes of any node to be retrieved and accessed through this unified identifier.

[0025] Specifically, the candidate solution generation module 120 is used to generate multiple candidate solutions for the current disaster situation based on coupled state objects, combined with the rules and case retrieval results of the vertical domain knowledge base and long short-term memory, and utilizing the probabilistic sampling mechanism of the large language model. This results in a set of candidate solutions including control actions such as lockdown, evacuation, diversion, and resource deployment. Correspondingly, during urban flooding disasters, the cascading failures between ground transportation and underground space exhibit characteristics of multi-objective trade-offs, cross-domain linkage, and rapid evolution. The disaster situation at a single moment often corresponds to multiple feasible control paths, including different types of control actions and combinations such as local lockdown, personnel evacuation, traffic diversion, and pre-deployment of emergency resources. Traditional rule-based optimization methods, when faced with the complex disaster situation with multiple dimensions and constraints, can usually only output a single fixed solution, making it difficult to cover differentiated control strategies under different solution space dimensions, and lacking dynamic adaptability to rapid changes in the disaster situation. At the same time, although the general large language model has strong reasoning and generation capabilities, it is prone to producing outputs inconsistent with actual engineering needs when lacking professional engineering knowledge constraints and historical experience guidance. Based on this, a candidate solution generation module 120 is introduced.

[0026] In one exemplary embodiment of this application, the candidate solution generation module includes: a decision context generation unit, used to convert coupled state objects into state description text based on a preset structured prompt word template, and to perform vector dimension concatenation and field fusion with rules retrieved from the vertical domain knowledge base and experience trajectories in long short-term memory to obtain a decision context; a decision context probability sampling unit, used to generate multi-branch actions based on probability sampling of the decision context to obtain an action sequence; and an action sequence encapsulation unit, used to extract key entities and structure encapsulate the action sequence, wherein the area of ​​action, execution time, and resource requirements are mapped to a standard action structure to obtain a candidate solution set.

[0027] The decision context generation unit first parses the dynamic risk nodes and cross-domain cascading topology contained in the coupled state object, transforming them into structured state description text. Specifically, the decision context generation unit traverses all nodes marked as high-risk and their associated cross-domain coupling edges in the coupled state object, and converts each high-risk node and its cascading impact links into natural language description statements one by one, following the format of "node identifier - risk level - risk parameter - associated node - cascading direction". Taking the hub station scenario in the previous module as an example, the comprehensive risk value of the northwest entrance / exit node in the coupled state object is 0.546, the population density in the middle of the station hall is 3.8 people per square meter, the vehicle speed on the north side road has dropped to 27% of the free flow speed, and the water depth in the west staircase area is 0.12 meters and continues to increase. The state description text generated after the above information is transformed is as follows: the risk value of the northwest entrance / exit node is 0.546, the external measured water depth is 0.32 meters, the flow velocity is 0.45 meters per second, and the flooding lasts for 18 minutes; the water depth of the west staircase node associated with the cross-domain coupling edge is 0.12 meters and is showing an upward trend; the population density in the middle of the station hall is 3.8 people per square meter, and the vehicle speed at the north road node has dropped to 27% of the free flow speed; one mobile drainage pump is available, with an estimated arrival time of 6 minutes.

[0028] After generating the state description text, the decision context generation unit retrieves hard engineering rule constraints that match the current disaster situation from the vertical domain knowledge base. The vertical domain knowledge base is a structured knowledge storage system pre-built by domain engineers based on professional specifications, design standards, operation procedures, and historical event review reports in the field of urban flood disaster management. Figure 4 This is a schematic diagram of the knowledge enhancement link of the vertical domain knowledge base in a vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application. Figure 4As shown, the construction of the vertical domain knowledge base follows a knowledge enhancement chain. The input sources of this chain cover four types of domain knowledge materials: standard specifications, historical cases, engineering experience, and emergency plans. After the above four types of materials undergo rule extraction, relationship establishment, and risk constraint structured encoding in the vertical domain knowledge graph construction stage, they form a graph-based knowledge representation with domain semantics. Finally, in the form of output constraint knowledge, it provides three types of structured constraint instructions to the downstream decision-making stage: prohibited actions, threshold conditions, and recommended actions. This ensures that the subsequent candidate solution generation process is always subject to the rigid constraints and flexible guidance of professional engineering rules.

[0029] The vertical domain knowledge base stores engineering rules such as drainage capacity thresholds, water accumulation hazard thresholds, underground space lockdown thresholds, facility failure thresholds, personnel evacuation rules, resource scheduling constraints, and recovery priority rules, as well as historical case knowledge such as disaster characteristics, control action sequences, verification results, and reasons for failure in historical events. The decision context generation unit uses the risk parameters of the current high-risk node as the search key to extract matching rule entries from the vertical domain knowledge base. For example, "a lockdown plan should be activated when the water depth outside the entrance / exit exceeds 0.30 meters" and "zonal evacuation should be activated when the crowd density in the station hall exceeds 4.0 people per square meter." Simultaneously, the vertical domain knowledge base stores water accumulation lockdown threshold rules. These rules set two levels of threshold constraints based on different water depth levels: the first level is the lockdown trigger threshold, set to activate the lockdown plan when the water depth outside the entrance / exit exceeds 0.30 meters. This is a soft trigger condition, indicating that the disaster has entered the warning zone where control measures should be taken. However, within this water depth range, limited transitional passage arrangements are still permitted, provided safety conditions are met. These include short-term one-way exit evacuation to avoid sudden overcrowding due to immediate hard closure. The second layer is the danger closure threshold, set at 0.50 meters outside the entrance / exit, which mandates immediate closure. This is a hard cutoff condition, and passage is not permitted under any circumstances. These two threshold layers together constitute the complete constraint framework for waterlogging control rules in the vertical domain knowledge base. In the subsequent rule consistency verification stage, the action parameters of candidate solutions must simultaneously satisfy the constraint logic of both threshold layers to pass the binary judgment at that level.

[0030] Meanwhile, the decision context generation unit extracts historical handling experience trajectories similar to the current disaster situation from long and short-term memory. The short-term memory is a rolling cache for the current event, responsible for storing context trajectory information such as the coupling state, candidate solutions, verification results, resilience scores and execution results of the most recent rounds in the current event, so as to support continuous deduction and rolling correction within the same event. The long-term memory is a persistent storage area for cross-event experience reuse, responsible for recording high-value experience entries in historical events after structured summarization, including successful action sequences, failure reasons, correction suggestions and applicable conditions.

[0031] Figure 5 This is a schematic diagram of the memory enhancement link of long short-term memory in a vertical domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation according to an embodiment of this application. Figure 5 As shown, the input sources of the memory enhancement link include the current event trajectory, historical cases, engineering experience, and emergency plans. After passing through the hierarchical aggregation mechanism of the memory enhancement link, the above information is injected into two levels: short-term memory and long-term memory. The short-term memory focuses on the immediate caching and fast access of the current round context to support the continuous construction and rolling update of the decision context within the same event. The long-term memory focuses on the persistent accumulation and structured archiving of historical successful and failed experiences to support strategy reuse and risk avoidance in cross-event scenarios. The two levels of memory work together through the above hierarchical aggregation architecture to form the complete information supply basis for the decision context generation unit when retrieving experience trajectories.

[0032] When extracting experience trajectories, the decision context generation unit uses the disaster characteristics, spatial topology characteristics, pedestrian and vehicle flow characteristics, and resource constraint characteristics of high-risk nodes in the current coupled state object as retrieval vectors. It performs similarity matching calculations in both short-term and long-term memory, extracting experience entries with matching scores exceeding a preset threshold in descending order of similarity. This yields historical handling experience trajectories most closely related to the current disaster situation, including successful action sequences and summaries of failure reasons from past events. The decision context generation unit then uses a preset structured prompt word template to concatenate the aforementioned state description text, rule constraints, and experience trajectories along vector dimensions and fuse fields to generate the decision context. The pre-designed structured prompt template is a multi-field text framework pre-designed by domain engineers based on the decision-making logic of urban flood disaster management. Its structure includes five areas: a state description area, used to fill in the state description text transformed from the current coupled state object; a rule constraint area, used to fill in engineering rule entries retrieved from the vertical domain knowledge base; an experience reference area, used to fill in historical handling experience trajectories extracted from long short-term memory; a target priority area, used to fill in the target ranking information for the current decision stage, such as placing personnel safety as the highest priority during the disaster phase; and an output format constraint area, used to constrain the model's output action description to include necessary fields such as the area of ​​effect, execution time, duration, resource requirements, and rollback conditions. Each area in the template has fixed field identifiers and placeholders, which are replaced by the current actual values ​​at runtime. The feature vectors of the state description text, the rule constraint entries, and the experience trajectory are concatenated along the vector dimension, and the structured fields of each area are merged and assembled according to the order defined in the template, ultimately generating a complete decision context text.

[0033] The large language model used in the decision context probabilistic sampling unit is a multi-layer encoder-decoder structure built on a transformer architecture. Its core components include a multi-head self-attention mechanism layer, a feedforward neural network layer, and a layer normalization module. The encoder transforms the input decision context text sequence into a high-dimensional feature vector word by word. Through multi-layer self-attention calculation, it captures the global dependencies between state descriptions, rule constraints, and experience trajectories, establishing a panoramic understanding feature representation of the urban disaster prevention scenario at the current moment. Based on the feature representation output by the encoder, the decoder generates corresponding disaster prevention operation step descriptions word by word through an autoregressive approach. In the pre-training stage, the weight matrix and bias parameters in the model are optimized by minimizing the next word prediction loss function on a large-scale general text corpus to obtain basic language understanding and generation capabilities. Subsequently, supervised fine-tuning is performed on labeled data in the field of urban flood disaster management. This labeled data contains disaster state descriptions and corresponding expert decision-making scheme pairs in historical flood events. By minimizing the domain task loss function, the weight parameters are further adjusted, enabling the model to acquire professional reasoning capabilities for disaster prevention and mitigation scenarios.

[0034] During the decoding and generation phase, the decision context probabilistic sampling unit employs a Monte Carlo sampling mechanism, performing multi-path parallel independent sampling on the word probability distribution output by the model. Specifically, the number of sampling branches is set to K, and in a specific setting, K=4, meaning four candidate solutions are generated in parallel. For the k-th sampling branch, the model is based on the input decision context. Candidate action sequences are generated step by step through autoregressive decoding. And calculate the joint generation probability of the candidate scheme sequence: ,in, This represents the candidate action sequence generated by the k-th sampling branch. This indicates the total length of the lexical units in the sequence. This represents the j-th word element in the sequence. Indicates in a given decision context and the previously generated The conditional generation probability of the j-th word under given conditions is calculated. A temperature parameter is introduced into the probability distribution to control the diversity of sampling; a larger value results in more dispersed sampling. In a specific setting, a value of 0.8 is used to ensure sufficient scheme differentiation while maintaining generation quality. Taking the aforementioned disaster scenario at a hub station as an example, the candidate action sequences generated by the four sampling branches may cover different control strategies, such as "immediately close the northwest entrance and exit and maintain one-way exit for 3 minutes before completely closing it; guide passengers in the west area of ​​the station hall to the central staircase for transfer, prioritizing the evacuation of vulnerable groups," "simultaneously close the northwest and west entrances and exits, divert all passengers to the southeast and northeast entrances and exits, suspend the platform escalator and initiate ground traffic control," "keep the northwest entrance and exit open but add flood barriers, only close the west gate group, some passengers in the central area of ​​the station hall remain waiting for subsequent transfer," and "stop all entry into the station and initiate tiered flow control outside the station, prioritizing the release of platform passengers to the east side of the station hall." The decision context probability sampling unit summarizes the output of all the above sampling branches to generate action sequences containing different solution space dimensions.

[0035] Figure 6 This is a schematic diagram illustrating the multi-branch candidate action tree expansion and trajectory classification based on the current coupling state in a vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, according to an embodiment of this application. Figure 6 As shown, with the current coupling state as the root node of the decision tree, the decision context probabilistic sampling unit expands four first-level candidate action branches (A, B, C, and D) from this root node through a Monte Carlo sampling mechanism. Each first-level branch corresponds to an independent candidate control policy direction. During the word-by-word generation process of autoregressive decoding, each first-level branch further splits into second-level sub-action nodes and third-level fine-grained action nodes, thus forming a hierarchical multi-branch action search tree that expands layer by layer from macro-policy direction to micro-execution steps. The number of branches at each level of the tree increases with the refinement of action granularity, fully characterizing the probabilistic sampling mechanism in different solution space dimensions. The parallel exploration path in the degree, where each leaf node represents the complete candidate action sequence finally generated by the branch, after the complexity identification and diversion deduction verification of the subsequent scheme screening module, the cross-interception verification and weighted scoring sorting of the resilience scoring result generation module, and the strategy determination and memory update of the scheme verification writing module, the trajectory of each leaf node in the action search tree is finally classified and marked as the optimal trajectory, the eliminated trajectory, and the alternative trajectory, respectively corresponding to the optimal control strategy finally output by the system, the infeasible schemes that are eliminated during the screening and verification process due to high matching with the failed cases or failure to pass the binary judgment, and the alternative scheme set retained for rapid switching in the event of sudden disaster changes.

[0036] The action sequence encapsulation unit performs key entity extraction and structured encapsulation operations. Each candidate solution in the action sequence exists in natural language text form, which contains descriptive modifiers and redundant dialogue characters, making it unsuitable for direct use in subsequent solution selection and resilience scoring. The action sequence encapsulation unit scans the natural language text of each candidate solution sentence by sentence, accurately stripping and extracting three types of key entities: the area of ​​action, i.e., the spatial object identifier to which the action is directed, such as the northwest entrance / exit, the west section of the station hall, or the central staircase; the execution time, i.e., the starting point of the action's execution time, such as immediate execution or execution in 3 minutes; and resource requirements, i.e., the type and quantity of disaster prevention resources required by the action, such as one mobile drainage pump and two station staff. The extraction of these key entities is achieved through named entity recognition, with recognition rules matching predefined field types in the standard action structure. The standard action structure is a predefined multi-field structured container. Its fields include at least: Action Type field, with values ​​ranging from lockdown, evacuation, diversion, or resource deployment; Scope field, recording the spatial node identifier to which the action is directed; Execution Time field, recording the start time of the action; Duration field, recording the expected duration of the action; Resource Requirement field, recording the type, quantity, and expected arrival time of the required resources; Constraints field, recording the preconditions on which the action depends; Priority field, recording the execution order of the action in the current plan; Fallback Plan field, recording the alternative measures when the action fails; and Evidence Source field, recording the rules or historical experience supporting the action.

[0037] The action sequence encapsulation unit maps each attribute extracted from the natural language text to the corresponding field of the aforementioned standard action structure, while filtering out meaningless and redundant modifiers. For each candidate solution generated by the sampling branch, a standard action structure is instantiated. To ensure the completeness of the evaluation, this stage can also include auxiliary handling solutions generated by other rule engines or threshold judgments within the same scrolling window, such as atomic actions like whether entry should be immediately stopped at the northwest entrance / exit. Finally, the action sequence encapsulation unit aggregates and assembles all instantiated standard action structure objects and the incorporated auxiliary handling solutions, packaging them into a more comprehensive set of candidate solutions that includes both macro-level strategies and micro-level actions.

[0038] Specifically, the scheme selection module 130 is used to identify the complexity of the candidate scheme set and perform triage and simulation verification to obtain a selected scheme set. It is understandable that the candidate scheme set output by the candidate scheme generation module 120 through the probabilistic sampling mechanism of the large language model contains multiple differentiated control schemes. These schemes exhibit significant differentiation in complexity: some schemes only involve threshold determination of a single node and local resource allocation, and feasibility determination can be completed solely based on static rules; other schemes involve complex elements such as multi-objective trade-offs, cross-domain cascading linkages, and dynamic evolution of disaster situations, requiring in-depth verification through historical case comparisons and external simulation tools. Performing simulations of equal depth on all candidate schemes would waste computational resources and prolong decision-making response time; relying solely on rule-based determination while ignoring the verification of complex schemes might overlook potential security risks. Therefore, the scheme selection module 130 is introduced.

[0039] In one exemplary embodiment of this application, the scheme screening module includes: a scheme parsing unit, used to perform scheme feature parsing and task complexity flow determination on each scheme in the candidate scheme set to obtain a direct verification subset and an auxiliary verification subset; a simulation subset generation unit, used to perform feature matching between each scheme in the auxiliary verification subset and historical disaster cases, and remove schemes that highly match the failure cases to obtain a simulation subset; and a subset merging unit, used to perform multi-branch simulation deduction on each scheme in the simulation subset, and after removing inefficient branches that expose safety risks, merge the effective branches with the direct verification subset to obtain a screened scheme set.

[0040] The scheme selection module 130 receives the candidate scheme set output by the candidate scheme generation module 120 as input. This candidate scheme set contains multiple structured and encapsulated standard action structure instances. Each instance includes fields such as action type, area of ​​effect, execution time, duration, resource requirements, constraints, priority, fallback plan, and evidence source. Taking the aforementioned hub station disaster scenario as an example, the candidate scheme set contains four candidate schemes: Scheme 1 is to immediately close the northwest entrance / exit function and maintain one-way exit for three minutes before completely closing it; passengers in the western area of ​​the station hall are guided to the central staircase for transfer, with priority given to evacuating vulnerable groups; Scheme 2 is to simultaneously close both the northwest and western entrances / exits, diverting all passengers to the southeast and northeast entrances / exits and suspending the platform's upward escalators; Scheme 3 is to keep the northwest entrance / exit open but add flood barriers, closing only the western turnstiles, with some passengers in the central area of ​​the station hall remaining and waiting for further transfer; Scheme 4 is to completely stop entry into the station and initiate tiered flow control outside the station, with platform passengers prioritized for release to the eastern station hall.

[0041] Example 1 of the scheme analysis unit: The scheme analysis unit extracts the scope field, resource requirement field, and expected target field from the standard action structure of each candidate scheme, and performs complexity identification logic on the scheme based on three complexity judgment dimensions. The first dimension is the multi-objective trade-off dimension, used to evaluate whether the scheme involves a conflict between the objectives of maintaining traffic efficiency and ensuring evacuation safety. The second dimension is the cross-domain linkage dimension, used to evaluate whether the scope of the scheme spans the coupling edge between ground traffic nodes and underground space nodes, that is, whether the execution of the scheme will trigger cross-domain cascading effects. The third dimension is the disaster dynamic evolution dimension, used to evaluate whether the effectiveness of the scheme depends on the future trends of water depth, flow velocity, or population density. For each dimension, the scheme analysis unit performs binary judgment based on the field content in the standard action structure. If the scheme does not involve any of the three dimensions, it is judged as a low-complexity scheme; if it involves any of the three dimensions, it is judged as a high-complexity scheme.

[0042] Taking the aforementioned four candidate solutions as examples, Solution 1 involves the redistribution of passenger flow in the station hall after the closure of the northwest entrance / exit, as well as the trade-off between prioritizing the evacuation of vulnerable groups and overall traffic efficiency. Furthermore, the effectiveness of its three-minute delayed closure strategy depends on the future evolution of the water depth, thus it is classified as a high-complexity solution. Solution 2 involves the interconnected impact of simultaneously closing both entrances / exits on ground traffic and underground passenger flow, and is also classified as a high-complexity solution. Solution 3 involves the risk decision of keeping high-risk entrances / exits open and arranging for personnel to remain, and is classified as a high-complexity solution. Solution 4 involves the chain reaction of the entire station's cessation of entry on external ground traffic and bus connections, and is also classified as a high-complexity solution. In this scenario, the solution parsing unit may also receive auxiliary handling solutions within the same scrolling time window, such as "whether the northwest entrance / exit should immediately stop entry," which only involve the determination of a single node threshold. These solutions can be determined solely by the water accumulation control threshold rules, and are therefore classified into the direct verification subset. Solutions 1 to 4 are all classified into the auxiliary verification subset.

[0043] The simulation subset generation unit performs feature matching. Historical disaster cases are a collection of case records extracted from post-disaster investigation reports, operational logs, and expert debriefing records of past urban flooding events, and stored in a vertical domain knowledge base after being structured and coded. Each case record contains four types of feature vectors: disaster feature vector, recording the water depth distribution, flow velocity range, inundation duration, and risk level of the historical event; spatial topology feature vector, recording the number of nodes, connectivity, entrance / exit locations, and channel layout of the affected area; pedestrian and vehicle flow feature vector, recording the population density, number of stranded individuals, proportion of vulnerable groups, and traffic congestion during the event; and resource constraint feature vector, recording the number of available drainage equipment, personnel configuration, and response time. Each case record is also labeled with a handling result tag, including success, partial success, or failure, along with the corresponding reasons for failure and applicable conditions. The simulation subset generation unit extracts the disaster characteristics, spatial topology characteristics, pedestrian and vehicle flow characteristics, and resource constraint characteristics of the current event in the coupled state object for each scheme in the auxiliary verification subset. These features are then compared with the similar feature vectors of historical disaster cases in the vertical domain knowledge base using cosine similarity calculations. A comprehensive matching value is calculated by introducing weight parameters. The formula for calculating this comprehensive matching value is as follows: in, This represents the overall matching value between the current event and the j-th historical case. and Let these represent the disaster feature vectors of the current event and the j-th historical case, respectively. and These represent spatial topological feature vectors, and These represent the feature vectors of pedestrian and vehicle traffic, respectively. and These represent resource constraint feature vectors, This function represents the cosine similarity calculation between two vectors. , , , The weight parameters are and satisfy the following conditions: The values ​​of the aforementioned weighting parameters are determined by domain engineers based on the degree of influence of various characteristics on disaster response effectiveness, and are selected within a specific setting. =0.30、 =0.25、 =0.25、 =0.20, where disaster characteristics have the highest weight, reflecting the dominant influence of hydrodynamic parameters such as water depth and flow velocity on the feasibility of the scheme. The simulation subset generation unit performs low-cost initial screening based on the above comprehensive matching value. For each scheme in the auxiliary verification subset, all historical cases marked as failed in the vertical domain knowledge base are traversed. If the comprehensive matching value of the scheme with a certain failed case is... If the circuit breaker threshold is exceeded, such as 0.85, the solution is determined to be highly consistent with known failure modes, and is directly marked as infeasible and eliminated.

[0044] Taking the aforementioned hub station scenario as an example, the core strategy of Scheme 3 is to keep high-risk entrances and exits open and arrange for personnel to remain there temporarily. The vertical domain knowledge base records a historical case marked as failed. In this case, a subway station remained open when the water depth outside the entrances and exits reached 0.28 meters, and station staff were arranged to wait on-site. Ultimately, due to insufficient water-blocking measures and excessively long waiting times, flooding intensified and personnel evacuation was severely delayed. The comprehensive matching value of Scheme 3 with this failed case in terms of disaster characteristics, spatial topology characteristics, and handling logic is 0.91, exceeding the circuit breaker threshold of 0.85. Therefore, Scheme 3 is directly eliminated. Schemes 1, 2, and 4 did not have a comprehensive matching value exceeding the circuit breaker threshold with any failed case, and therefore were retained and output as a subset to be simulated.

[0045] Example 2 of the scheme analysis unit Specifically, in the resilience management scenario of the transportation-underground space coupled system under urban flooding disasters, various emergency response actions, such as road closures, subway entrance and exit shutdowns, crowd evacuation, and resource allocation, are not executed independently in isolated vacuum environments. On the contrary, they work together on the same complex physical network highly interwoven with elements such as road nodes, underground passages, commercial facilities, and public transportation hubs. The scheme evaluation paradigm in Example 1 examines only the characteristics of each scheme in isolation, such as whether it involves multi-objective trade-offs between safety and efficiency, or whether it is affected by rapidly changing disaster conditions, thereby determining its complexity. This independent evaluation paradigm implies a key but fragile assumption: that the complexity of each scheme is internalized and closed, unaffected by other coexisting schemes. However, in real emergency scenarios, there are profound coupling relationships between schemes, including competition for limited emergency resources, overlapping interference in physical space, and even strict dependence on the execution order. Among these, the most potentially dangerous is the cascading connectivity disruption effect: localized containment measures that appear safe in two independent assessments may unexpectedly sever the only life-saving evacuation route in the entire area when implemented together, creating a catastrophic emergent high risk. This fatal safety hazard caused by combined risks is completely invisible from the perspective of traditional independent assessments and is easily mistakenly categorized into a low-cost direct verification path, thus bypassing the necessary in-depth simulation verification process and posing a huge hidden danger to the engineering reliability of the final control strategy.

[0046] Therefore, the core purpose of introducing an explicit quantification mechanism for determining the flow of task complexity between schemes is to overcome the limitations of the aforementioned independent evaluation paradigm. By constructing a computational framework that can accurately identify and quantify the coupling relationship between schemes, it reveals and amplifies the potential risks caused by the combined execution of multiple schemes. This ensures that even those seemingly simple but actually highly coupled dangerous schemes can be accurately identified and forcibly routed to a high-precision verification path that includes historical case matching and external simulation, thereby effectively intercepting combined risks.

[0047] In another exemplary embodiment of this application, the scheme parsing unit is used for: The inherent complexity vector is obtained by quantifying the inherent complexity of the candidate solution set. The specific implementation process of this decision-making mechanism involves first quantifying the inherent complexity of the candidate solution set to obtain an inherent complexity vector. This step aims to isolate the interactions between solutions, assigning only an initial scalar value to the inherent complexity of each solution itself. The implementation method involves extracting and evaluating features for each solution from three dimensions: multi-objective trade-offs, cross-system linkage, and disaster dynamic evolution. These features are then normalized using a weighted summation method and the Sigmoid function to generate a continuous scalar value between zero and one. The calculation formula is expressed as follows: in, For the plan The inherent complexity value; This is a Sigmoid normalization function to ensure a uniform output value range; , , The inherent complexity of the scheme in balancing conflicting objectives, impacts across system boundaries, and adapting to changes in disaster situation is represented by the numerical values ​​obtained after quantitatively evaluating the characteristics of the scheme in terms of conflicting objectives, cross-domain impacts, and dynamic adaptability based on preset rules and scoring models. , , The corresponding weight coefficients, assigned by domain experts based on historical disaster data, are 0.3, 0.4, and 0.3, respectively. Since the Sigmoid function outputs 0.5 when the input is zero, for solutions with complexity far below the baseline level, the scores for all three dimensions are significantly negative. Linearly combining these values ​​and inputting them into the Sigmoid function yields an output approaching zero, thus allowing the inherent complexity value to cover the low-end range close to zero. Conversely, for solutions involving complex cross-domain linkages and multi-objective trade-offs, the linear combination produces a larger positive value, corresponding to a Sigmoid output approaching one. This transforms qualitative solution characteristics into computable continuous scalars, laying the foundation for subsequent quantitative fusion.

[0048] The coupling coefficients of candidate solutions are calculated to obtain the solution coupling matrix. It should be understood that the true risk of a solution does not originate solely from itself, but rather a greater extent from its interactions with other solutions. This step aims to capture and quantify the coupling strength between any two solutions across four dimensions: resource, spatial, temporal, and network connectivity. The implementation involves, for each pair of solutions... The overlap in emergency resource requirements, the intersection in physical spatial scope, and the dependence in execution order of the two schemes are calculated and then summed using a weighted average. Based on this, a multiplicative structure is introduced as a risk amplifier, driven by the cascaded connectivity loss rate, to measure the degree of damage to the global evacuation network when the two schemes are executed jointly. The formula for calculating this coupling coefficient is as follows: in, For the plan and The coupling coefficient between them; , and , These are the resource requirements set and spatial scope of the proposed solution, respectively. It is a time-dependent index; , , The weights for each basic coupling element, It is a function used to calculate the geometric area or volume of a given two-dimensional or three-dimensional spatial region (defined by Z). The multiplicative structure in the above formula... It carries the core engineering semantics: the linear weighted sum of the basic coupling constitutes the base metric for the mutual influence between the two schemes, while the cascaded connectivity loss rate It then acts as a risk amplifier on this foundation.

[0049] In urban flooding scenarios, the combined implementation of two lockdown measures may sever several previously unobstructed emergency evacuation routes within the coupled network, leaving stranded individuals without accessible exits. This emergent combined risk far exceeds the sum of the effects of the two measures acting independently. If the combined implementation does not disrupt any evacuation routes, then... When the value is zero, the amplifier is ineffective, retaining only the three basic couplings: resource contention, spatial overlay, and temporal dependency; however, when joint execution causes a large number of evacuation paths to be blocked... As the value approaches 1, the amplifier will significantly increase the coupling coefficient of the scheme pair, forcing it to be effectively identified in subsequent steps. The calculation formula is as follows: , here This represents the total set of available evacuation paths in the network without intervention. This represents the set of available paths remaining after the two schemes are executed together. This ratio accurately characterizes the destructive effect of the combined measures on global connectivity. As a cascade amplification factor, used to nonlinearly enhance the impact of the destructive effect, it is a hyperparameter, such as 2.0, determined through retrospective simulation of historical disaster cases or stress testing in a simulated environment. It should be noted that when both the basic coupling term and the cascade connectivity loss rate take large values... The calculated result may exceed the [0,1] interval, which is mathematically permissible because the coupling coefficient here serves as a relative strength weight in the coupling matrix of the schemes, rather than a normalized probability. Its absolute value reflects the relative magnitude of the combination risk between scheme pairs. In the subsequent weighted aggregation operation to enhance coupling complexity, both the numerator and denominator contain... This allows the establishment of an N×N coupling matrix, where each element... All of them precisely quantified the combined risks of a pair of options, especially assigning extremely high coefficients to high-risk combinations that could cut off critical escape routes.

[0050] The coupling enhancement complexity vector is obtained by performing graph-weighted propagation-based calculation on the inherent complexity vector and the scheme coupling matrix. Previous steps have obtained the node information (inherent complexity) and the edge information (coupling coefficient) between schemes. Therefore, it is necessary to effectively fuse these two types of information so that the evaluation of the complexity of a single scheme can be aware of the global impact from the entire scheme network. The implementation method is to construct an undirected graph with schemes as nodes and coupling coefficients as edge weights, and perform neighborhood-weighted aggregation operations on the graph. The final complexity of each scheme is obtained by fusing its own inherent complexity with the complexity of all other neighboring schemes through weighted propagation via coupling edge weights. The calculation formula is as follows: in, For the plan Final complexity after information augmentation; As its own weight, it controls the fusion ratio of its own information and the information propagated from the neighboring domain. It is an empirical value pre-set by domain experts based on the differences in the importance attached to the complexity of the solution itself and the coupling effect between solutions at specific disaster stages (such as before, during, and after the disaster). The value range is (0,1), such as 0.1. To prevent division by a constant, This represents the total number of candidate solutions. The core mechanism of this formula lies in the neighborhood weighted average term: when a solution has a large coupling coefficient with one or more solutions with high inherent complexity, the high-complexity neighbor occupies a larger weight in the weighted average, thereby passing on the higher complexity information to the current solution, thus enhancing its complexity. Significantly higher than its inherent complexity Such an inherent complexity A very low-complexity solution, if it has a strong coupling relationship with one or more high-complexity solutions, i.e. The value is very large, resulting in increased complexity. This will be significantly enhanced, thereby exposing its hidden combined risks.

[0051] An adaptive routing process based on coupling density is applied to the coupling enhancement complexity vector and the scheme coupling matrix to obtain a direct verification subset and an auxiliary verification subset. This step aims to establish a dynamic decision threshold that adapts to the overall chaos of the current disaster situation, thereby achieving intelligent routing of schemes. Its implementation first involves calculating the global coupling density based on the coupling matrix. , This density is used to macroscopically characterize the overall strength of mutual checks and balances and conflicts among all current candidate solutions. Subsequently, a baseline routing threshold determined by historical tuning is applied based on this density. Perform dynamic corrections and generate adaptive routing thresholds. ,in To adjust the density sensitivity, The baseline routing threshold determined by historical optimization is, for example, 0.50. When the global coupling density... A higher threshold – meaning there is a strong mutual constraint relationship among the various solutions in the current candidate solution space – is the threshold. The proactive elevation of the threshold causes the increased coupling complexity of more solutions to fall below the threshold range. Solutions that might otherwise be pushed into the simulation verification path are appropriately pulled back to the direct verification path. This avoids the time escalation of disaster emergency response caused by excessively pushing solutions into high-cost simulation stages, while ensuring safety. When the global coupling density is low—such as when solutions are independent in the pre-disaster prevention phase—the threshold is close to the baseline value. The flow diversion behavior degenerates into a classic decision pattern based on inherent complexity, maintaining compatibility with the first embodiment.

[0052] Finally, increasing the coupling between each solution increases complexity. With this dynamic threshold Compare and perform a binary route determination: Solutions below the threshold are considered to have manageable risk and are included in the low-cost direct verification subset; while Schemes that exceed or equal the threshold are considered high-risk or have significant potential coupling and must be included in a subset requiring further verification, which includes in-depth validation. This approach, while ensuring safety, avoids indiscriminately pushing all schemes into high-cost simulation stages, thus improving the efficiency and robustness of the overall decision-making process.

[0053] Taking a scenario of backflow during a rainstorm at a rail transit hub as an example, consider two candidate solutions: Solution 1 (delaying the closure of the northwest entrance and guiding evacuation) and Solution 5 (closing section K of the main ground road on the north side of the hub). Solution 1 involves timing control, cross-regional passenger flow management, and multi-objective trade-offs. All three dimensions have relatively large positive scores, and the inherent complexity is obtained after linear combination and Sigmoid mapping. =0.78. Option 5, however, involves only a single traffic control action, does not involve cross-system coordination, has no conflicting objectives, and has extremely low dependence on the evolution of the disaster. Its scores across all three dimensions are significantly negative, resulting in a large negative value after linear combination. The inherent complexity is obtained after Sigmoid mapping. =0.15. Under the independent evaluation paradigm, Scheme 5 will be directly fed into the verification path. However, in calculating the coupling coefficient... It was discovered that although the two have almost no overlap in resources and space, the K section is the only main road for personnel to evacuate from the area after other exits of the hub station (such as the northeast exit).

[0054] When Scheme 1 and Scheme 5 are implemented together, it will result in a cascading connectivity loss rate in this region. Up to 0.90. In With a setting of 2.0, this makes the coupling coefficient... This is amplified to a higher value, such as 0.85. When calculating the coupling enhancement complexity, for scheme five, although its own... =0.15 is very low, but due to the inherent complexity of Scheme 1 =0.78 is relatively high and the coupling coefficient between the two is also high. The value of 0.85 is relatively large. The neighborhood weighted average term transfers the high-complexity information of Scheme 1 to Scheme 5 with a large weight, thus increasing its complexity. Significantly higher than its inherent complexity, calculated =0.62. Meanwhile, due to this high-strength coupling, the global coupling density... This also increases the adaptive routing threshold. From the benchmark Appropriately improve. There is a high-strength coupling between Scheme 1 and Scheme 5, resulting in a high global coupling density. It was pulled up, Adaptive routing threshold with appropriate settings From the benchmark value =Increased from 0.50 to 0.58. The increased complexity of Scheme 5. =0.62 higher than =0.58, therefore it was correctly routed to the subset to be assisted in the verification process. The potential fatal risk of blocking the only evacuation main road when it was executed in conjunction with Scheme 1 was carefully evaluated by in-depth verification process.

[0055] The subset merging unit receives the subset to be simulated and establishes a virtual simulation branch for each scheme. It automatically calls external professional computing tools as needed for multi-branch simulation. Example 1 above will be used as an example. The external professional computing tools include a flood evolution simulation tool and a personnel evacuation simulation tool. The flood evolution simulation tool predicts the water depth and flow velocity changes at each node within a few minutes after a specific control action is implemented. The personnel evacuation simulation tool predicts the peak passenger flow density and evacuation completion time at each passage and entrance after a specific evacuation scheme is implemented. The flood evolution simulation tool is built on a coupled hydrodynamic model architecture of urban surface-underground space and developed using mature hydrodynamic numerical calculation methods in the field of urban disaster prevention engineering. The personnel evacuation simulation tool is developed based on a fusion architecture of a multi-agent model and a social force model, and trained and optimized using practical data from urban rail transit personnel evacuation engineering. The subset merging unit selectively calls the tools according to the characteristics of the scheme.

[0056] Taking Scheme 1 as an example, the safety of its three-minute delayed closure strategy depends on whether the water depth at the northwest entrance / exit will exceed the danger threshold within that time window. Therefore, the subset merging unit calls a flood evolution simulation tool to predict the water depth change at the northwest entrance / exit node within the next three minutes, and simultaneously calls a personnel evacuation simulation tool to predict the density change trend of passengers in the western area of ​​the station hall after being transferred via the central staircase. Schemes 2 and 4 have higher certainty in terms of safety because they directly close high-risk entrances / exits. The subset merging unit can complete the auxiliary judgment through vertical domain knowledge base retrieval and empirical reasoning, without calling high-cost simulation tools. The subset merging unit receives objective numerical results from external computing tools and performs safety and resource feasibility assessments on each simulation branch. If the simulation results of a certain simulation branch show that there are safety risks such as the water depth at a node exceeding the safety threshold, the passenger flow density in the passage exceeding the congestion limit, or emergency resources failing to arrive within the specified time during the implementation of the scheme, then the branch is marked as an ineffective branch and is eliminated.

[0057] Taking Scheme 1 as an example, the output of the flood evolution simulation tool shows that the water depth at the northwest entrance node will rise from 0.32 meters to 0.38 meters within the next three minutes, which does not exceed the danger threshold of 0.50 meters. The output of the personnel evacuation simulation tool shows that the peak density of the central staircase transfer path is 4.2 people per square meter, and it can be reduced to below 3.0 people per square meter within eight minutes. Neither of the above results exposes any safety risks. Therefore, the derivation branch of Scheme 1 is retained as a valid branch. After all derivation branches have been verified, the subset merging unit will perform data format alignment and merging with the directly verified subset output by the scheme analysis unit through the verified valid branches, and encapsulate each scheme into a standard action structure format, finally outputting the filtered scheme set.

[0058] Specifically, the resilience scoring result generation module 140 is used to perform cross-interception verification on the selected solution set in sequence based on rule consistency, state consistency, security, and goal achievement, and to determine the resilience score result based on a weighted score and ranking of the solutions using multi-dimensional resilience indicators. It should be understood that the selected solution set output by the solution selection module 130 includes multiple candidate solutions retained after complexity triage, historical case matching, and simulation verification. Although these solutions have passed the initial screening, they have not yet undergone item-by-item compliance verification under a unified verification framework regarding engineering rule consistency, current physical state consistency, secondary security, and goal achievement. At the same time, different candidate solutions have their own advantages and disadvantages in multiple dimensions such as personnel safety, evacuation efficiency, access network connectivity, vulnerable group protection, resource consumption, and secondary risks. A comprehensive comparison and objective ranking of solutions cannot be achieved solely based on a single indicator from simulation. Therefore, the resilience scoring result generation module 140 is introduced.

[0059] In one exemplary embodiment of this application, the resilience scoring result generation module includes: a scheme binary determination unit, used to perform binary determination of rule consistency, state consistency, security and goal achievement of each scheme in the selected scheme set in sequence through a unified checker to obtain a set of feasible schemes; and a resilience assessment determination unit, used to perform multi-dimensional resilience weighted assessment and ranking determination on the set of feasible schemes to obtain a resilience scoring result.

[0060] Taking the aforementioned disaster scenario at the hub station as an example, the selected solutions include three candidate solutions that have been verified by simulation or rules: Solution 1 is to immediately close the northwest entrance and exit, maintain one-way exit for three minutes, and then completely close it, guiding passengers in the western area of ​​the station hall to the central staircase for transfer, and prioritizing the evacuation of vulnerable groups; Solution 2 is to simultaneously close the northwest and west entrances and exits, divert all passengers to the southeast and northeast entrances and exits, and suspend the platform's upward escalator; Solution 4 is to stop all entry into the station and activate the external layered interception, prioritizing the release of platform passengers to the east side of the station hall.

[0061] The binary decision unit performs the judgment. The unified checker is a layer-by-layer verification engine based on rules and state constraints. Its architecture includes four cascaded verification levels. Each level performs a binary judgment, and the output value is either 1 for pass or 0 for fail. The first level is rule consistency verification, which is used to determine whether the action parameters of the candidate scheme violate the engineering safety hard thresholds recorded in the vertical domain knowledge base, including water accumulation control thresholds, lower limits of traffic capacity, resource scheduling time limits, and evacuation priority constraints. The second level is state consistency verification, which is used to determine whether the area of ​​action, facility status, and resource occupancy pointed to by each action in the candidate scheme are consistent with the real-time physical status recorded in the current coupled state object, including the actual opening and closing status of entrances and exits, the passable direction of stairs, the actual available number of drainage equipment, and the expected arrival time. The third level is safety verification, used to determine whether the candidate plan will cause secondary casualties during implementation. This includes whether people are guided to high-risk areas where water depth is continuously rising, whether additional passenger flow is added to passages where density is already close to the congestion limit, and whether the only exit of a vulnerable group's area is closed before the vulnerable group has been evacuated. The fourth level is target achievement verification, used to determine whether the candidate plan can achieve the preset disaster prevention targets within the specified rolling decision-making window. This includes whether the reduction in risk values ​​at high-risk nodes meets expectations, whether the evacuation completion rate of stranded people reaches the phased target, and whether the capacity restoration rate of key passages meets the minimum requirements.

[0062] The scheme binary determination unit performs the above four binary checks sequentially for each candidate scheme in the selected scheme set. Taking Scheme 1 as an example, at the rule consistency level, the three-minute delay closure strategy of Scheme 1 needs to be compared with the water accumulation control threshold rule. The flood evolution simulation results of the scheme selection module show that the water depth at the northwest entrance node will rise from 0.32 meters to 0.38 meters in the next three minutes, which does not exceed the danger threshold of 0.50 meters. Therefore, the rule consistency check result is... The value is 1; at the state consistency level, the northwest entrance / exit in the current coupled state object is open and the central staircase has remaining passage capacity, which is consistent with the action direction of Scheme 1. Therefore... The value is set to 1; at the safety level, the personnel evacuation simulation results show that the peak density of the transfer path on the central staircase is 4.2 people per square meter, and it can be reduced to below 3.0 people per square meter within eight minutes. There is no action to guide personnel to the rising water depth area, therefore... The value is 1; at the target achievement level, Option 1 can reduce the risk value of the northwest entrance / exit from 0.546 to below 0.30 within a 30-minute rolling window, and the stranded people in the west area of ​​the station hall can be transferred within 15 minutes. The value is 1.

[0063] The scheme two-value determination unit calculates the overall executable judgment value of the scheme using a logical multiplication method, and the formula is as follows: in, This represents the overall executable value of the i-th candidate solution. This indicates the result of the rule consistency check. This indicates the state consistency check result. This indicates the security verification result. This represents the result of the goal achievement verification. All four verification results are binary: 1 for passing and 0 for failing. The formula uses a logical multiplication structure; failure of any one verification result in... =0, thus achieving hard truncation filtering. The executable judgment value for Scheme 1 is... After undergoing the same four checks, the executable judgment values ​​for Scheme 2 and Scheme 4 are respectively... =1 and =1. The scheme binary determination unit packages all schemes with an executable judgment value of 1 and outputs them as a feasible scheme set.

[0064] The resilience assessment unit performs multi-dimensional resilience weighted assessment and ranking. For each feasible solution, the resilience assessment unit comprehensively extracts a multi-dimensional evaluation attribute matrix from the simulation results in the solution selection module and the attributes of the coupled state objects. This multi-dimensional evaluation attribute matrix includes five positive evaluation dimensions and three negative penalty dimensions. The five positive evaluation dimensions are: personnel safety score. This is used to measure the degree of reduction in exposure risk for people in high-risk areas after the implementation of the plan; evacuation and traffic efficiency score. This is used to measure the evacuation rate of stranded people and the recovery rate of critical passageways after the implementation of the plan; network connectivity maintenance score. This is used to measure the number and proportion of entrances and exits that allow for normal passage between surface transportation and underground spaces after the implementation of the plan; and the fairness protection score for vulnerable groups. This is used to measure whether the plan prioritizes evacuation measures for the elderly, children, and people with mobility impairments; critical facility resilience score. This is used to measure the estimated time it will take for affected facilities to return to normal operation after the plan is implemented. The three negative penalty dimensions are: emergency resource consumption. This is used to measure the ratio between the total amount of emergency resources (such as drainage pumps, station personnel, and ground traffic controllers) used during the implementation of a plan and the total amount available; secondary risk exposure penalty value. This is used to measure whether there are secondary risks during the implementation of the plan, such as guiding people to areas with continuously rising water depth or adding passenger flow to passages that are already close to their congestion limits; and the expected value of the risk of failure again after recovery. This is used to measure the probability that high-risk nodes will trigger lockdowns or fail again within a subsequent rolling time window after the plan is implemented. The scores for each dimension are normalized to between 0 and 1. Higher values ​​for positive dimensions indicate better performance in that dimension, while higher values ​​for negative dimensions indicate greater cost or risk. The scores for each dimension are obtained as follows: Personnel safety score and evacuation efficiency score are extracted from the passenger flow density change curve, evacuation completion time, and risk value reduction feedback from the simulation tool in the plan selection module; network connectivity maintenance score is calculated from the proportion of remaining passable entrances / exits and paths in the coupled state objects; the vulnerability group fairness protection score is assigned based on whether the standard action structure of the plan includes priority evacuation actions for vulnerable groups; the critical facility recovery capability score is estimated from the historical recovery time data of similar facilities in the vertical domain knowledge base; resource consumption is calculated from the resource demand field of the standard action structure and the emergency resource status sub-object of the coupled state objects; and the secondary risk exposure penalty value and the expected risk of recurrence are extracted from the risk evolution prediction of subsequent periods feedback from the simulation tool.

[0065] The resilience assessment unit adaptively adjusts the weighting coefficients of each dimension based on the current disaster prevention stage. This implementation scenario is in the mid-disaster phase, where personnel safety and evacuation efficiency are the primary objectives. Therefore, the weights for personnel safety and evacuation efficiency scores should be higher than other dimensions. The specific values ​​of the weighting coefficients are pre-determined by domain engineers based on the differences in objective priorities among the pre-disaster, mid-disaster, and post-disaster phases, and are selected within a specific setting for the mid-disaster phase. =0.24、 =0.18、 =0.14、 =0.10、 =0.08、 =0.10、 =0.10、 =0.06. The resilience assessment unit calculates the overall resilience score for each feasible option using a weighted scoring formula, which is: in, Let represent the overall resilience score of the i-th feasible solution. to The weighting coefficients for each dimension are calculated by adding positive dimensions to the total score and subtracting negative dimensions from the total score, so that the overall resilience score reflects both the positive benefits and negative costs of the solution.

[0066] Taking the three feasible solutions in the aforementioned disaster scenario at the hub station as examples, the resilience assessment unit extracts the multi-dimensional evaluation attributes of each solution and calculates the comprehensive resilience score. Solution 1's delayed closure and guided transfer strategy effectively reduces the risk of personnel exposure at the northwest entrance / exit in terms of safety, resulting in a high personnel safety score. The value is 0.88; the passenger density in the west area of ​​the station hall can be reduced to a safe range within eight minutes after being transferred via the central staircase, resulting in an evacuation efficiency score. The value is 0.82; with only one entrance / exit closed and the other three entrances / exits remaining operational, the network connectivity score is maintained. The value is 0.75; the plan clearly stipulates that vulnerable groups should be given priority in evacuation, and the fairness protection score is [not specified]. The value is 0.90; the northwest entrance / exit is expected to reopen to traffic within 30 minutes after the drainage pumps are in place, hence the recovery capability score. The value was set at 0.70; in terms of resource consumption, only one mobile drainage pump and a small number of station staff were used. The value was set to 0.35; the simulation results did not show any secondary risks. The value is 0.15; the probability of the window being blocked again during subsequent scrolling is low. The value is 0.20. Substituting this into the above formula, we get... =0.5478. The same calculations were performed on Schemes Two and Four, yielding the same result. =0.4452, =0.3730. The resilience assessment unit sorted the three feasible solutions in descending order based on their comprehensive resilience scores, resulting in the following ranking: Solution 1 (0.5478) ranked first, Solution 2 (0.4452) ranked second, and Solution 4 (0.3730) ranked third. The resilience assessment unit then packaged the ranking results, along with the detailed scores for each dimension, weighting, and ranking labels for each solution, into a resilience score.

[0067] Specifically, the scheme verification and writing module 150 is used to determine strategies and update memory based on resilience score results. It outputs the scheme with the highest score that meets the constraints as the optimal control strategy and retains a set of alternative schemes. Simultaneously, it writes the current round of decision-making trajectory and verification conclusions into long short-term memory after structured summarization. In other words, the evolution of urban flood disasters is characterized by both persistence and abrupt changes. During the rolling decision-making process, the ranking results generated by each round of resilience scores need to be transformed into control strategies that can be implemented in practice. At the same time, a set of alternative schemes needs to be retained to cope with sudden deterioration of the disaster situation or changes in resource conditions in subsequent time windows. Meanwhile, if the status information, scheme selection, verification conclusions, and execution effects of each round of decisions are discarded after use in that round, subsequent rolling time windows cannot obtain the contextual information of previous decisions, and experience reuse across events is impossible, making it difficult to support integrated and continuous control of pre-disaster prevention, in-disaster response, and post-disaster recovery. Therefore, the scheme verification and writing module 150 is introduced to provide real-time contextual support for continuous rolling simulations within the same event, and to provide persistent knowledge accumulation for experience transfer and strategy reuse across events.

[0068] In one exemplary embodiment of this application, the scheme verification and writing module includes: a scheme splicing unit, used to extract the top-ranked scheme output that meets resource constraints as the optimal control strategy based on the descending order of resilience score results, extract the remaining scheme outputs that meet the safety threshold as a candidate scheme set, and splice the above scheme metadata to generate a scheme result set; a key-value association binding unit, used to perform key-value association binding with coupled state objects to obtain structured event slices, and push them to a short-term cache with a rolling window elimination mechanism to obtain short-term memory; and a writing unit, used to calculate the long-term retention priority of each experience item in the short-term memory according to reuse frequency, effect improvement, applicable scope and time decay, and write the experience items exceeding the threshold into long-term memory after structured summarization.

[0069] The scheme splicing unit, based on the descending order of resilience scores, first extracts the top-ranked scheme and determines whether it meets the current hard resource constraints. Hard resource constraints refer to the rigid boundary conditions formed by the currently available resource quantity, estimated arrival time, and resource occupancy status recorded in the emergency resource status sub-object of the coupled state object. The scheme splicing unit compares each resource type and quantity listed in the resource requirement field of the standard action structure of the top-ranked scheme with the currently available quantity in the emergency resource status sub-object of the coupled state object. Taking Scheme 1 as an example, its resource requirements are one mobile drainage pump expected to arrive within six minutes, two station staff to guide passengers in the west area of ​​the station hall, and one emergency broadcast system to prioritize the evacuation of vulnerable groups. The coupled state object records one currently available mobile drainage pump with an estimated arrival time of six minutes, three station staff on duty in the station hall, and an emergency broadcast system that is available. All of these resource requirements are within the range of currently available resources; therefore, Scheme 1 meets the hard resource constraints. The scheme assembly unit then transforms the standard action structure of Scheme 1 into an optimal control strategy output that can be executed by external hardware and personnel scheduling. This optimal control strategy includes all executable fields such as action type, area of ​​effect, execution time, duration, resource allocation instructions, vulnerable group priority identifier, and rollback conditions. After extracting the optimal control strategy, the scheme assembly unit determines whether each of the second-best and subsequent schemes meets the basic safety threshold. The basic safety threshold refers to the scheme's passing result in the binary judgment dimension of safety verification and its personnel safety score not lower than the preset minimum safety baseline value, such as 0.60. Taking Scheme 2 as an example, its personnel safety score is 0.92 and its safety verification result is passing, meeting the basic safety threshold, so it is extracted and loaded into the candidate scheme reserve pool; Scheme 4's personnel safety score is 0.95 and its safety verification result is passing, also meeting the basic safety threshold, and it is also loaded into the candidate scheme reserve pool. The scheme assembly unit packages all schemes in the reserve pool of alternative schemes into a set of alternative schemes, which is used for rapid switching in the event of sudden disaster changes. Subsequently, the scheme assembly unit merges the optimal control strategy with the metadata of the alternative scheme set to generate a scheme result set. Taking this scenario as an example, the scheme result set contains three records: Scheme 1 is marked as optimal, Scheme 2 is marked as alternative 1, and Scheme 4 is marked as alternative 2.

[0070] The key-value association binding unit performs key-value association binding to generate structured event slices. The specific method of key-value association binding is as follows: a unique event slice key is generated by combining the timestamp of the current decision round with the event stage identifier. All sub-objects in the coupled state object (including ground traffic state sub-objects, underground space state sub-objects, flood risk state sub-objects, pedestrian and vehicle flow state sub-objects, emergency resource state sub-objects, and cross-domain coupling edge sets) are used as the state-side values. All records in the solution result set, the simulation tool logs called in this round (including the node water depth prediction values ​​returned by the flood evolution simulation tool and the density change curves returned by the personnel evacuation simulation tool), and the weight configuration and ranking position in the resilience score results are used as the decision-side values. The above state-side and decision-side values ​​are mapped and bound through the event slice key. The bound full-process context data is packaged and encapsulated into a structured event slice. The key-value association binding unit then pushes the above structured event slice to the short-term buffer to generate short-term memory. Short-term memory is a rolling buffer oriented towards the current event, and its architecture is a queue structure with fixed capacity and a first-in-first-out (FIFO) elimination mechanism. The queue's capacity is preset by domain engineers based on the typical duration of a single flood event and the frequency of rolling decisions. In a specific setting, the queue capacity is set at twenty event slices, meaning it retains all context information from the most recent twenty rounds of decisions in the current event. When a new event slice is pushed to the tail of the queue, if the queue has reached its capacity limit, the oldest event slice at the head of the queue is automatically discarded. Through this rolling window discarding mechanism, short-term memory always maintains real-time context coverage of the most recent rounds of decisions for the current event, supporting the extraction of experience trajectories from the decision context generation units in subsequent rolling windows. Taking this scenario as an example, after the structured event slice generated in this round of decisions is pushed into the short-term buffer, the short-term memory contains the entire decision trajectory of the current event from the initial moment to the current moment.

[0071] The writing unit performs long-term retention priority calculations and writes the data into long-term memory. Long-term memory is a persistent storage area for cross-event experience reuse and strategy transfer. Its architecture adopts a structured template format. Each long-term memory record includes an event feature summary, action trajectory summary, tool call result summary, resilience evaluation result summary, successful experience entries, failure reason entries, correction suggestions, and applicable condition tags. The storage capacity of long-term memory has no fixed upper limit, but its content quality and retrieval efficiency are controlled through the writing unit's priority calculation and elimination mechanism. The writing unit first scans all event slices in short-term memory and extracts candidate experience entries from each event slice. Candidate experience entries include the action sequence and resilience score of the optimal solution in this round, the failure reasons and verification conclusions of the eliminated solutions, key return values ​​of the simulation tool, and key characteristics of the disaster state.

[0072] For each candidate experience item, the writing unit comprehensively evaluates the following four indicators to calculate its long-term retention priority. The first indicator is reuse frequency, representing the cumulative number of times the experience item has been retrieved and actually referenced in subsequent decision rounds within short-term memory. This value is obtained from the short-term memory retrieval and call logs. A higher reuse frequency indicates that the experience item has higher reference value in the continuous decision-making of the current event. The second indicator is effect improvement, representing the decrease in risk value or the increase in resilience score of the target node after adopting the solution corresponding to the experience item. This value is extracted from the difference in resilience score or the change in risk value between the two rounds recorded in the event slice. A greater effect improvement indicates that the decision-making model carried by the experience item has a stronger actual improvement effect. The third indicator is scope of application, representing the number of scenario types covered by the disaster characteristics and spatial topology characteristics corresponding to the experience item in the historical cases of the vertical domain knowledge base. This value is obtained by performing similarity matching between the disaster characteristic vector of the experience item and all historical cases in the vertical domain knowledge base, and counting the number of cases with a matching degree exceeding a preset threshold. A wider scope of application indicates that the experience item has cross-scenario transfer value. The fourth indicator is time decay, which represents the degree of decay in reference value caused by the time interval from the time of generation of the experience item to the current time. This value is calculated by linear normalization based on the difference between the generation timestamp of the experience item and the current timestamp. The longer the time interval, the greater the decay value, indicating that the timeliness of the experience item is weaker.

[0073] Based on the above four indicators, the writing unit calculates the long-term retention priority value for each candidate experience item using a weighted summation formula, which is: ,in, This represents the long-term retention priority value of the j-th candidate experience item. This represents the normalized value of the reuse frequency of this experience item. This indicates the normalized value representing the improvement in decision-making effectiveness brought about by this experience item. This indicates the normalized value representing the scope of application of the experience entry. This represents the time decay normalized value of the experience entry. , , , This refers to the weighting coefficients. The first three terms in the formula are added together, reflecting the positive retention value of the experience item; the fourth term is subtracted, reflecting the negative attenuation effect of time on the reference value of the experience item. The values ​​of the weighting coefficients are determined by domain engineers based on the emphasis placed on the stability of experience reuse and decision-making effectiveness in the field of urban flood disaster management, and are selected within a specific setting. =0.35、 =0.30、 =0.20、 =0.15, where the reuse frequency and the improvement of decision-making effect have higher weights, so as to prioritize the retention of experience items that have stable reference value in multiple decisions and can significantly improve the quality of decision-making.

[0074] Taking the successful experience item of Option 1 in the aforementioned hub station scenario as an example, the item's content is the action pattern of "delaying the northwest entrance / exit by three minutes before closing it for one-way exit, guiding passengers in the western area of ​​the station hall to transfer via the central staircase, and prioritizing the evacuation of vulnerable groups," with a resilience score of 0.5478. In the subsequent three rounds of rolling decision-making for the current event, this experience item was retrieved and cited twice, with a reuse frequency normalized value of [value missing]. The value was set to 0.67; after the implementation of this scheme, the risk value of the target node decreased from 0.546 to 0.28, and the effect improved the normalized value. The value is 0.78; the disaster feature vector of this experience item matches four of the six types of subway station flooding cases in the vertical domain knowledge base with a degree exceeding the threshold, and the applicable scope is normalized. The value is 0.67; this experience entry was generated only three decision cycles ago, and the time decay normalized value is... The value is 0.15. Substituting this into the above formula, we get... =0.5800. The write unit performs a threshold determination on the calculation result. When When the value is greater than or equal to 0.60, the experience entry is determined to be a high-priority entry, and after structured summary generation is initiated, it is written into long-term memory; when... When the value is greater than or equal to 0.40 and less than 0.60, the experience entry is classified as a medium-priority entry and, after compression, is retained in long-term memory in a low-priority form; when... When the score is less than 0.40, the experience entry is classified as a low-priority entry and will be either downgraded for storage or deleted. Taking the successful experience entry from Scheme 1 as an example, its... The value is 0.5800, which is in the medium priority range, so it is compressed and written to long-term memory in a low priority form.

[0075] Meanwhile, the reason for the failure of the aforementioned Option 3 (keeping the northwest entrance and exit open and arranging for personnel to remain there) was highly reused and its effectiveness was significantly improved (avoiding the misadoption of high-risk options) because it closely matched historical failure cases in this round and was repeatedly retrieved in the subsequent two rounds of decision-making to exclude similar options. After calculation... The value is 0.63, exceeding the high-priority threshold of 0.60. Therefore, the writing unit initiates a structured summary for this failure reason entry, compressing its core content into a summary text: "When the water depth outside the entrance / exit exceeds 0.30 meters and continues to rise, the strategy of keeping it open and arranging for personnel to remain inside the station is highly consistent with historical failure patterns, and should prioritize the implementation of lockdown and guided evacuation." This summary text, along with the corresponding disaster characteristic tags, applicable conditions, and failure reason identifiers, is written into long-term memory. During the writing process into long-term memory, the writing unit performs structured summarization and content compression on experience entries, compressing the lengthy simulation numerical sequences, time-by-time state snapshots, and detailed verification logs in the original event slices into core rule entries and action paradigm summaries to avoid disordered accumulation of historical information and improve subsequent retrieval efficiency. As the current event continues to evolve, the writing unit organizes and optimizes short-term and long-term memory at fixed intervals, prioritizing the long-term retention of experience entries with high reuse frequency and stable decision-making effects, while compressing or deleting experience entries whose timeliness has significantly decreased and whose reference value is weak.

[0076] In summary, a vertical-domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation, based on embodiments of this application, has been elucidated. Addressing the core issues of complex multi-system coupling, diverse decision-making objectives, and the difficulty of achieving unified cross-system representation and continuous rolling management in urban flood disaster scenarios using existing methods, this invention proposes a vertical-domain intelligent agent-driven management and control system for urban engineering disaster prevention and mitigation. Specifically, based on structured coupled state modeling, multi-source heterogeneous data is spatiotemporally aligned and quality-verified. Through entity feature extraction, node dynamic risk assessment, and cascading relationship fusion, a unified coupled state object is constructed, overcoming the shortcomings of existing technologies where the analysis of each subsystem is isolated and unable to represent the impact of cross-system cascading failures. Furthermore, a probabilistic sampling mechanism of a large language model drives the generation of multiple candidate solutions, combined with vertical-domain knowledge base rules, historical case retrieval, and long short-term memory, to generate a candidate solution set, solving the problems of insufficient professional capabilities and inference illusions inherent in general-purpose large language models. Subsequently, candidate solutions are screened through complexity identification and streamlining simulation, followed by cross-interception verification. A weighted scoring and ranking based on multi-dimensional resilience indicators ensures the output solution is optimal in terms of safety, efficiency, fairness, and resource consumption. Finally, a long short-term memory linkage writing mechanism is used to structure and store the decision trajectory and verification conclusions, supporting continuous rolling decision-making and cross-event experience reuse, achieving integrated resilience management of pre-disaster prevention, in-disaster response, and post-disaster recovery.

[0077] Those skilled in the art will understand that the steps, measures, and schemes in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted; furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted; furthermore, the steps, measures, and schemes in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above-described embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent for the embodiments of this disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this disclosure, and these all fall within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the appended claims. As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and detail can be made without departing from the spirit and scope of the present invention as defined in the appended claims.

[0078] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation, characterized in that, include: The heterogeneous data coupling module is used to perform coupling state modeling on multi-source heterogeneous data, including rainfall and water accumulation monitoring, road traffic operation, underground space facility topology, pedestrian and vehicle flow, and emergency resource allocation, through spatiotemporal alignment and data quality verification, in order to obtain coupled state objects. The candidate solution generation module is used to generate multiple candidate solutions for the current disaster situation based on coupled state objects, combined with the rules and case retrieval results of the vertical domain knowledge base and long short-term memory, and using the probability sampling mechanism of the large language model to obtain a set of candidate solutions including lockdown, evacuation, diversion and resource deployment control actions. The solution selection module is used to identify the complexity of the candidate solution set and perform flow deduction and verification to obtain the selected solution set. The resilience score generation module is used to cross-check the selected scheme set in terms of rule consistency, state consistency, security and goal achievement, and to determine the resilience score result by weighting and ranking the schemes based on multi-dimensional resilience indicators. The scheme verification and writing module is used to determine strategies and update memory based on resilience score results. It outputs the scheme with the highest score that meets the constraints as the optimal control strategy and retains the set of alternative schemes. At the same time, it writes the decision trajectory and verification conclusion of this round into long short-term memory after structured summarization.

2. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 1, characterized in that, The heterogeneous data coupling module includes: Heterogeneous data structuring unit, used to perform acquisition frequency alignment and spatial dimension unified encoding on multi-source heterogeneous data to obtain structured input objects; The entity extraction risk assessment unit is used to extract entity features and perform dynamic risk assessment on structured input objects to obtain entity relationship sets and node risk sets; The Coupled State Graph Construction Unit is used to perform cascading relationship fusion and coupled state graph construction on entity relationship sets and node risk sets to obtain coupled state objects.

3. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 2, characterized in that, The entity extraction risk assessment unit includes: The entity relation set construction subunit is used to extract the attributes and topological connections of nodes from the structured input object to obtain the entity relation set; The risk quantification subunit is used to quantify the risk of each node in the structured input object, including water depth, flow velocity, inundation duration, and population density, to obtain the node risk set.

4. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 1, characterized in that, The candidate solution generation module includes: The decision context generation unit is used to convert coupled state objects into state description text based on preset structured prompt word templates, and then perform vector dimension splicing and field fusion with the rules retrieved from the vertical domain knowledge base and the experience trajectories in long short-term memory to obtain the decision context. The decision context probability sampling unit is used to generate multi-branch actions based on probability sampling of the decision context to obtain an action sequence; The action sequence encapsulation unit is used to extract key entities and encapsulate the action sequence in a structured manner. The scope, execution time and resource requirements are mapped to the standard action structure to obtain a set of candidate solutions.

5. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 1, characterized in that, The scheme selection module includes: The scheme analysis unit is used to analyze the scheme features and determine the task complexity flow of each scheme in the candidate scheme set to obtain the direct verification subset and the auxiliary verification subset; The simulation subset generation unit is used to perform feature matching between each scheme in the auxiliary verification subset and historical disaster cases, and to remove schemes that are highly matched with failed cases to obtain the simulation subset; The subset merging unit is used to perform multi-branch simulations on each scheme in the subset to be simulated, and after eliminating inefficient branches that expose safety risks, merge the effective branches with the directly verified subset to obtain the set of selected schemes.

6. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 1, characterized in that, The toughness score generation module includes: The scheme binary judgment unit is used to perform binary judgments on each scheme in the selected scheme set in turn through a unified checker to determine the rule consistency, state consistency, safety and goal achievement in order to obtain a set of feasible schemes. The resilience assessment and determination unit is used to perform multi-dimensional resilience weighted assessment and ranking of the set of feasible solutions to obtain resilience score results.

7. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 1, characterized in that, The scheme verification and writing module includes: The scheme splicing unit is used to extract the top-ranked scheme that meets the resource constraints as the optimal control strategy based on the descending sort of resilience score results, extract the remaining schemes that meet the safety threshold as the candidate scheme set, and splice the above scheme metadata to generate the scheme result set. The key-value association binding unit is used to associate and bind the solution result set with the coupled state object to obtain structured event slices, and push them to a short-term cache with a rolling window eviction mechanism to obtain short-term memory. The writing unit is used to calculate the long-term retention priority of each experience item in short-term memory according to the frequency of reuse, effect enhancement, scope of application and time decay, and write the experience items that exceed the threshold into long-term memory after structured summarization.

8. The vertical domain intelligent agent-driven control system for urban engineering disaster prevention and mitigation according to claim 5, characterized in that, The scheme parsing unit is used for: The inherent complexity of the candidate solution set is quantified to obtain an inherent complexity vector; The coupling coefficients of the candidate schemes are calculated to obtain the scheme coupling matrix; The coupling enhancement complexity vector is obtained by performing graph-weighted propagation-based calculation on the inherent complexity vector and the scheme coupling matrix. Adaptive routing based on coupling density is performed on the coupling enhancement complexity vector and the scheme coupling matrix to obtain the direct verification subset and the auxiliary verification subset.

Images