A method and system for multi-scenario management and control of railway stations based on a large model
By dividing the station into zones and generating situation tensors, feasible domains and policy drafts are constructed, solving the problems of decision generalization and adaptability of existing station management methods in complex scenarios. This achieves compliance and security of station management and meets the traceability requirements of public transportation safety.
Patent Information
- Application Number
- CN202611618343.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing station management methods suffer from several problems when facing complex scenarios, including weak generalization ability of decision kernels, lack of physical coupling and priority arbitration between scenarios, ineffective use of large models, difficulty in online rolling of optimization solution methods, lack of adaptive correction and operational behavior modeling in simulation models, static solidification of knowledge and contingency plan bases, and lack of execution effect feedback and strategy switching jitter suppression.
By dividing the station into control zones, a station-wide situation tensor is generated. Based on the situation tensor, a fingerprint vector of the current scene is generated, a physical and rule-feasible domain is constructed, a candidate strategy draft is generated and its feasibility is determined, short-term domain inference and optimal strategy selection are performed, and a dual-channel update mechanism of hierarchical authorization and deviation-driven is combined to realize automatic switching and adaptive optimization of strategies.
It ensures compliance and effectiveness of station management in complex scenarios, avoids frequent fluctuations in policy switching, ensures safety and response speed, and enables adaptive updates of the knowledge base and contingency plans, meeting the traceability requirements for public transportation safety responsibilities.
Smart Images

Figure CN123117735A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transportation hub management and control technology, and in particular to a method and system for multi-scenario management and control of stations based on a large model. Background Technology
[0002] Railway stations are a type of strongly coupled dynamic system driven by train arrivals and departures in terms of time and by passenger flow in terms of space. Their management and control have the following characteristics that distinguish them from general industrial objects: First, the scenarios are highly complex and coexist concurrently. Scenarios such as regular passenger transport, morning and evening rush hours, transfer connections, train delays, equipment failures, sudden large passenger flows, severe weather, and sudden emergency evacuations do not occur in a mutually exclusive manner. Instead, they often coexist in a way that one main scenario leads to several sub-scenarios. For example, the main scenario of large passenger flows during the morning rush hour can simultaneously lead to three sub-scenarios: congestion in transfer channels, platform congestion, and overflowing queues at security checkpoints.
[0003] Secondly, the constraints have hard safety attributes. Station management involves actions that directly affect personal safety, such as flow control, station closure, evacuation, and switching of turnstile and escalator operating directions. Any strategy must meet hard constraints such as traffic capacity, area capacity, action sequence, and authority specifications. It is not permissible to substitute high probability of correctness for certain feasibility.
[0004] Third, decisions must be traceable and auditable. Stations are the main responsible parties for public transportation safety. Once control actions are implemented, they must be able to answer four questions: what situation is it based on, what rules are referenced, who authorized it, and what are the expected results?
[0005] The above solutions represent the current mainstream approaches: interface and function integration, threshold and rule triggering, optimization and prediction model-based solutions, and large model-assisted solutions. Each of these approaches has the following unresolved shortcomings: (1) The decision-making core is still threshold-rule-prepared plan, which has weak generalization ability for complex scenarios outside the rules.
[0006] (2) There is a lack of physical coupling and priority arbitration between scenarios.
[0007] (3) Large models are either excluded from the decision-making loop or placed in an uncontrollable position.
[0008] (4) The parameters of the optimization solution method are fixed and difficult to scroll online.
[0009] (5) The simulation model lacks adaptive correction and operational behavior modeling.
[0010] (6) Static solidification of knowledge and contingency plan database.
[0011] (7) Lack of execution effect feedback, strategy switching jitter suppression and human-machine hierarchical authorization.
[0012] Therefore, it is evident that the existing station management methods described above still have inconveniences and shortcomings in their use, and urgently need further improvement. How to create a new multi-scenario station management method based on a large model has become a pressing goal for the industry. Summary of the Invention
[0013] In view of this, the present disclosure provides a station multi-scenario management and control method based on a large model, which at least partially solves the problems existing in the prior art.
[0014] In a first aspect, embodiments of this disclosure provide a multi-scenario management and control method for railway stations based on a large model, the method comprising the following steps: The station is divided into control zones, and the collected data is aggregated into situation features according to the zones to generate a station-wide situation tensor; Based on the overall situation tensor and its average rate of change within the observation window, a fingerprint vector for the current scene is generated; the fingerprint vector for the current scene is compared with the prototype fingerprint vector of a predefined scene to obtain a set of main scenes and sub-scenes. Using the partitions involved in the main scene and the set of sub-scenes as constraints, a physical and rule-feasible domain is constructed; Using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content, a structured prompt is constructed; based on the structured prompt, candidate strategy drafts and the self-evaluation confidence of each candidate strategy draft are generated; Feasibility is determined and the feasibility rate is calculated based on the candidate strategy drafts; wherein, candidate strategy drafts with a feasibility rate lower than the feasibility rate threshold are rejected as a whole and required to be regenerated, and candidate strategy drafts with a feasibility rate not lower than the feasibility rate threshold are projected and mapped into the physical and rule feasible domain, and the candidate strategy after projection repair is output. Using the candidate strategies after projection repair as the inference objects, short-time domain inferences are performed respectively to calculate the optimal candidate strategy and its comprehensive utility. The utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is executed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic deployment is executed according to the comprehensive confidence of the optimal candidate strategy in intervals, and the execution results are output. The measured utility value is recalculated based on the measured data after execution, and the relative prediction deviation between the measured utility value and the predicted utility value is calculated; the symbolic side channel and the neural side channel are updated simultaneously using the relative prediction deviation as the driving force.
[0015] According to a specific implementation of this disclosure, the step of dividing the station into control zones and aggregating the collected data into situational features according to the zones to generate a station-wide situational tensor includes: The station is divided into several control zones according to the station's architectural plan and passenger flow lines, with the zone boundaries defined by physical cross-sections; wherein, the physical cross-sections include at least one of the following: turnstile groups, stairwells, and passageway intersections; Establish partition adjacency relationships and determine the upstream partition set for each partition; Collect data; preprocess the data and aggregate it into situational features by region to form the overall situational tensor of the station; wherein the data includes: video passenger flow detection data, AFC entry and exit records, gate passage count, security checkpoint queue length, escalator and elevator operating status, environmental sensor data, train arrival and departure times and platform occupancy information, and meteorological and surrounding activity information.
[0016] According to a specific implementation of this disclosure, the step of generating a current scene fingerprint vector based on the overall situation tensor and its average rate of change within the observation window; and obtaining a set of main scenes and sub-scenes by comparing the current scene fingerprint vector with the prototype fingerprint vector of a predefined scene, including: The whole-station situation tensor and the average rate of change tensor of the whole-station situation tensor within the observation window are concatenated and input into the pre-trained scene fingerprint encoder to obtain the current scene fingerprint vector; the average rate of change tensor is the difference between the whole-station situation tensor at the current time and the whole-station situation tensor at the beginning of the observation window divided by the duration of the observation window. The cosine similarity is calculated based on the fingerprint vector of the current scene and the prototype fingerprint vector of each predefined scene. The scene membership degree of each predefined scene is obtained by taking the exponent of the cosine similarity divided by the temperature coefficient, and then summing the cosine similarity over all predefined scenes. The predefined scene with the highest scene membership degree is selected as the main scene; The set of sub-scenes is formed by taking the values of the adjacent elements of the scene dependency graph to represent predefined scenes that are coupled and whose scene membership degree is not lower than the lower limit of the membership degree of the linked sub-scenes. The scene dependency graph is jointly determined by the station topology and operation procedures. If two predefined scenes involve the same control zone or have a passenger flow transfer relationship, then the adjacent elements of the two predefined scenes are values that indicate that they are coupled.
[0017] According to a specific implementation of this disclosure, the step of constructing a physically and rule-feasible domain, using the partitions involved in the main scene and the set of sub-scenes as constraints, includes: Taking all partitions involved in the main scene and the set of sub-scenes as the constraint range, the intersection of the following five types of constraints is taken as the physical and rule feasible domain, and recalculated periodically according to the equipment status and train arrival and departure schedule; A capacity constraint is set, requiring that the simulation density of any involved zone in any simulation step does not exceed the congestion critical density, which is determined by the station design documents and service level classification. A capacity constraint is set, requiring that the calculated flow of any involved zone in any calculation step does not exceed the actual capacity. The actual capacity is the product of the equipment status reduction factor and the design capacity. The equipment status reduction factor is the ratio of the number of available turnstiles, escalators, and passages under the jurisdiction of the zone to the rated number of available turnstiles, escalators, and passages, and is one when all equipment is available. The design capacity is taken from the station design documents. Set timing constraints, requiring that the effective delay of any action item does not exceed the maximum allowable effective delay, and that the effective time of the action falls within the simulation time domain. The maximum allowable effective delay is determined by the longest measured time of on-site broadcast response, guide screen switching and personnel arrival. Set mutual exclusion and interlocking constraints, requiring that no action pair in the same strategy belongs to the set of mutually exclusive action pairs; Set permissions and norms to require that the permission level required for any action item is no higher than the currently available authorization level. The permission level is divided according to the degree of direct impact of the action on personal safety, with the highest level being used for actions such as traffic restriction, site blocking, and evacuation. The physical and rule-feasible domain is formed by using candidate control strategies that simultaneously satisfy five types of constraints.
[0018] According to a specific implementation of this disclosure, the step of constructing structured prompts using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content; and generating candidate strategy drafts and self-assessed confidence levels for each candidate strategy draft based on the structured prompts, includes: Based on the current scene fingerprint vector, retrieve the historical cases with the highest cosine similarity in the case library as the similar historical cases, and record the highest case similarity. The highest case similarity is the maximum value among the cosine similarities between the current scene fingerprint vector and the scene fingerprint vectors of each historical case. The structured prompt is constructed using the main scene and the set of sub-scenes, the verbal description of the five types of constraints, the obtained several similar historical cases and their handling results and actual effects, and the output format as the intention layer structured strategy draft, and the format constraints that provide evidence for each item as the content. The verbal description includes the current reduced passage capacity value, the list of available actions, and the list of mutual exclusion pairs. Using the structured prompts as input, the station management big data model outputs several candidate strategy drafts and the self-assessed confidence level of each candidate strategy draft. The candidate strategy drafts are composed of several action items in sequence. Each action item includes the action object, action type, area of effect, start time and duration, and supporting evidence.
[0019] According to a specific implementation of this disclosure, the feasibility determination and calculation of the feasibility rate are based on the candidate strategy drafts; wherein, candidate strategy drafts with a feasibility rate lower than a feasibility rate threshold are rejected entirely and required to be regenerated, and candidate strategy drafts with a feasibility rate not lower than a feasibility rate threshold are projected and mapped to the physical and rule-based feasible domain, and the projected and repaired candidate strategies are output, including: For each candidate strategy draft, its action items are judged sequentially according to four levels: syntax, permissions, topology interlocks, and physical mechanisms. First, it is checked whether the action item format and device identifier can be parsed. Then, it is checked whether the permission level required by the action item does not exceed the currently available authorization level. Next, it is checked whether the action item falls into the set of mutually exclusive action pairs with other action items in the same strategy, and whether it conflicts with the upstream and downstream states of the device. Finally, it is checked whether the action item's effective partition satisfies the capacity constraints and traffic capacity constraints in the extrapolation time domain. If any level fails, the feasibility judgment result of the action item is 0; if all levels pass, the result is 1. The feasibility rate of a draft is defined as the ratio of the number of actions with a feasibility assessment result of 1 to the total number of actions in that draft. If the feasibility rate is lower than the feasibility rate threshold, the overall structure of the draft is deemed invalid, the draft is rejected, and the station management model is required to be regenerated based on the reasons for rejection. If the feasibility rate is not lower than the feasibility rate threshold, then within the physical and rule-based feasible domain, the strategy with the smallest difference from the original candidate strategy draft is selected as the candidate strategy after projection repair. The difference is calculated by summing the equivalent modification costs for each of the three types of operations: adding, deleting, and modifying action items. The equivalent modification costs are taken from the station action cost table.
[0020] According to a specific implementation of this disclosure, the step of performing short-time domain inferences on the candidate strategies after projection repair as the inference objects, and calculating the optimal candidate strategy and its comprehensive utility, includes: For each candidate strategy after projection repair, several projection steps are advanced from the current time according to the projection step size. Each step is calculated and iterated to the end of the projection time domain according to the following relationship: the passenger flow density is obtained by dividing the number of people staying in the partition by the effective passage area of the partition; the average passenger flow speed is obtained by multiplying the free flow walking speed by the difference between the passenger flow density and the congestion density, where the free flow walking speed is the statistical average of the pedestrian free flow speed measured at the station, and the congestion density is the density when the speed drops to zero; the passenger flow density, the average passenger flow speed, and the effective passage area of the partition are used to calculate the passenger flow density. The passenger flow rate is obtained by multiplying the passage width, the three factors, and the passenger flow volume. The passenger arrival rate is obtained by adding the sum of the passenger flow of the upstream adjacent zones to the external passenger flow inflow rate of that zone. The external passenger flow inflow rate is estimated from the ticket check records at the entrance, the passenger flow detection at the entrance and exit, and the number of passengers disembarking from the train. The smaller value between the actual passage capacity and the number of passengers that can pass is taken as the passenger flow throughput rate. The number of passengers that can pass is the sum of the number of stranded passengers divided by the projection step length and the passenger flow arrival rate. The difference between the passenger flow arrival rate and the passenger flow throughput rate is multiplied by the projection step length and accumulated to the number of stranded passengers to obtain the number of stranded passengers in the next projection step. The relative speed decay rate is obtained by taking the positive part of the ratio of the difference between the average walking speed of two adjacent steps to the reference value of walking speed, where the reference value of walking speed is the statistical average of walking speed during the normal off-peak period of the zone; the congestion risk index is obtained by multiplying the square of the ratio of passenger flow density to congestion critical density by the product of the relative speed decay rate and the speed decay sensitivity coefficient. The difference between the passenger flow of a partition in the simulation step when implementing the candidate strategy and the baseline flow when no intervention is taken is accumulated over all simulation steps and all partitions and multiplied by the simulation step length to obtain the number of people passing through; the average station-wide risk cost is obtained by averaging the congestion risk index of a partition in the simulation step when implementing the candidate strategy over all simulation steps and all partitions. The service disruption cost is obtained by summing the product of the number of people affected by each action and the duration of the disruption. The number of people affected is calculated based on the average number of people staying in the affected area within the projected time domain. The equivalent operating cost is obtained by summing the equivalent operating costs of each action. The equivalent operating cost is calculated from the station cost quota table according to the action type and duration. The following four normalized components are obtained by dividing the number of people passing through, the average station-wide risk cost, the service interference cost, and the equivalent operating cost by the reference value of the number of people passing through, the target value of the risk index, the reference value of the service interference, and the reference value of the equivalent operating cost, respectively: the reference value of the number of people passing through is the station's capacity under the congestion critical density; the reference value of the service interference is the product of the capacity and the projected time domain length; the reference value of the equivalent operating cost is the product of the equivalent operating manpower that can be deployed per shift and the number of shift minutes; and the target value of the risk index is specified by the station's safe operation procedures. The comprehensive utility is obtained by multiplying the efficiency weight by the normalized number of people, the reduced risk weight by the normalized risk weight, the reduced service interference weight by the normalized service interference weight, and the reduced cost weight by the normalized cost weight. The sum of the four weights is one and all are non-negative, and their initial values are given by the station operation regulations. The candidate strategy with the highest overall utility is selected as the optimal candidate strategy.
[0021] According to a specific implementation of this disclosure, the utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is performed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic deployment is performed according to intervals based on the comprehensive confidence of the optimal candidate strategy, and the execution results are output, including: The utility gain is obtained by subtracting the utility of the current execution strategy, which is recalculated using the same utility function and the current situation, from the comprehensive utility of the optimal candidate strategy. The overall confidence level is the product of the feasibility rate and a weighted term, whereby the weighted term is obtained by weighting the self-assessed confidence level and the highest case similarity according to the case support fusion coefficient. The current execution strategy shall be replaced by the optimal candidate strategy if and only if the utility gain is not lower than the switching hysteresis threshold and the difference between the current time and the last strategy switching time is not less than the minimum dwell time; otherwise, the current execution strategy shall be maintained. Execution is selected from the following intervals based on the overall confidence level: Execution is automatically initiated when the overall confidence level is not lower than the automatic execution threshold; Execution is automatically initiated when the overall confidence level is not lower than the manual confirmation lower limit but lower than the automatic execution threshold, and is pushed to the duty officer along with the evidence chain; if the duty officer does not withdraw the request within the specified time, execution is automatic; Execution is initiated when the overall confidence level is lower than the manual confirmation lower limit, and is executed after explicit confirmation by the duty officer; The manual confirmation lower limit is less than the automatic execution threshold.
[0022] According to a specific implementation of this disclosure, the step of recalculating the measured utility value based on the measured data after execution, and calculating the relative prediction deviation between the measured utility value and the predicted utility value; and simultaneously updating the symbolic side channel and the neural side channel using the relative prediction deviation as the driving variable, includes: After the strategy has been executed for a full simulation time domain, the measured utility value is recalculated based on the actual flow, density, number of people staying, and actual cost. The absolute value of the difference between the measured utility value and the predicted utility value is divided by the sum of the absolute value of the predicted utility value and the zero-prevention constant to obtain the relative prediction deviation. The predicted utility value is the comprehensive utility of the optimal candidate strategy calculated at the decision time. Update the weights in the symbol-side channel; using the difference between the predicted utility value and the measured utility value as the error, and the normalized component vector of the selected candidate as the gradient direction, update the four weights according to the utility weight learning rate, and then project the updated weights onto the weight simplex. Update the throughput capacity in the symbol-side channel; when the zone meets the calibration conditions of continuous upstream supply, non-zero number of people staying in the zone, and all equipment under its jurisdiction is available, revise the design throughput capacity by multiplying the throughput capacity calibration learning rate by the difference between the measured throughput rate and the design throughput capacity. Update the case retrieval weights in the neural side channel; after attenuating the retrieval weights of historical cases by the forgetting factor, add the product of the cosine similarity between the current scene fingerprint vector and the scene fingerprint vector of the historical case, and a negative exponential term with the relative prediction deviation as the exponent; Update the cue constraints in the neural side channel; write the current period's situation snapshot, rejected actions and their reasons for rejection, repaired actions and their repair costs into the constraint appendix of the cue template; Generate audit logs; the audit logs include: situation snapshot and main / sub-scenario determination results, feasible domain parameters, original text of candidate strategy drafts and supporting evidence, item-by-item feasibility determination results and items to be fixed, utility and sub-items of each candidate, confidence level and authorization level, authorizer, execution time and measured results.
[0023] Secondly, embodiments of this disclosure provide a station multi-scenario management and control system based on a large model, the system comprising: The partitioning module is configured to divide the station into control zones and aggregate the collected data into situation features according to the zones, generating a station-wide situation tensor. The scene module is configured to generate a current scene fingerprint vector based on the overall situation tensor and its average rate of change within the observation window; and to obtain a set of main scenes and sub-scenes by comparing the current scene fingerprint vector with the prototype fingerprint vector of a predefined scene. Using the partitions involved in the main scene and the set of sub-scenes as constraints, a physical and rule-feasible domain is constructed; Furthermore, a structured prompt is constructed using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content; and a candidate strategy draft and the self-evaluation confidence of each candidate strategy draft are generated based on the structured prompt; The judgment module is configured to determine the feasibility of the candidate strategy draft and calculate the feasibility rate; wherein, the candidate strategy draft with a feasibility rate lower than the feasibility rate threshold is rejected as a whole and required to be regenerated, and the candidate strategy draft with a feasibility rate not lower than the feasibility rate threshold is projected and mapped into the physical and rule feasible domain, and the candidate strategy after projection repair is output. Using the candidate strategies after projection repair as the inference objects, short-time domain inferences are performed respectively to calculate the optimal candidate strategy and its comprehensive utility. The utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is executed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic deployment is executed according to the comprehensive confidence of the optimal candidate strategy in intervals, and the execution results are output. The update module is configured to recalculate the measured utility value based on the measured data after execution, and calculate the relative prediction deviation between the measured utility value and the predicted utility value; and update the symbolic side channel and the neural side channel simultaneously with the relative prediction deviation as the driving quantity.
[0024] The station multi-scenario management and control method based on a large model in this embodiment has the following beneficial effects: (1) The large model generates a draft under the condition of known feasible domain boundaries, and is deterministically repaired by the projection operator, so that the action items entering the execution stage are all compliant in four aspects: syntax, permissions, topological interlock and physical mechanism, thus eliminating the possibility of physical illusion type instructions being allowed from the mechanism. (2) The scenario dependency graph determines the linkage range and resource allocation priority when multiple scenarios are concurrent by the coupling strength, avoiding instruction discreteness and mutual cancellation; (3) The risk index takes into account both density and walking speed decay, so that signs of stampede such as stop-walking waves can be captured in advance, and warnings are given before congestion occurs. (4) The delayed switching condition deterministically suppresses the frequency of strategy switching, avoiding passenger confusion and order disruption caused by frequent reversals of on-site guidance information; (5) Hierarchical authorization allows high-risk actions to retain a manual confirmation process, while low-risk actions are executed automatically, balancing response speed and accountability. (6) Deviation-driven dual-channel evolution enables the feasible domain parameters, utility weights, and case weights to be adaptively updated with station renovations, timetable changes, and seasonal passenger flow characteristics, overcoming the defects of static and fixed knowledge base and contingency plan base. (7) Generate audit records throughout the process, including situation snapshots, supporting evidence, confidence level, authorized person and expected results, to meet the traceability requirements of the public transportation safety responsibility entities. Attached Figure Description
[0025] Figure 1 A schematic diagram of a multi-scenario station management and control method based on a large model provided in this disclosure embodiment; Figure 2 A bidirectional coupling relationship diagram between a large model, feasible region, and twin inference provided in this embodiment of the disclosure; Figure 3A schematic diagram of a scene dependency graph and main / sub-scene determination provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of a station multi-scenario control system based on a large model, provided as an embodiment of the present disclosure. Detailed Implementation
[0026] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.
[0027] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0028] It should be noted that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. Furthermore, this device and / or this method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.
[0029] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.
[0030] This invention provides a multi-scenario station management method based on a large model. It constructs a bidirectional coupled closed loop: large model intent generation—feasible region projection repair—twin-based optimization—hysteresis-driven dual-channel evolution—deviation-driven dual-channel evolution. Figure 2 As shown, where: (1) The interaction between the feasible region and the large model is bidirectional, rather than unidirectional. On the one hand, the feasible region is written into the large model as structured language constraints, so that the generation process "a priori" falls near the feasible region; on the other hand, the draft generated by the large model is mapped back to the feasible region through the projection operator, and the deviation generated by the projection (the proportion of action terms that are corrected) serves as a feedback signal on whether the feasible region is clearly expressed, driving the linguistic expression of the feasible region constraints and parameter revision. The two are each other's inputs and outputs.
[0031] (2) The interaction between twin inference and the large model is bidirectional. Inference provides utility scores for multiple candidates generated by the large model. The scores are not only the basis for selection, but also serve as a sample to be fed back into the case library to determine "which candidates are adopted and what their actual effects are", thus changing the prior distribution of subsequent retrieval and generation.
[0032] (3) The evolutionary path acts on both the symbolic and neural sides. The execution bias updates the feasible domain parameters and utility weights (symbolic side, which can be interpreted and audited by rules) on the one hand, and updates the case retrieval weights and prompt templates (neural side, which changes the generation distribution) on the other hand. The two paths are driven by the same bias amount and share the same audit record.
[0033] Figure 1 This is a schematic diagram of the process of a station multi-scenario management and control method based on a large model provided in an embodiment of this disclosure.
[0034] like Figure 1 As shown, in step S110, the station is divided into control zones, and the collected data is aggregated into situation features according to the zones to generate the overall station situation tensor.
[0035] In this embodiment of the invention, the step of dividing the station into control zones and aggregating the collected data into situational features according to the zones to generate a station-wide situational tensor includes: dividing the station into several control zones according to the station's architectural plan and passenger flow lines, with the zone boundaries being physical cross-sections; wherein the physical cross-sections include at least one of: turnstile groups, stairwells, and passageway intersections; establishing zone adjacency relationships and determining the upstream zone set of each zone; collecting data; preprocessing the data and aggregating it into situational features according to the zones to constitute the station-wide situational tensor; wherein the data includes: video passenger flow detection data, AFC entry and exit records, turnstile passage counts, security checkpoint queue lengths, escalator and elevator operating status, environmental sensor data, train arrival and departure times and platform occupancy information, and meteorological and surrounding activity information.
[0036] More specifically, the station partition topology construction and multi-source situation tensor generation include the following steps: Based on the station's architectural layout and passenger flow patterns, the station is divided into... Each zone is a control zone, with its boundaries defined by physical cross-sections (turnstile groups, stairwells, and passageway intersections). This allows the passenger flow within each zone to be described by a set of lumped parameters. Zone adjacency relationships are established, and the upstream zone set for each zone is determined. .
[0037] Data from multiple sources, including video passenger flow detection, AFC entry and exit records, turnstile passage counts, security checkpoint queue lengths, escalator and elevator operating status, environmental sensors, train arrival and departure times and platform occupancy, and meteorological and surrounding activity information, is collected. After time alignment, spatial mapping, and anomaly removal, the data is aggregated by region. The status characteristics constitute the overall site status tensor. .
[0038] More specifically, we now proceed to step S120.
[0039] In step S120, a current scene fingerprint vector is generated based on the overall situation tensor and its average rate of change within the observation window; the current scene fingerprint vector is compared with the prototype fingerprint vector of a predefined scene to obtain a set of main scenes and sub-scenes.
[0040] In this embodiment of the invention, the step of generating a current scene fingerprint vector based on the global situation tensor and its average rate of change within the observation window; and obtaining a set of main scenes and sub-scenes by comparing the current scene fingerprint vector with the prototype fingerprint vectors of predefined scenes, includes: concatenating the global situation tensor and the average rate of change tensor of the global situation tensor within the observation window and inputting them into a pre-trained scene fingerprint encoder to obtain the current scene fingerprint vector; the average rate of change tensor is the difference between the current global situation tensor and the global situation tensor at the start of the observation window divided by the duration of the observation window; and comparing the current scene fingerprint vector with the prototype fingerprint vectors of each predefined scene. Calculate the cosine similarity; take the exponent by dividing the cosine similarity by the temperature coefficient, and then sum the cosine similarity over all predefined scenarios to obtain the scenario membership degree of each predefined scenario; take the predefined scenario with the largest scenario membership degree as the main scenario; according to the scenario dependency graph, take the predefined scenarios whose adjacent element values represent coupling and whose scenario membership degree is not lower than the lower limit of the membership degree of the linked sub-scenario, and form the sub-scenario set; the scenario dependency graph is jointly determined by the station topology and operation procedures. If two predefined scenarios involve the same control zone or have a passenger flow transfer relationship, then the adjacent elements corresponding to the two predefined scenarios are values that indicate coupling.
[0041] More specifically, such as Figure 3 As shown, the scene fingerprint encoding and master / sub-scene determination specifically include the following steps: First, a scene fingerprint is generated from the state tensor and its rate of change within the observation window: ...(1) in, for The scene fingerprint vector at any given moment, with dimensions of ; For a pre-trained scene fingerprint encoder, the input is the concatenation of the current situation tensor and the situation change rate tensor; for The overall site status tensor at any given moment; Let the state tensor be the average rate of change tensor within the observation window. ,in For the number of observation window steps, To deduce the step size. If only using If the encodings are identical to the fingerprints, it will lead to the retrieval of incorrect historical cases.
[0042] Secondly, calculate the similarity between the current fingerprint and prototype fingerprints for various scenarios. : ...(2) in, Cosine similarity; For the first The prototype fingerprint vector of a scenario is determined by the cluster center of historical samples of that scenario. This is the fingerprint vector for the current scene.
[0043] Convert the similarity of the prototype fingerprints into scene membership: ...(3) in, For the scene exist Membership degree at any given moment; The smaller the value, the more concentrated the membership distribution is in the most similar scenario; The total number of predefined scenarios; To sum dummy variables.
[0044] In actual station operation, multiple scenarios coexist. Rigid categorization would miss secondary but developing risks. The affiliation system retains this information for use in the next step of linkage determination.
[0045] Next, determine the scope of linkage based on the scene dependency graph: ... (4) in, for The main scene sequence number of the moment; For the scene The degree of membership.
[0046] ...(5) in, A collection of sub-scenes that are driven by the main scene; This is the element in the adjacency matrix of the scene dependency graph that corresponds to the main scene. A value of 1 indicates that the scene is coupled with the main scene. This is the lower bound of the membership degree of linked sub-scenes, used to filter out scenes that are coupled but have not yet been developed.
[0047] The scenario dependency graph is jointly determined by the station topology and operating procedures: if two scenarios involve the same zone or have a passenger flow relationship, the corresponding adjacent elements are set to 1. For example, in a transfer station, the transfer passage collision and the platform congestion share the same platform zone, and the adjacent elements of the two are set to 1.
[0048] The main scenario determines the primary objective of this control measure, while the set of sub-scenarios determines the scope of linkage and the priority of resource allocation, thereby avoiding the problem of independent triggering of each scenario and mutual cancellation of instructions in existing technologies.
[0049] Next, proceed to step S130.
[0050] In step S130, a physical and rule-feasible domain is constructed using the partitions involved in the main scene and the set of sub-scenes as constraints.
[0051] In this embodiment of the invention, the construction of the physical and rule-feasible domain using the partitions involved in the main scene and the set of sub-scenes as the constraint range includes: using all partitions involved in the main scene and the set of sub-scenes as the constraint range, taking the intersection of the following five types of constraints as the physical and rule-feasible domain, and recalculating it periodically according to equipment status and train arrival and departure schedules; setting capacity constraints, requiring that the simulation density of any involved partition in any simulation step does not exceed the congestion critical density, the congestion critical density being determined by station design documents and service level classification; setting traffic capacity constraints, requiring that the simulation flow of any involved partition in any simulation step does not exceed the actual traffic capacity, the actual traffic capacity being the product of the equipment status reduction factor and the designed traffic capacity, the equipment status reduction factor being the number of turnstiles, escalators, and passageways under the jurisdiction of the partition. The ratio of the available number of units or equipment to the rated number of units or equipment, and the ratio when all equipment is available, is one. The designed traffic capacity is taken from the station design documents. Timing constraints are set, requiring that the effective delay of any action item does not exceed the maximum allowable effective delay, and that the effective time of the action falls within the projected time domain. The maximum allowable effective delay is determined by the longest measured time consumed by on-site broadcast response, guide screen switching, and personnel arrival. Mutual exclusion and interlocking constraints are set, requiring that no action pair in the same strategy belongs to a set of mutually exclusive action pairs. Permission and specification constraints are set, requiring that the permission level required for any action item is not higher than the currently available authorization level. The permission level is divided according to the direct impact of the action on personal safety, with the highest level used for flow restriction, station closure, and evacuation actions. The physical and rule-feasible domain is formed by candidate control strategies that simultaneously satisfy all five types of constraints.
[0052] More specifically, the construction of the feasible region, which is dynamically updated according to the situation, includes the following steps: The feasible region is defined by the intersection of the following five types of constraints: (C1) Capacity constraint: For any partition involved With any deduction step The calculated density must not exceed the critical congestion density, i.e. .in For partitioning Passenger flow density, The critical congestion density is determined by the station design documents and service level classification.
[0053] (C2) Capacity constraints: For any partition involved With any deduction step The projected traffic flow must not exceed the actual traffic capacity, i.e. Furthermore, the actual traffic capacity is reduced based on the equipment status: ... (6) in, For partitioning exist Actual traffic capacity at any given moment (people / s); This is the equipment condition reduction factor, with a value of [value missing]. The ratio is determined by the ratio of the available number of turnstiles, escalators, and passageways under the jurisdiction of the zone to the rated number; when all equipment is available, the ratio is 1. For partitioning The designed throughput capacity (people / s) is taken from the station design documents.
[0054] (C3) Timing constraint: any action item Effective delay It must not exceed the maximum allowed value. Furthermore, the time at which the action takes effect must fall within the deduction time domain. For action items The effective time delay, The maximum allowable effective time delay is determined by the longest measured time taken for on-site broadcast response, directional screen switching, and personnel arrival. This constraint excludes pseudo-feasible solutions that are correct in strategy but cannot be implemented in time.
[0055] (C4) Mutual Exclusion and Interlocking Constraints: If two actions form a mutual exclusion pair, they cannot both appear in the same policy. Let the set of mutually exclusive action pairs be denoted as . Then it requires For any pair of actions in the strategy to be true. A typical mutually exclusive pair is closing an entrance / exit and guiding passenger flow through that entrance / exit to exit the station.
[0056] (C5) Permissions and Normative Constraints: Any action item Required permission level It must not exceed the currently available authorization level. ,Right now Permission levels are divided according to the degree of direct impact of actions on personal safety, with the highest level being used for actions such as traffic restriction, site closure, and evacuation.
[0057] ... (7) in, for The feasible domain at any given time; Let be any candidate control strategy to be tested.
[0058] The feasible domain is not a static configuration table, but rather... Train arrival and departure plans are recalculated periodically with the current authorization level; at the same time, the feasible region will be linguisticized to form a priori constraints on the large model, which is the first edge of the bidirectional coupling between the large model and the feasible region in this invention.
[0059] Next, proceed to step S140.
[0060] In step S140, a structured prompt is constructed using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content; and a candidate strategy draft and the self-evaluation confidence of each candidate strategy draft are generated based on the structured prompt.
[0061] In this embodiment of the invention, the construction of structured prompts using the main scene, the set of sub-scenes, the physical and rule-feasible domains, and similar historical cases retrieved based on the current scene fingerprint vector as content; and the generation of candidate strategy drafts and self-assessed confidence scores for each candidate strategy draft based on the structured prompts, includes: retrieving several historical cases with the highest cosine similarity in the case library based on the current scene fingerprint vector as the similar historical cases, and recording the highest case similarity, where the highest case similarity is the maximum value among the cosine similarities between the current scene fingerprint vector and the scene fingerprint vectors of each historical case; using the main scene and the set of sub-scenes, The structured prompt is constructed using the following content: a verbal description of five types of constraints, several similar historical cases and their handling results and actual effects, and a format constraint in which the output is a structured strategy draft at the intent layer with supporting evidence for each item. The verbal description includes the current reduced capacity value, a list of available actions, and a list of mutually exclusive pairs. Using the structured prompt as input, the station management big model outputs several candidate strategy drafts and the self-assessed confidence level of each candidate strategy draft. The candidate strategy draft consists of several action items in sequence, and each action item includes the action object, action type, area of effect, start time and duration, and supporting evidence.
[0062] More specifically, the generation of candidate policies for large models under feasible region constraints includes the following steps: This step is used to prevent the large model from freely generating instructions under unconstrained conditions. Instead, it incorporates the linguistic description of the feasible domain, retrieved similar historical cases, and the main and sub-scene sets into the structured hints. The large model outputs a structured strategy draft for the intent layer, requiring each strategy to be supported by evidence and a self-assessed confidence level.
[0063] First, retrieve the most similar fingerprint from the case database based on the scenario fingerprint. A historical case: ... (8) in, The highest similarity score obtained in this search. For the first case in the case library Scene fingerprint vectors of historical cases; Cosine similarity; This is the fingerprint vector for the current scene.
[0064] Secondly, construct structured suggestions and generate candidate drafts: ... (9) in, The first generated for the large model Each candidate strategy draft consists of several action items in sequence. Each action item includes the action object, action type, area of effect, start time and duration. The self-assessment confidence level of the large model for this draft is given by [value]. ; The total number of candidate strategies; A large-scale model for station management; The structured suggestions include: the main scene and sub-scene sets, the verbal descriptions of feasible regions C1 to C5 (including the current reduced accessibility values, a list of available actions, and a list of mutual exclusion pairs), and the retrieved... The document outlines a format constraint that requires similar cases, their handling outcomes, and actual test results, as well as that the output must be a draft at the intent level with supporting evidence provided for each point.
[0065] Next, proceed to step S150.
[0066] In step S150, feasibility is determined and the feasibility rate is calculated based on the candidate strategy draft; wherein, candidate strategy drafts with a feasibility rate lower than the feasibility rate threshold are rejected as a whole and required to be regenerated, and candidate strategy drafts with a feasibility rate not lower than the feasibility rate threshold are projected and mapped into the physical and rule feasible domain, and the candidate strategy after projection repair is output.
[0067] In this embodiment of the invention, the feasibility determination and calculation of the feasibility rate based on the candidate strategy drafts are as follows: Candidate strategy drafts with a feasibility rate lower than the feasibility rate threshold are rejected entirely and required to be regenerated; candidate strategy drafts with a feasibility rate not lower than the feasibility rate threshold are projected onto the physical and rule-based feasible domain, and the projected and repaired candidate strategies are output. This includes: for each candidate strategy draft, its action items are judged sequentially according to four levels: syntax, permissions, topology interlocks, and physical mechanisms. First, the format of the action item and the device identifier are checked to see if they are parsable; then, the permission level required by the action item does not exceed the currently available authorization level; next, the action item is checked to see if it falls into a set of mutually exclusive action pairs with other action items in the same strategy, and whether it conflicts with the upstream and downstream states of the device; finally, the function of the action item is checked. Whether the partition satisfies capacity constraints and traffic capacity constraints in the extrapolation time domain; if any level fails, the feasibility judgment result of that action item is 0, and if all levels pass, it is 1; the ratio of the number of actions with a feasibility judgment result of 1 in a draft to the total number of actions in the draft is used as the feasibility rate of the draft; if the feasibility rate is lower than the feasibility rate threshold, the overall structure of the draft is determined to be invalid, rejected, and the station control big model is required to be regenerated according to the reason for rejection; if the feasibility rate is not lower than the feasibility rate threshold, then within the physical and rule feasible domain, the strategy with the smallest difference from the original candidate strategy draft is taken as the candidate strategy after projection repair. The difference is calculated by summing the equivalent modification costs for the three types of operations of adding, deleting, and modifying action items. The equivalent modification costs are taken from the station action cost table.
[0068] More specifically, feasible region projection and minimal modification repair include the following steps: This step is used to perform deterministic verification and repair on each draft output by S140.
[0069] First, determine the feasibility of each item: ... (10) in, For action items The feasibility assessment results; This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The judgment is performed sequentially at four levels: syntax form, permissions, topology interlocks, and physical mechanism. First, it checks if the action item format and device identifier are parsable; then it checks if its permission level does not exceed [the specified threshold]. Then check whether it falls into the set of mutually exclusive pairs with other actions of the same strategy. The system checks for conflicts with upstream and downstream states of the equipment, and finally verifies whether the functional partition satisfies capacity and throughput constraints within the extrapolated time domain. A failure at any level results in a value of 0.
[0070] ... (11) in, For draft The feasibility rate; For draft The total number of actions; For action items The feasibility assessment result. If ,in If the feasibility threshold is met, the overall structure of the draft is deemed invalid (rather than individual actions requiring adjustment), the draft is rejected, and the large model is required to be regenerated based on the reasons for rejection; if Then perform the minimum modification projection: ... (12) in, The first after projection restoration One candidate strategy; It is the minimum modification projection operator over the feasible region; This is the feasible region; Any strategy within the feasible domain; This is the original draft; To determine the difference between the two strategies, the equivalent modification costs for each of the three types of operations (addition, deletion, and modification) are calculated and then summed. The costs are taken from the station action cost table.
[0071] Most of the actions in the draft are generally reasonable, with only a few items going out of bounds (such as switching the passage direction of a certain gate group to bidirectional, but the hardware of that group does not support bidirectional). Discarding the whole item would discard the rest of the reasonable content and increase the latency of a large model call. The minimal changes retain the original draft's intent, only replacing out-of-bounds items with the equivalent action with the lowest cost within the feasible domain, thus maintaining the consistency of the processing logic.
[0072] Next, proceed to step S160.
[0073] In step S160, short-time domain deductions are performed on the candidate strategies after projection repair as the deduction objects to calculate the optimal candidate strategy and its comprehensive utility.
[0074] In this embodiment of the invention, the step of performing short-time domain deductions on the candidate strategies after projection repair as the deduction objects, and calculating the optimal candidate strategy and its comprehensive utility, includes: for each candidate strategy after projection repair, advancing several deduction steps according to the deduction step size starting from the current time, and calculating and iterating to the end of the deduction time domain according to the following relationship in sequence: obtaining the passenger flow density by dividing the number of people staying in the partition by the effective passage area of the partition; obtaining the average passenger flow speed by multiplying the free-flow walking speed by the difference between the passenger flow density and the congestion density, wherein the free-flow walking speed is the statistical average of the pedestrian free-flow speed measured at the station, and the congestion density is the density when the speed drops to zero; and calculating the passenger flow density, the average passenger flow speed, and the effective passage width of the partition by... The passenger flow rate is calculated by multiplying the passenger flow of the upstream adjacent zones and adding the external passenger flow inflow rate of that zone. The external passenger flow inflow rate is estimated from the ticket check records at the entrance, the passenger flow detection at the entrance and exit, and the number of passengers disembarking from the train. The passenger flow throughput rate is the smaller of the actual capacity and the number of passengers that can pass through. The number of passengers that can pass through is the sum of the number of stranded passengers divided by the projection step length and the passenger flow arrival rate. The number of stranded passengers is calculated by multiplying the difference between the passenger flow arrival rate and the passenger flow throughput rate by the projection step length. The number of stranded passengers in the next projection step is calculated by multiplying the difference between the passenger flow arrival rate and the passenger flow throughput rate by the projection step length. The relative speed decay rate is calculated by the positive part of the ratio of the difference between the average passenger flow speeds of two adjacent steps and the reference speed value. The reference speed value is the statistical average speed of the passenger flow during the normal off-peak period of that zone. The congestion risk index is obtained by multiplying the square of the ratio of passenger flow density to congestion critical density by the product of the relative speed decay rate and the speed decay sensitivity coefficient; the number of passengers passing through is obtained by summing the difference between the passenger flow of the partition in the simulation step when implementing the candidate strategy and the baseline flow when no intervention is taken, and multiplying it by the simulation step length over all simulation steps and all partitions; the average station-wide risk cost is obtained by averaging the congestion risk index of the partition in the simulation step when implementing the candidate strategy over all simulation steps and all partitions; the service interference cost is obtained by summing the products of the number of people affected by each action and the duration, wherein the number of people affected is calculated based on the average number of people staying in the partition affected by the action in the simulation time domain; the sum of the equivalent operating costs of each action is obtained. The equivalent operating cost is calculated from the station cost quota table according to the action type and duration. The equivalent operating cost is obtained by dividing the number of passengers, the average station-wide risk cost, the service interference cost, and the equivalent operating cost reference value by the reference value of the number of passengers, the risk index control target value, the service interference reference value, and the equivalent operating cost reference value, respectively, to obtain four normalized components. The reference value of the number of passengers is the station's capacity under the congestion critical density. The reference value of the service interference is the product of the capacity and the projected time domain length. The reference value of the equivalent operating cost is the product of the equivalent operating manpower available for a single shift and the shift minutes. The risk index control target value is specified in the station's safe operation procedures.The comprehensive utility is obtained by multiplying the efficiency weight by the normalized number of participants, the reduced risk weight by the normalized risk weight, the reduced service interference weight by the normalized service interference weight, and the reduced cost weight by the normalized cost weight. The sum of the four weights is one and all are non-negative, and their initial values are given by the station operation regulations. The candidate strategy with the largest comprehensive utility is selected as the optimal candidate strategy.
[0075] More specifically, twin inference and multi-objective utility optimization include the following steps: This step is used for the output of S150. Each feasible candidate is analyzed in a short time domain and the best one is selected.
[0076] Passenger flow propagation deduction (analytic and queuing hybrid model) for each candidate strategy ,since Starting step length Advance step: ... (13) in, For partitioning Passenger density (people / m²) 2 ); For partitioning The number of people stranded; For partitioning Effective passage area (m²) 2 ).
[0077] ... (14) in, For partitioning The average walking speed of the passenger flow (m / s); The free-flow walking speed (m / s) is determined by the statistical average of the measured free-flow pedestrian speeds at the station. Passenger flow density; Blockage density (persons / m³) 2 The density when the velocity drops to zero is obtained by extrapolation from the velocity-density scatter plot measured at the station.
[0078] ... (15) in, For partitioning Passenger flow (people / s); For partitioning Effective passage width (m).
[0079] ... (16) in, For partitioning Passenger arrival rate (people / s); For partitioning The set of upstream adjacent partitions; upstream partition Passenger flow (people / s); For partitioning The external passenger flow inflow rate (people / s) is estimated from the ticket check records at the entrance, passenger flow detection at the entrance and exit, and the number of passengers disembarking from the train.
[0080] ... (17) in, For partitioning Passenger throughput (people / s); Actual traffic capacity (people / s); The number of people stranded; To deduce the step size; Let this be the arrival rate (people / s). We take the smaller of the two values to express the physical fact that "the throughput is limited by both the upper limit of the channel capacity and the number of people that can pass through".
[0081] ... (18) in, The number of people stranded in the next simulation step. Equations (13) to (18) constitute a closed-loop calculation for one step of the simulation. By iterating step by step, the sequence of density, velocity, flow and number of people stranded in the entire simulation time domain can be obtained.
[0082] Risk index with velocity decay term (the core difference of Invention Point 5) ... (19) in, For partitioning The relative velocity decay rate; and These are the average walking speeds (m / s) of the passenger flow in the previous and current simulation steps, respectively. The reference value for travel speed (m / s) is the statistical average of travel speed during the normal off-peak period of this zone.
[0083] ... (20) in, For partitioning The congestion risk index; Passenger density (people / m²) 2 ); Congestion threshold density (people / m²) 2 ); The speed attenuation sensitivity coefficient is determined based on empirical values obtained from reviewing historical incidents at the station. This represents the relative velocity decay rate.
[0084] The direct precursor to stampedes is often not high density itself, but rather the stop-and-go waves and counter-currents that occur within high-density crowds—at this point, the density has not yet exceeded the limit, but the sudden drop in speed and the stop-and-go movement indicate localized flow instability. Risk indices constructed solely based on density (such as threshold determinations using existing technology) do not trigger alarms at this stage, missing the optimal intervention window; by introducing a relative velocity decay rate, the risk index becomes sensitive to flow instability and can trigger control measures before the density exceeds the limit.
[0085] Multi-objective utility and optimality ……(twenty one) in, Candidate strategies The increase in the number of people passing through the time domain relative to the no-intervention baseline; To execute candidate Time zone In the Step-by-step passenger flow (people / s); Baseline flow (people / s) without any intervention; To deduce the number of time-domain steps; Total number of partitions; To deduce the step length.
[0086] ……(twenty two) in, Candidate strategies The average total risk cost over the simulation time domain; To execute candidate Time zone In the The risk index of crowding at the foot of the road.
[0087] ……(twenty three) in, Candidate strategies Service disruption cost (person·s); For action items The number of passengers affected is calculated based on the average number of passengers stranded in the affected area during the simulation time domain. For action items Duration.
[0088] ……(twenty four) in, Candidate strategies Equivalent operating cost (equivalent manpower per minute); For action items The equivalent operating cost is calculated from the station cost quota table according to the action type and duration.
[0089] ... (25) in, Candidate strategies The overall effect; The four weights are efficiency, risk, service disruption, and cost, respectively, to meet the requirements. All values are non-negative, and the initial values are given by the station operation regulations; To obtain a reference value for the number of people, the maximum number of people that the entire station can accommodate at the critical congestion density is taken, i.e. ,in Effective passage area of the zone (m²) 2 ); The target value for risk index control is stipulated in the station's safe operation regulations; For the service interference reference value (person·s), take That is, the product of the station's critical capacity and the projected time domain length; The reference value for equivalent operating cost (equivalent manpower per minute) is calculated as the product of the equivalent operating manpower available for a single shift at the station and the number of shift minutes. The selection of the four denominators ensures that each component is normalized to a dimensionless quantity of the same magnitude, thereby making the four weights comparable and avoiding weight distortion caused by different dimensions.
[0090] ... (26) in, The optimal candidate strategy number; Candidate strategies The overall effect.
[0091] Existing technologies mostly solve for a single objective or solve offline after configuring simulation parameters once. This invention replaces large-space search with a small number of high-quality candidates generated by a large model, and replaces time-consuming solution with short-time domain analysis-queueing hybrid inference, thus enabling recalculation in a second-level rhythm as the situation progresses.
[0092] Next, proceed to step S170.
[0093] In step S170, the utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time, the switch is executed; automatic deployment is executed according to the interval based on the comprehensive confidence of the optimal candidate strategy, and the execution results are output.
[0094] In this embodiment of the invention, the utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is executed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic distribution is executed according to the comprehensive confidence of the optimal candidate strategy in intervals, and the execution results are output, including: subtracting the utility of the current execution strategy recalculated according to the same utility function and the current situation from the comprehensive utility of the optimal candidate strategy to obtain the utility gain; taking the product of the feasibility rate and a weighted term as the comprehensive confidence, the weighted term being obtained by weighting the self-assessed confidence and the highest case similarity according to the case support fusion coefficient; if and only When the utility gain is not lower than the switching hysteresis threshold and the difference between the current time and the last policy switching time is not less than the minimum dwell time, the current execution policy is replaced by the optimal candidate policy; otherwise, the current execution policy is maintained. Execution is selected based on the interval to which the comprehensive confidence level belongs: execution is automatically initiated when the comprehensive confidence level is not lower than the automatic execution threshold; execution is automatically initiated when the comprehensive confidence level is not lower than the manual confirmation lower limit but lower than the automatic execution threshold, and is pushed to the duty officer along with a chain of evidence; if the duty officer does not withdraw within the specified time, execution is automatic; execution is initiated when the comprehensive confidence level is lower than the manual confirmation lower limit, and is executed after explicit confirmation by the duty officer; the manual confirmation lower limit is less than the automatic execution threshold.
[0095] More specifically, the delayed handover and tiered authorization execution includes the following steps: ... (27) in, For utility gain; The overall utility of the optimal candidate strategy; The utility of the currently executing strategy is recalculated using the same utility function as the current situation.
[0096] The overall confidence level is calculated by combining the feasibility rate, the self-assessed confidence level of the large model, and the case support level using the following formula: ... (28) in, Candidate strategies The overall confidence level; This represents the feasibility rate of the candidate. Assess the self-confidence of the large model; The highest similarity found in the retrieved case; This is the case support fusion coefficient; a larger value indicates a greater reliance on the support strength of historical cases. Feasibility is introduced in a product form, meaning that the confidence level of a draft that requires extensive revisions to become feasible is proportionally lowered. Switching and authorization are performed according to the following rules: (R1) Switching condition: if and only if and Only when... Replace the current execution strategy; otherwise, maintain the original strategy. To switch the hysteresis threshold, For the minimum stay duration, This is the last time the strategy was switched.
[0097] (R2) Level 3 Authorization: When Automatically issued and executed; when The system will push the notification to the duty officer along with the chain of evidence. If the duty officer does not withdraw the notification within the specified time, it will be automatically executed. Actions must be explicitly confirmed by the duty officer before execution. To automatically apply the threshold, The lower limit is manually confirmed, and .
[0098] Once station control actions (changing signage on screens, changing announcements, and relocating staff) are implemented, it takes time for passengers to perceive them and for them to have a real effect. If strategies are switched every decision cycle due to minor differences in utility, on-site guidance information will be frequently reversed, creating confusion. Delays and pauses mean that switching only occurs when the new strategy is significantly better and the previous strategy has been fully effective.
[0099] In industrial and public safety scenarios, the most reliable role of large models is interpretation, supervision, and decision support, and key actions must retain a human confirmation step; this invention classifies according to confidence level, so that routine evacuation actions with high confidence can be executed automatically in seconds, while high-risk actions such as flow restriction, station closure, and evacuation are locked under human confirmation, achieving both response speed and accountability.
[0100] Next, proceed to step S180.
[0101] In step S180, the measured utility value is recalculated based on the measured data after execution, and the relative prediction deviation between the measured utility value and the predicted utility value is calculated; the symbolic side channel and the neural side channel are updated simultaneously using the relative prediction deviation as the driving quantity.
[0102] In this embodiment of the invention, the step of recalculating the measured utility value based on the actual measured data after execution, and calculating the relative prediction deviation between the measured utility value and the predicted utility value; simultaneously updating the symbolic side channel and the neural side channel with the relative prediction deviation as the driving force, includes: after the strategy has been executed for a full extrapolation time domain, recalculating the measured utility value according to the measured flow, density, number of people staying, and actual cost, taking the absolute value of the difference between the measured utility value and the predicted utility value, dividing it by the sum of the absolute value of the predicted utility value and the zero-prevention constant, to obtain the relative prediction deviation, where the predicted utility value is the comprehensive utility of the optimal candidate strategy calculated at the decision time; updating the weights in the symbolic side channel; using the difference between the predicted utility value and the measured utility value as the error, and the normalized component vector of the selected candidate as the gradient direction, updating the four weights according to the utility weight learning rate, and then projecting the updated weights onto the weight simplex; updating the throughput in the symbolic side channel; and satisfying the upstream conditions in the partition. Under the calibration conditions of continuous supply, non-zero number of people remaining in the area, and all available equipment, the designed traffic capacity is revised by multiplying the traffic capacity calibration learning rate by the difference between the measured pass rate and the designed traffic capacity; the case retrieval weights in the neural side channel are updated; the retrieval weights of historical cases are attenuated by the forgetting factor, and then the product of the cosine similarity between the current scene fingerprint vector and the scene fingerprint vector of the historical case, and a negative exponential term with the relative prediction deviation as the exponent is added; the prompt constraints in the neural side channel are updated; the situation snapshot of this period, the rejected action items and their reasons for rejection, the repaired action items and their repair costs are written into the constraint appendix of the prompt template; an audit record is generated; the audit record includes: situation snapshot and main / sub-scene judgment results, feasible domain parameters, original text of candidate strategy draft and supporting evidence, item-by-item feasibility judgment results and repaired items, each candidate utility and sub-item, confidence level and authorization level, authorizer, execution time and measured effect.
[0103] More specifically, bias-driven two-channel evolution includes the following steps: Bias calculation is performed by recalculating the utility based on measured data after the strategy has been executed for a full simulation time domain. ... (29) in, This represents the relative prediction bias; The predicted utility value obtained from the deduction, i.e., the value calculated at the decision time according to equation (25), is... ; The utility measured value is recalculated based on the actual measured data after execution. Its algorithm is the same as that of equation (25), but each item is calculated based on the actual flow, density, number of people staying and actual cost. To prevent zero constants, which are used to avoid numerical divergence caused by zero denominators, a normal value much smaller than the order of magnitude of the typical absolute value of utility is taken.
[0104] Symbolic side channel: Utility weight update. Since utility has a linear relationship with weights, the gradient can be given analytically: ... (30) in, Candidate strategies The normalized component vector; , , , These are the four normalized terms in equation (25), with the negative sign corresponding to the subtraction term in the utility.
[0105] ... (31) in, This is the weight vector; The learning rate is the utility weight; For the projection operator onto the weighted simplex, ensure that the update still satisfies And all terms are non-negative; The normalized component vector of the selected candidate.
[0106] If the predicted utility is higher than the actual utility ( If the projection is too optimistic, then the weight of the component that increases utility will be reduced, and the weight of the component that decreases utility will be increased, so that subsequent decisions will be more cautious about the overestimated benefits.
[0107] Symbol-side channel: Online calibration of traffic capacity when partitioning When the calibration conditions of continuous upstream supply, non-zero number of stranded personnel in the area, and full availability of all equipment under its jurisdiction are met, the design capacity shall be revised using the measured pass rate. ... (32) in, For partitioning Design throughput capacity (people / s); The learning rate is calibrated for the throughput capacity, and its value is taken from... ; For partitioning The measured pass rate (people / s). This revision directly writes back the feasible domain of S130, so that the actual capabilities after station renovation, equipment update and flow adjustment are gradually absorbed into hard constraints, overcoming the defects of static solidification of knowledge base and contingency plan base (6).
[0108] Neural Lateral Pathway: Case Studies and Tips Update ... (33) in, For example Search weight; The forgetting factor takes values of ; For current fingerprints and cases Cosine similarity of fingerprints; This represents the relative prediction bias. Factor Records where the handling effect deviates significantly from the expectation are entered into the case library with a lower weight, avoiding the repeated retrieval of a single failed handling as a high-priority experience. At the same time, the current situation snapshot, the rejected actions and their reasons for rejection, the corrected actions and their costs are written into the constraint appendix of the prompt template, so that the next round of generation will a priori avoid actions that have been disproven. This is the closing point of the bidirectional coupling closed loop of generation-projection-writeback.
[0109] Audit logging generates one audit record per decision-making cycle, including: a situation snapshot. With the main / sub-scene determination results and feasible region parameters (including) and This includes the original draft of the large model and supporting evidence, the feasibility assessment results for each item and the items to be corrected, the utility and sub-items of each candidate model, the confidence level and authorization level, the authorizing person, the execution time and the measured results. This record is unmodifiable and is used for subsequent review and accountability.
[0110] The necessity of channel coordination lies in the symbolic side (feasibility domain parameters, utility weights) and the neural side (case weights, cue constraint appendix), both determined by the same deviation. The driving force ensures that the revision directions of the two channels are consistent: an overly optimistic deviation in one inference increases the weight of risk items (symbolic side, interpretable and auditable) and decreases the retrieval weight of similar cases (neural side, changing the generation distribution). Only the combination of these two factors produces the effect of preventing the recurrence of similar errors. If only one aspect is addressed, the other side will still pull the system back to its original behavior pattern. This is the technical manifestation of the closed loop in this invention where the large model, feasible region, and inference are mutually input and output.
[0111] Figure 4 The present invention illustrates a station multi-scenario management and control system 400 based on a large model, including a partitioning module 410, a scenario module 420, a judgment module 430, and an update module 440.
[0112] The partitioning module 410 is used to divide the station into control zones and aggregate the collected data into situation features according to the zones to generate a station-wide situation tensor. The scene module 420 is used to generate a current scene fingerprint vector based on the overall situation tensor and its average rate of change within the observation window; and to obtain a set of main scenes and sub-scenes by comparing the current scene fingerprint vector with the prototype fingerprint vector of a predefined scene. Using the partitions involved in the main scene and the set of sub-scenes as constraints, a physical and rule-feasible domain is constructed; Furthermore, a structured prompt is constructed using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content; and a candidate strategy draft and the self-evaluation confidence of each candidate strategy draft are generated based on the structured prompt; The judgment module 430 is used to determine the feasibility and calculate the feasibility rate based on the candidate strategy draft; wherein, the candidate strategy draft with a feasibility rate lower than the feasibility rate threshold is rejected as a whole and required to be regenerated, and the candidate strategy draft with a feasibility rate not lower than the feasibility rate threshold is projected and mapped into the physical and rule feasible domain, and the candidate strategy after projection repair is output. Using the candidate strategies after projection repair as the inference objects, short-time domain inferences are performed respectively to calculate the optimal candidate strategy and its comprehensive utility. The utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is executed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic deployment is executed according to the comprehensive confidence of the optimal candidate strategy in intervals, and the execution results are output. The update module 440 is used to recalculate the measured utility value based on the measured data after execution, and calculate the relative prediction deviation between the measured utility value and the predicted utility value; and update the symbolic side channel and the neural side channel simultaneously with the relative prediction deviation as the driving quantity.
[0113] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A large model-based station multi-scene management and control method, characterized in that, The method includes the following steps: The station is divided into control zones, and the collected data is aggregated into situation features according to the zones to generate a station-wide situation tensor; Based on the overall situation tensor and its average rate of change within the observation window, a fingerprint vector for the current scene is generated; the fingerprint vector for the current scene is compared with the prototype fingerprint vector of a predefined scene to obtain a set of main scenes and sub-scenes. Using the partitions involved in the main scene and the set of sub-scenes as constraints, a physical and rule-feasible domain is constructed; Using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content, a structured prompt is constructed; based on the structured prompt, candidate strategy drafts and the self-evaluation confidence of each candidate strategy draft are generated; Feasibility is determined and the feasibility rate is calculated based on the candidate strategy drafts; wherein, candidate strategy drafts with a feasibility rate lower than the feasibility rate threshold are rejected as a whole and required to be regenerated, and candidate strategy drafts with a feasibility rate not lower than the feasibility rate threshold are projected and mapped into the physical and rule feasible domain, and the candidate strategy after projection repair is output. Using the candidate strategies after projection repair as the inference objects, short-time domain inferences are performed respectively to calculate the optimal candidate strategy and its comprehensive utility. The utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is executed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic deployment is executed according to the comprehensive confidence of the optimal candidate strategy in intervals, and the execution results are output. The measured utility value is recalculated based on the measured data after execution, and the relative prediction deviation between the measured utility value and the predicted utility value is calculated; the symbolic side channel and the neural side channel are updated simultaneously using the relative prediction deviation as the driving force.
2. The large model-based station multi-scene management and control method according to claim 1, characterized in that, The process of dividing the station into control zones and aggregating the collected data into situational features according to the zones to generate a station-wide situational tensor includes: The station is divided into several control zones according to the station's architectural plan and passenger flow lines, with the zone boundaries defined by physical cross-sections; wherein, the physical cross-sections include at least one of the following: turnstile groups, stairwells, and passageway intersections; Establish partition adjacency relationships and determine the upstream partition set for each partition; Collect data; preprocess the data and aggregate it into situational features by region to form the overall situational tensor of the station; wherein the data includes: video passenger flow detection data, AFC entry and exit records, gate passage count, security checkpoint queue length, escalator and elevator operating status, environmental sensor data, train arrival and departure times and platform occupancy information, and meteorological and surrounding activity information.
3. The large model-based station multi-scene management and control method according to claim 1, characterized in that, The current scene fingerprint vector is generated based on the overall situation tensor and its average rate of change within the observation window; Using the current scene fingerprint vector and the prototype fingerprint vector of a predefined scene as comparison objects, a set of main scenes and sub-scenes is obtained, including: The whole-station situation tensor and the average rate of change tensor of the whole-station situation tensor within the observation window are concatenated and input into the pre-trained scene fingerprint encoder to obtain the current scene fingerprint vector; the average rate of change tensor is the difference between the whole-station situation tensor at the current time and the whole-station situation tensor at the beginning of the observation window divided by the duration of the observation window. The cosine similarity is calculated based on the fingerprint vector of the current scene and the prototype fingerprint vector of each predefined scene. The scene membership degree of each predefined scene is obtained by taking the exponent of the cosine similarity divided by the temperature coefficient, and then using the sum of the cosine similarity over all predefined scenes as the denominator. The predefined scene with the highest scene membership degree is selected as the main scene; The set of sub-scenes is formed by taking the values of the adjacent elements of the scene dependency graph to represent predefined scenes that are coupled and whose scene membership degree is not lower than the lower limit of the membership degree of the linked sub-scenes. The scene dependency graph is jointly determined by the station topology and operation procedures. If two predefined scenes involve the same control zone or have a passenger flow transfer relationship, then the adjacent elements of the two predefined scenes are values that indicate that they are coupled.
4. The large model-based station multi-scene management and control method according to claim 1, characterized in that, The construction of a physically and rule-feasible domain, using the partitions involved in the main scene and the set of sub-scenes as constraints, includes: Taking all partitions involved in the main scene and the set of sub-scenes as the constraint range, the intersection of the following five types of constraints is taken as the physical and rule feasible domain, and recalculated periodically according to the equipment status and train arrival and departure schedule; A capacity constraint is set, requiring that the simulation density of any involved zone in any simulation step does not exceed the congestion critical density, which is determined by the station design documents and service level classification. A capacity constraint is set, requiring that the calculated flow of any involved zone in any calculation step does not exceed the actual capacity. The actual capacity is the product of the equipment status reduction factor and the design capacity. The equipment status reduction factor is the ratio of the number of available turnstiles, escalators, and passages under the jurisdiction of the zone to the rated number of available turnstiles, escalators, and passages, and is one when all equipment is available. The design capacity is taken from the station design documents. Set timing constraints, requiring that the effective delay of any action item does not exceed the maximum allowable effective delay, and that the effective time of the action falls within the simulation time domain. The maximum allowable effective delay is determined by the longest measured time of on-site broadcast response, guide screen switching and personnel arrival. Set mutual exclusion and interlocking constraints, requiring that no action pair in the same strategy belongs to the set of mutually exclusive action pairs; Set permissions and norms to require that the permission level required for any action item is no higher than the currently available authorization level. The permission level is divided according to the degree of direct impact of the action on personal safety, with the highest level being used for actions such as traffic restriction, site blocking, and evacuation. The physical and rule-feasible domain is formed by using candidate control strategies that simultaneously satisfy five types of constraints.
5. The large model-based station multi-scene management and control method according to claim 1, characterized in that, The structured prompt is constructed using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content; Based on the structured prompts, candidate strategy drafts and self-assessed confidence levels for each candidate strategy draft are generated, including: Based on the current scene fingerprint vector, retrieve the historical cases with the highest cosine similarity in the case library as the similar historical cases, and record the highest case similarity; the highest case similarity is the maximum value among the cosine similarities between the current scene fingerprint vector and the scene fingerprint vectors of each historical case. The structured prompt is constructed using the main scene and the set of sub-scenes, the verbal description of five types of constraints, several similar historical cases and their handling results and actual effects, and the output format as the draft of the intent layer structured strategy and the format constraints that provide evidence for each item. The verbal description includes the current reduced passage capacity value, the list of available actions, and the list of mutually exclusive pairs. Using the structured prompts as input, the station management big data model outputs several candidate strategy drafts and the self-assessed confidence of each candidate strategy draft. The candidate strategy drafts are composed of several action items in sequence. Each action item includes the action object, action type, area of effect, start time and duration, and supporting evidence.
6. The large model-based station multi-scene management and control method according to claim 1, characterized in that, The feasibility is determined and the feasibility rate is calculated based on the candidate strategy drafts; wherein, candidate strategy drafts with a feasibility rate lower than the feasibility rate threshold are rejected as a whole and required to be regenerated, and candidate strategy drafts with a feasibility rate not lower than the feasibility rate threshold are projected and mapped to the physical and rule-based feasible domain, and the projected and repaired candidate strategies are output, including: For each candidate strategy draft, its action items are judged sequentially according to four levels: syntax, permissions, topology interlocks, and physical mechanisms. First, it is checked whether the action item format and device identifier can be parsed. Then, it is checked whether the permission level required by the action item does not exceed the currently available authorization level. Next, it is checked whether the action item falls into the set of mutually exclusive action pairs with other action items in the same strategy, and whether it conflicts with the upstream and downstream states of the device. Finally, it is checked whether the action item's effective partition satisfies the capacity constraints and traffic capacity constraints in the extrapolation time domain. If any level fails, the feasibility judgment result of the action item is 0; if all levels pass, the result is 1. The feasibility rate of a draft is defined as the ratio of the number of actions with a feasibility assessment result of 1 to the total number of actions in that draft. If the feasibility rate is lower than the feasibility rate threshold, the overall structure of the draft is deemed invalid, the draft is rejected, and the station management big model is required to be regenerated based on the reasons for rejection. If the feasibility rate is not lower than the feasibility rate threshold, then within the physical and rule-based feasible domain, the strategy with the smallest difference from the original candidate strategy draft is selected as the candidate strategy after projection repair. The difference is calculated by summing the equivalent modification costs for the three types of operations: adding, deleting, and modifying action items. The equivalent modification costs are taken from the station action cost table.
7. The large model-based station multi-scene management and control method according to claim 1, characterized in that, The step of performing short-time domain inferences on the candidate strategies after projection repair, and calculating the optimal candidate strategy and its comprehensive utility, includes: For each candidate strategy after projection repair, several projection steps are advanced from the current time according to the projection step size. Each step is calculated and iterated to the end of the projection time domain according to the following relationship: the passenger flow density is obtained by dividing the number of people staying in the partition by the effective passage area of the partition; the average passenger flow speed is obtained by multiplying the free flow walking speed by the difference between the passenger flow density and the congestion density, where the free flow walking speed is the statistical average of the pedestrian free flow speed measured at the station, and the congestion density is the density when the speed drops to zero; the passenger flow density, the average passenger flow speed, and the effective passage area of the partition are used to calculate the passenger flow density. The passenger flow rate is obtained by multiplying the passage width, the three factors, and the passenger flow volume. The passenger arrival rate is obtained by adding the sum of the passenger flow of the upstream adjacent zones to the external passenger flow inflow rate of that zone. The external passenger flow inflow rate is estimated from the ticket check records at the entrance, the passenger flow detection at the entrance and exit, and the number of passengers disembarking from the train. The smaller value between the actual passage capacity and the number of passengers that can pass is taken as the passenger flow throughput rate. The number of passengers that can pass is the sum of the number of stranded passengers divided by the projection step length and the passenger flow arrival rate. The difference between the passenger flow arrival rate and the passenger flow throughput rate is multiplied by the projection step length and accumulated to the number of stranded passengers to obtain the number of stranded passengers in the next projection step. The relative speed decay rate is obtained by taking the positive part of the ratio of the difference between the average walking speed of two adjacent steps to the reference value of walking speed, where the reference value of walking speed is the statistical average of walking speed during the normal off-peak period of the zone; the congestion risk index is obtained by multiplying the square of the ratio of passenger flow density to congestion critical density by the product of the relative speed decay rate and the speed decay sensitivity coefficient. The difference between the passenger flow of a partition in the simulation step when implementing the candidate strategy and the baseline flow when no intervention is taken is accumulated over all simulation steps and all partitions and multiplied by the simulation step length to obtain the number of people passing through; the average station-wide risk cost is obtained by averaging the congestion risk index of a partition in the simulation step when implementing the candidate strategy over all simulation steps and all partitions. The service disruption cost is obtained by summing the product of the number of people affected by each action and the duration of the disruption. The number of people affected is calculated based on the average number of people staying in the affected area within the projected time domain. The equivalent operating cost is obtained by summing the equivalent operating costs of each action. The equivalent operating cost is calculated from the station cost quota table according to the action type and duration. The following four normalized components are obtained by dividing the number of passengers, the average station-wide risk cost, the service interference cost, and the equivalent operating cost by the reference value of the number of passengers, the target value of the risk index, the reference value of the service interference, and the reference value of the equivalent operating cost, respectively. The reference value of the number of passengers is the number of passengers that the station can accommodate under the critical congestion density. The reference value of the service interference is the product of the number of passengers that can be accommodated and the projected time domain length. The reference value of the equivalent operating cost is the product of the equivalent operating manpower that can be deployed per shift and the number of shift minutes. The target value of the risk index is specified by the station's safe operation procedures. The comprehensive utility is obtained by multiplying the efficiency weight by the normalized component of the number of people, the reduced risk weight by the normalized risk component, the reduced service interference weight by the normalized service interference component, and the reduced cost weight by the normalized cost component. The sum of the four weights is one and none of them are negative. The candidate strategy with the highest overall utility is selected as the optimal candidate strategy.
8. The station multi-scenario management and control method based on a large model according to claim 1, characterized in that, The utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time, the switch is executed. Based on the comprehensive confidence score of the optimal candidate strategy, automatic deployment is performed in intervals, and the execution results are output, including: The utility gain is obtained by subtracting the utility of the current execution strategy, which is recalculated using the same utility function and the current situation, from the comprehensive utility of the optimal candidate strategy. The overall confidence level is the product of the feasibility rate and a weighted term, whereby the weighted term is obtained by weighting the self-assessed confidence level and the highest case similarity according to the case support fusion coefficient. The current execution strategy shall be replaced by the optimal candidate strategy if and only if the utility gain is not lower than the switching hysteresis threshold and the difference between the current time and the last strategy switching time is not less than the minimum dwell time; otherwise, the current execution strategy shall be maintained. Execution is selected from the following intervals based on the overall confidence level: Execution is automatically initiated when the overall confidence level is not lower than the automatic execution threshold; Execution is automatically initiated when the overall confidence level is not lower than the manual confirmation lower limit but lower than the automatic execution threshold, and is pushed to the duty officer along with the evidence chain; if the duty officer does not withdraw the request within the specified time, execution is automatic; Execution is initiated when the overall confidence level is lower than the manual confirmation lower limit, and is executed after explicit confirmation by the duty officer; The manual confirmation lower limit is less than the automatic execution threshold.
9. The station multi-scenario management and control method based on a large model according to claim 1, characterized in that, The process involves recalculating the measured utility value based on the post-execution measured data and calculating the relative prediction deviation between the measured utility value and the predicted utility value; simultaneously updating the symbolic side channel and the neural side channel using the relative prediction deviation as the driving variable, including: After the strategy has been executed for a full simulation time domain, the measured utility value is recalculated based on the actual flow, density, number of people staying, and actual cost. The absolute value of the difference between the measured utility value and the predicted utility value is divided by the sum of the absolute value of the predicted utility value and the zero-prevention constant to obtain the relative prediction deviation. The predicted utility value is the comprehensive utility of the optimal candidate strategy calculated at the decision time. Update the weights in the symbol-side channel; using the difference between the predicted utility value and the measured utility value as the error, and the normalized component vector of the selected candidate as the gradient direction, update the four weights according to the utility weight learning rate, and then project the updated weights onto the weight simplex. Update the throughput capacity in the symbol-side channel; when the zone meets the calibration conditions of continuous upstream supply, non-zero number of people staying in the zone, and all equipment under its jurisdiction is available, revise the design throughput capacity by multiplying the throughput capacity calibration learning rate by the difference between the measured throughput rate and the design throughput capacity. Update the case retrieval weights in the neural side channel; after attenuating the retrieval weights of historical cases by the forgetting factor, add the product of the cosine similarity between the current scene fingerprint vector and the scene fingerprint vector of the historical case, and a negative exponential term with the relative prediction deviation as the exponent; Update the cue constraints in the neural side channel; write the current period's situation snapshot, rejected actions and their reasons for rejection, repaired actions and their repair costs into the constraint appendix of the cue template; Generate audit logs; the audit logs include: situation snapshot and main / sub-scenario determination results, feasible domain parameters, original text of candidate strategy drafts and supporting evidence, item-by-item feasibility determination results and items to be fixed, utility and sub-items of each candidate, confidence level and authorization level, authorizer, execution time and measured results.
10. A multi-scenario station management and control system based on a large model, characterized in that, The system includes: The partitioning module is configured to divide the station into control zones and aggregate the collected data into situation features according to the zones, generating a station-wide situation tensor. The scene module is configured to generate a current scene fingerprint vector based on the overall situation tensor and its average rate of change within the observation window; and to obtain a set of main scenes and sub-scenes by comparing the current scene fingerprint vector with the prototype fingerprint vector of a predefined scene. Using the partitions involved in the main scene and the set of sub-scenes as constraints, a physical and rule-feasible domain is constructed; Furthermore, a structured prompt is constructed using the main scene, the set of sub-scenes, the physical and rule-feasible domain, and similar historical cases retrieved according to the fingerprint vector of the current scene as content; and a candidate strategy draft and the self-evaluation confidence of each candidate strategy draft are generated based on the structured prompt; The judgment module is configured to determine the feasibility of the candidate strategy draft and calculate the feasibility rate; wherein, the candidate strategy draft with a feasibility rate lower than the feasibility rate threshold is rejected as a whole and required to be regenerated, and the candidate strategy draft with a feasibility rate not lower than the feasibility rate threshold is projected and mapped into the physical and rule feasible domain, and the candidate strategy after projection repair is output. Using the candidate strategies after projection repair as the inference objects, short-time domain inferences are performed respectively to calculate the optimal candidate strategy and its comprehensive utility. The utility gain is calculated based on the comprehensive utility of the optimal candidate strategy and the current execution strategy; a switch is executed when the utility gain is not lower than the switching hysteresis threshold and the time since the last strategy switch is not less than the minimum dwell time; automatic deployment is executed according to the comprehensive confidence of the optimal candidate strategy in intervals, and the execution results are output. The update module is configured to recalculate the measured utility value based on the measured data after execution, and calculate the relative prediction deviation between the measured utility value and the predicted utility value; and update the symbolic side channel and the neural side channel simultaneously with the relative prediction deviation as the driving quantity.
