Online self-optimization method and system for passenger flow prediction model

Abstract

The present application relates to a kind of passenger flow prediction model online self-optimization method and system.The passenger flow prediction model online self-optimization method includes: obtaining relevant multi-source associated data when receiving alarm event;Calibrate the position of alarm event according to multi-source associated data;Extract scene characteristic parameters according to the position after calibration;Scene-based online training is carried out to passenger flow prediction model based on scene characteristic parameters, to obtain the optimized model adapted to current scene;Based on the optimized passenger flow prediction model, obtain the passenger flow prediction result of current scene.The present application is helpful to realize the online self-optimization of passenger flow prediction model, adapt to the characteristics of the instantaneous change of urban public transport passenger flow scene, greatly improve the timeliness of passenger flow prediction model online training, provide high adaptability technical support for the real-time, precision of urban public transport operation scheduling.

Description

Technical Field

[0001] This invention relates to the field of traffic passenger flow forecasting technology, and more specifically, to an online self-optimization method and system for passenger flow forecasting models. Background Technology

[0002] With the development of urban public transportation networks and their scaling up, passenger flow continues to rise, and the spatiotemporal characteristics of passenger flow distribution are becoming increasingly complex. Sudden surges in passenger flow and congestion place higher demands on the real-time and accurate operation and scheduling of urban public transportation. Passenger flow forecasting technology, as the core support for public transportation operation and scheduling, can provide data references for capacity allocation, passenger flow management, and emergency response, and has become a key technological link in the construction of smart public transportation.

[0003] Among existing passenger flow forecasting technologies, deep learning-based prediction models have become one of the mainstream solutions due to their strong ability to capture spatiotemporal features. These solutions typically build models based on historical passenger flow data and perform offline training. The trained static prediction model is then applied to actual passenger flow forecasting scenarios. By mining the spatiotemporal correlation patterns of passenger flow data, it can predict the future scale and distribution of passenger flow. Compared to traditional statistical forecasting methods, it offers significant improvements in prediction accuracy and adaptability to complex scenarios. Correspondingly, its output passenger flow forecast results mainly extend from historical passenger flow patterns, focusing on predicting the scale and distribution of passenger flow during regular periods, which can meet the scheduling reference needs under stable daily operation scenarios.

[0004] In the process of developing this invention, the inventors discovered that current deep learning-based passenger flow prediction models adopt an "offline training, static application" model. After the passenger flow prediction model is trained, the weight parameters are fixed and no longer iterated, making it unable to adapt to the dynamic changes of actual passenger flow scenarios in real time. When facing large-scale event dispersal and other sudden passenger flow scenarios, the prediction accuracy drops significantly, resulting in a large deviation between the output passenger flow prediction results and the actual passenger flow situation. At the same time, if such prediction models are directly trained online, there is also a core problem of insufficient training timeliness, making it difficult to match the instantaneous changes in passenger flow scenarios. As a result, the optimized model cannot adapt to the current passenger flow scenario in a timely manner, and the output passenger flow prediction results cannot quickly respond to scenario changes. This makes it impossible to provide accurate data support for real-time capacity allocation and emergency passenger flow management, posing a serious challenge to the precise and real-time operation and scheduling of urban public transportation. Summary of the Invention

[0005] Based on this, and to address the aforementioned problems, this invention provides an online self-optimization method and system for passenger flow prediction models. By acquiring multi-source correlated data of alarm events, calibrating alarm locations and extracting scene feature parameters, conducting scenario-based online training of passenger flow prediction models, and adaptively adjusting edge deployment and training frequency, this method helps to achieve online self-optimization of passenger flow prediction models. It adapts to the instantaneous changes in urban public transportation passenger flow scenarios, significantly improves the timeliness of online training of passenger flow prediction models, and provides highly adaptable technical support for the real-time and accurate operation scheduling of urban public transportation.

[0006] In a first aspect, the present invention provides an online self-optimization method for a passenger flow prediction model. The online self-optimization method for the passenger flow prediction model includes: obtaining relevant multi-source correlation data when an alarm event is received; calibrating the location of the alarm event based on the multi-source correlation data; extracting scene feature parameters based on the calibrated location; performing scene-specific online training on the passenger flow prediction model based on the scene feature parameters to obtain an optimized model adapted to the current scene; and obtaining the passenger flow prediction result for the current scene based on the optimized passenger flow prediction model.

[0007] Optionally, in this embodiment of the invention, the extraction of scene feature parameters based on the calibrated location includes: extracting the inherent scene characteristics and real-time passenger flow status of the location based on the calibrated location; and converting the inherent scene characteristics and real-time passenger flow status into scene feature parameters through preset quantization rules. By extracting inherent scene characteristics and real-time passenger flow status through precise location association after calibration, and then converting them into standardized and computable scene feature parameters through preset quantization rules, this provides high-quality prior feature inputs that fit the actual scene for online training of the passenger flow prediction model. It also replaces the autonomous learning process of basic scene knowledge, significantly reducing the feature learning and convergence cost of the model, effectively improving the relevance and timeliness of the online training of the model, ensuring that the optimized model can adapt to the instantaneous changes in urban public transportation passenger flow scenarios in a timely manner, thereby improving the prediction accuracy in sudden passenger flow scenarios, and helping to achieve accurate matching of passenger flow prediction results with the real-time and accurate requirements of urban public transportation operation scheduling.

[0008] In the above implementation process, the inherent scene characteristics and real-time passenger flow status are transformed into scene feature parameters, including: transforming the inherent scene characteristics into evacuation difficulty coefficient, related scene coefficient, and clustering pattern coefficient, and transforming the real-time passenger flow status into a fluctuation amplitude coefficient; combining the evacuation difficulty coefficient, related scene coefficient, clustering pattern coefficient, and fluctuation amplitude coefficient for weighted calculation to obtain the scene comprehensive risk level value; and integrating the evacuation difficulty coefficient, related scene coefficient, clustering pattern coefficient, fluctuation amplitude coefficient, and scene comprehensive risk level value to form a structured scene feature parameter. By quantifying inherent scene characteristics into evacuation difficulty coefficients, related scene coefficients, and clustering pattern coefficients, and quantifying real-time passenger flow status into fluctuation amplitude coefficients, and then weighted calculations to obtain a comprehensive scene risk level value and integrating them into structured scene feature parameters, abstract scene characteristics and dynamic passenger flow status can be transformed into standardized features that the passenger flow prediction model can directly identify and calculate. This replaces the redundant learning process of the passenger flow prediction model for complex scene information, significantly reducing the computational cost and convergence time of the model training. Furthermore, by constructing a comprehensive scene risk level value, the model can accurately determine the degree of risk of passenger flow scenes, providing a clear and quantitative basis for the accurate selection of subsequent incremental training samples and the adaptation and adjustment of dynamic training strategies. This enables various optimization methods for online training of the passenger flow prediction model to accurately match passenger flow scenes with different risk levels and scene characteristics, effectively improving the targeting and efficiency of online training, shortening training time, and ensuring that the model can adapt to the instantaneous changes in urban public transportation passenger flow scenes in a timely manner.

[0009] Optionally, in this embodiment of the invention, the passenger flow prediction model is trained online in a scenario-specific manner based on scenario feature parameters to obtain an optimized model adapted to the current scenario. This includes: screening incremental passenger flow training samples based on scenario feature parameters to form a high-quality training sample set adapted to the current scenario; adjusting the passenger flow prediction model training strategy and algorithm training parameters based on scenario feature parameters to form a scenario-specific training configuration scheme; and using the high-quality training sample set and the scenario-specific training configuration scheme to perform targeted online training on the passenger flow prediction model to obtain an optimized model adapted to the current scenario. This approach not only eliminates redundant and invalid data through sample screening, significantly reducing the amount of training data processing and computation time, but also leverages scenario-based training configurations to ensure that the passenger flow prediction model focuses on the core characteristics and changing patterns of the current passenger flow scenario. This avoids indiscriminate full-parameter iteration, improving the targeting and efficiency of training and effectively addressing the pain point of insufficient timeliness in traditional online training. It ensures that the optimized model can adapt to the instantaneous changes in urban public transportation passenger flow scenarios in a timely manner. At the same time, this training method enables the passenger flow prediction model to quickly learn the spatiotemporal characteristics and evolution patterns of passenger flow in the current scenario, significantly improving its adaptability to sudden passenger flow scenarios such as the dispersal of large events, commuting peak connections, and the dispersal of shopping districts. It greatly improves the problem of low prediction accuracy and large result deviation of traditional models in sudden scenarios, enabling the output passenger flow prediction results to accurately match actual passenger flow changes.

[0010] Optionally, in this embodiment of the invention, the online self-optimization method of the passenger flow prediction model further includes: extracting the comprehensive risk level value of the scene from the scene feature parameters to determine the final risk level of the alarm event; establishing a mapping relationship of online training frequency based on the final risk level, and configuring differentiated initial online training frequencies for alarm events of different risk levels. By providing scene risk level as an objective basis, the training frequency is adapted to the actual risk level and change characteristics of the passenger flow scene, avoiding the waste of computing resources and increased energy consumption caused by indiscriminate training, and matching a higher initial training frequency for high-risk sudden passenger flow scenes, ensuring rapid iterative optimization of the model, accurately capturing the instantaneous change patterns of passenger flow, effectively improving the response speed and timeliness of online training in high-risk scenes, and solving the pain points of insufficient timeliness and low resource utilization efficiency of traditional online training.

[0011] Optionally, in this embodiment of the invention, upon receiving an alarm event, relevant multi-source associated data is obtained, including: collecting passenger flow operation data of stations and passage areas through real-time passenger flow monitoring; setting a preset passenger flow threshold, comparing the passenger flow operation data with the preset passenger flow threshold, and automatically triggering an alarm event when the passenger flow operation data exceeds the preset passenger flow threshold; recording the basic alarm information to obtain multi-source associated data spatiotemporally related to the alarm event. By using quantifiable passenger flow operation indicators as the basis for alarm judgment, the alarm triggering for abnormal scenarios such as sudden passenger flow gathering and congestion has an objective standard, effectively avoiding false alarms and missed alarms, ensuring timely capture of abnormal dynamics of urban public transportation passenger flow, and by collecting multi-source associated data spatiotemporally bound to the alarm event, eliminating irrelevant data interference, providing comprehensive, realistic, and actual abnormal scenario-appropriate raw data support for subsequent accurate alarm location calibration and scene feature parameter extraction, ensuring the accuracy, relevance, and effectiveness of the entire online self-optimization process of the passenger flow prediction model from the data source.

[0012] In the aforementioned implementation process, multi-source associated data includes GPS positioning, site geographic information, surveillance video, equipment distribution, site topology, and historical passenger flow data. This achieves multi-dimensional, full-scenario aggregation of alarm event-related data, integrating static basic data such as spatial location, geographic attributes, and hardware layout, while also incorporating dynamic associated data such as video dynamic monitoring and historical passenger flow patterns. This enables multi-source associated data to possess spatial completeness, temporal relevance, and scenario richness. Compared to the limitations of traditional solutions that rely solely on passenger flow data, this multi-source data system provides a multi-dimensional cross-validation data source for subsequent accurate alarm location calibration, effectively improving the accuracy of location calibration and avoiding scenario positioning errors caused by single data biases. Simultaneously, it provides comprehensive original evidence for scene feature parameter extraction, allowing for the mining of scene characteristics from multiple dimensions such as spatial layout, hardware conditions, historical patterns, and real-time dynamics. This ensures that the extracted scene feature parameters better match the real characteristics of urban public transportation passenger flow scenarios, significantly improving the accuracy and comprehensiveness of feature parameters.

[0013] Optionally, in this embodiment of the invention, the online self-optimization method of the passenger flow prediction model further includes: collecting on-site business processing data of alarm events, merging the on-site business processing data, scene feature parameters, and high-quality training sample sets into an incremental training sample library, and updating the sample library; when a new alarm event is triggered, repeatedly executing the optimization of the passenger flow prediction model based on the updated sample library. Through the continuous updating of the sample library and the cyclical execution of model optimization, the passenger flow prediction model can continuously learn the evolution patterns of different sudden passenger flow scenarios and the actual operational handling logic, achieving continuous iterative upgrades of prediction capabilities and completely breaking through the rigid limitations of traditional models' "offline training and static application".

[0014] Secondly, the present invention also provides an online self-optimization system for a passenger flow prediction model, the online self-optimization system for a passenger flow prediction model comprising: a data acquisition module, used to acquire relevant multi-source correlation data when an alarm event is received; a parameter extraction module, used to calibrate the location of the alarm event based on the multi-source correlation data; and extract scene feature parameters based on the calibrated location; a model training module, used to perform scenario-based online training on the passenger flow prediction model based on the scene feature parameters to obtain an optimized model adapted to the current scenario; and a passenger flow prediction module, used to obtain the passenger flow prediction result for the current scenario based on the optimized passenger flow prediction model.

[0015] Optionally, in this embodiment of the invention, the system further includes a model iteration optimization module, which is used to collect on-site business processing data of alarm events, integrate the on-site business processing data with scene feature parameters and high-quality training sample sets into an incremental training sample library, and update the sample library; when a new alarm event is triggered, the optimization of the passenger flow prediction model is repeatedly executed based on the updated sample library.

[0016] The online self-optimization method and system for passenger flow prediction models provided by this invention achieves online training of the passenger flow prediction model specifically for the alarm event scenario by acquiring multi-source correlated data of alarm events, calibrating alarm locations and extracting scene feature parameters, conducting scenario-based online training of the passenger flow prediction model, and adaptively adjusting edge deployment and training frequency. This facilitates online self-optimization of the passenger flow prediction model, adapts to the instantaneous changes in urban public transportation passenger flow scenarios, improves the timeliness of online training of the passenger flow prediction model, and provides highly adaptable technical support for the real-time and accurate operation scheduling of urban public transportation, effectively improving the overall reliability and performance of the system. To improve model training efficiency to meet the needs of online training, this invention provides corresponding optimization methods, which will be further described below. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the online self-optimization method for the passenger flow prediction model provided in the embodiments of the present invention; Figure 2 This is a flowchart of the scene feature parameter extraction method provided in the embodiments of the present invention; Figure 3This is a flowchart of the passenger flow prediction model training method provided in the embodiments of the present invention; Figure 4 A flowchart of the training frequency setting method provided in the embodiments of the present invention; Figure 5 This is a flowchart of the multi-source correlation data acquisition method provided in the embodiments of the present invention; Figure 6 This is a flowchart of the iterative optimization method for the passenger flow prediction model provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the online self-optimization system for the passenger flow prediction model provided in the embodiments of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will now be described with reference to the accompanying drawings. For example, the flowcharts and block diagrams in the drawings illustrate the architecture, functions, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, program segment, or part of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions. In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0020] Urban public transportation, as the core carrier of urban residents' travel, is a crucial infrastructure for alleviating urban traffic congestion, improving residents' travel efficiency, and promoting high-quality urban development. Its networked and large-scale development has become an important trend in modern urban construction. With the continuous agglomeration of urban population and the continuous expansion of urban space, the passenger flow of urban public transportation is experiencing explosive growth, and the spatiotemporal characteristics of passenger flow distribution are becoming increasingly complex. Situations such as tidal passenger flow during peak hours and sudden passenger congestion after large events occur frequently, placing extremely high demands on the real-time, accurate, and emergency response capabilities of public transportation operation scheduling. Passenger flow forecasting technology, as the core support of the public transportation operation scheduling system, can predict the scale, distribution, and evolution trend of passenger flow in advance, providing key data references for dynamic allocation of transport capacity, precise passenger flow guidance, and emergency response to sudden scenarios. It is a core technological link in realizing the construction of smart public transportation, and its forecasting accuracy and real-time performance directly determine the efficiency, safety, and service quality of public transportation operations.

[0021] Among existing urban public transportation passenger flow prediction technologies, deep learning-based prediction models have become the mainstream solution in the industry due to their powerful ability to capture spatiotemporal features. Their core working principle is as follows: using historical passenger flow data of urban public transportation as the core training sample, and combining auxiliary data such as station geographic information and station topology, a deep learning prediction model is constructed. Offline training and weight parameter optimization of the model are completed in the cloud using massive amounts of historical data. After the model training converges and the accuracy reaches the target, the static model is deployed to the public transportation operation monitoring system. In actual passenger flow prediction scenarios, real-time collected basic passenger flow data is directly input. Based on the trained weight parameters of the model, the spatiotemporal correlation patterns of passenger flow data are mined to predict information such as passenger flow scale, distribution, and peak nodes for specific future periods. Compared with traditional statistical prediction methods, this technology overcomes the limitations of linear analysis and can better fit the nonlinear spatiotemporal variation patterns of passenger flow. It shows significant improvements in prediction accuracy and adaptability to complex scenarios. Its output prediction results mainly extend from historical passenger flow patterns and can meet the scheduling reference needs of stable operation scenarios such as daily off-peak and regular peak periods.

[0022] The inventors discovered that existing deep learning passenger flow prediction models, which rely on an "offline training, static application" model, have fundamental flaws. They are unable to adapt to the dynamic changes in urban public transportation passenger flow scenarios and struggle with the timeliness challenges of direct online training. Consequently, they fail to meet the precise and real-time scheduling requirements of networked urban public transportation operations. Specifically, the drawbacks are: First, after training, the model's weight parameters remain fixed without an iterative optimization mechanism. It can only learn the inherent patterns of historical passenger flow and cannot adapt to the dynamic changes in actual passenger flow scenarios in real time. Prediction accuracy drops significantly when facing sudden passenger flow scenarios not fully covered by historical data. Second, when directly training such models online, the passenger flow scenarios are not considered. The design optimization strategy for the instantaneous change characteristics has a core problem of insufficient training timeliness. Data preprocessing and model iteration take too long, and often by the time the model optimization is completed, the corresponding passenger flow scenario has changed, and the optimized model cannot adapt to the current scenario in a timely manner. Third, the model training is disconnected from the actual operation and scheduling business. It only relies on passenger flow monitoring data to complete offline training without incorporating feedback data from on-site operation and disposal, which cannot form a closed-loop mechanism for model optimization, and the predictive ability is difficult to continuously improve. Fourth, the model deployment does not consider the computing power constraints of urban public transportation station / station edge devices. It adopts the cloud-based full-structure model design, which consumes a lot of computing resources and energy when training directly online, and cannot achieve efficient deployment and real-time optimization at the edge.

[0023] Based on this, the online self-optimization method for a passenger flow prediction model provided in this application helps to achieve online self-optimization of the passenger flow prediction model by acquiring multi-source correlation data of alarm events, calibrating alarm locations and extracting scene feature parameters, conducting scenario-based online training of the passenger flow prediction model, and adaptively adjusting edge deployment and training frequency. This adapts to the instantaneous changes in urban public transportation passenger flow scenarios, improves the timeliness of online training of the passenger flow prediction model, provides highly adaptable technical support for the real-time and accurate operation scheduling of urban public transportation, and effectively improves the overall reliability and performance of the system.

[0024] Please refer to Figure 1 , Figure 1 A flowchart of an online self-optimization method for a passenger flow prediction model provided in an embodiment of the present invention; the online self-optimization method for the passenger flow prediction model includes: Step S100: Obtain relevant multi-source correlation data when an alarm event is received.

[0025] In step S100 above, passenger flow monitoring system can collect passenger flow operation data in real time, including at least one of passenger flow density, passage area occupancy rate, and instantaneous passenger flow velocity in the station and passage area. Among them, passage area occupancy rate is the core quantitative indicator representing the degree of passenger flow congestion in urban public transportation scenarios. It refers to the proportion of open / semi-open passage space (such as entrance distribution area, drop-off area, waiting area, etc.) in public transportation stations (including bus hubs, passenger transport centers, etc.) that is actually occupied by passenger flow. Its calculation logic can be "the area actually occupied by passenger flow in the passage area at a specific time ÷ the total effective passage area of ​​the passage area", and the value range is set to 0-1 (the closer the value is to 1, the more congested the passage space is).

[0026] The passenger flow data is compared with a preset passenger flow threshold. When the passenger flow data exceeds the preset threshold, an alarm event is automatically triggered. At the same time, basic alarm information including the alarm time, initial alarm location, and type of passenger flow threshold exceeding the threshold is recorded. Simultaneously, a multi-source data acquisition interface is called to collect various multi-source associated data related to the alarm event in time and space, including GPS positioning data, site geographic information data, monitoring video frame data, equipment distribution data, site topology data, and historical passenger flow data from the same period. This completes the aggregation and integration of basic alarm information and multi-source associated data, forming a complete data set that can provide comprehensive data support for subsequent alarm location calibration and scene feature parameter extraction.

[0027] Step S200: Calibrate the location of the alarm event based on the multi-source correlation data; extract scene feature parameters based on the calibrated location.

[0028] In step S200 above, the multi-source associated data is preprocessed by format standardization and noise removal to eliminate format differences and interference data from different data sources. A multi-source data fusion algorithm can be used to perform weighted fusion calculation on the preprocessed multi-source associated data. Cross-validation is performed by combining the spatial coordinates of GPS positioning data with the visual features of monitoring video frame data and the spatial correlation of site topology data to correct the deviation of the initial alarm location and accurately locate the specific physical area where the alarm event occurred.

[0029] Based on the calibrated precise location, the inherent scene characteristics and real-time passenger flow status of that location are extracted synchronously and quickly. The inherent scene characteristics may include passenger flow evacuation difficulty determined by the width of the passage area, the number of entrances and exits, and the presence of obstacles; passenger flow aggregation patterns determined by historical data from the same period and real-time passenger flow growth rate; and surrounding related scenes such as whether it is near large venues, office buildings, or schools. The real-time passenger flow status is the dynamic change characteristics of passenger flow. Through preset quantification rules, the inherent scene characteristics and real-time passenger flow status are transformed into quantifiable feature parameters. The passenger flow evacuation difficulty is quantified into evacuation efficiency coefficient and related scene coefficient, and the passenger flow aggregation pattern is quantified into aggregation pattern coefficient. At the same time, the real-time passenger flow status is transformed into a fluctuation amplitude coefficient. The evacuation efficiency coefficient, related scene coefficient, aggregation pattern coefficient, and fluctuation amplitude coefficient are combined and weighted to obtain the scene comprehensive risk level value. Finally, the evacuation efficiency coefficient, aggregation pattern coefficient, related scene coefficient, fluctuation amplitude coefficient, and scene comprehensive risk level value are integrated to form a structured scene feature parameter. This feature parameter can be directly used as prior information for training the passenger flow prediction model, replacing the model's autonomous learning process of basic scene knowledge and the corresponding detailed training iteration process.

[0030] Step S300: Perform scenario-based online training on the passenger flow prediction model based on scenario feature parameters to obtain an optimized model that adapts to the current scenario.

[0031] In step S300 above, a precise mapping relationship between scene types and sample simplification rules is constructed based on the extracted structured scene feature parameters. The quantified scene features are transformed into programmable sample filtering logic to conduct refined filtering of incremental passenger flow training samples. Redundant and invalid data in non-alarm areas and without significant fluctuations are eliminated, and only effective data segments and peak data points strongly correlated with the current scene features are retained, forming a high-quality training sample set adapted to the current passenger flow scene. Subsequently, according to the passenger flow scene characteristics corresponding to the scene feature parameters, the training parameters of the online gradient descent algorithm (SGD) and the training priority of the ST-GNN+multi-scale Transformer passenger flow prediction model that has completed basic offline training are dynamically adjusted. For high-risk scenarios with high evacuation difficulty and drastic passenger flow fluctuations, the prediction weight of the model for passenger flow peaks is optimized and the number of training iterations for non-core parameters is reasonably reduced. For low-risk scenarios with low evacuation difficulty and gradual passenger flow changes, the training priority is appropriately reduced. The model is optimized to maintain training accuracy while retaining core parameters, forming a scenario-based training configuration scheme that fits the current scene. Simultaneously, the model is dynamically optimized for edge deployment constraints by incorporating scene feature parameters. This retains the spatiotemporal convolutional layer of ST-GNN responsible for capturing spatial dependencies in passenger flow and the core attention layer of multi-scale Transformer for mining long- and short-term temporal correlations, while eliminating redundant feature fusion branches and unnecessary activation modules. Furthermore, the model input dimension can be adjusted as needed based on the scene's comprehensive risk level, achieving a balance between edge device computing power constraints and model prediction accuracy. Finally, using this high-quality training sample set and scenario-based training configuration scheme, targeted online training is conducted on the lightweight optimized ST-GNN+multi-scale Transformer model, adjusting the model's weight parameters in real time. After training, an optimized model adapted to the current passenger flow scenario is obtained, ensuring that the model training time matches the instantaneous changes in the passenger flow scenario, achieving training synchronization with the scene.

[0032] Step S400: Based on the optimized passenger flow prediction model, obtain the passenger flow prediction results for the current scenario.

[0033] In step S400 above, the structured scene feature parameters extracted after calibration, along with core passenger flow operation data such as passenger flow density, passage area occupancy rate, and instantaneous flow velocity collected in real time under the current passenger flow scenario, are imported as inputs into the ST-GNN+multi-scale Transformer optimization model that has completed scenario-based online training. Relying on the weight parameters of this model after targeted training and optimization, the model fully leverages ST-GNN's ability to accurately capture the spatial topological relationships and basic temporal changes of urban public transportation passenger flow, and the multi-scale Transformer's ability to mine the long-term evolution patterns of passenger flow at multiple time scales such as minutes and hours, to optimize evacuation efficiency coefficients and aggregation... By deeply integrating prior feature parameters such as speed coefficient and associated scenario coefficient with real-time passenger flow data, the model accurately uncovers the core patterns of spatial distribution and transmission of passenger flow and multi-scale temporal evolution in the current scenario. This allows for the prediction of key information such as subsequent changes in passenger flow scale, peak periods, passenger flow density distribution in key areas, and passenger flow dispersal trends in the current passenger flow scenario. Ultimately, the model outputs refined passenger flow prediction results that closely match the actual characteristics of the current passenger flow scenario. These results can directly provide real-time and accurate data references for the dynamic allocation of transport capacity, precise passenger flow guidance, and emergency response to sudden scenarios in urban public transportation operations. At the same time, the model will simultaneously record relevant data from this prediction, preserving basic data for subsequent model iteration and optimization.

[0034] Therefore, the online self-optimization method for passenger flow prediction model provided by the embodiments of the present invention helps to realize the online self-optimization of passenger flow prediction model, adapts to the instantaneous changes in urban public transportation passenger flow scenarios, greatly improves the timeliness of online training of passenger flow prediction model, and provides highly adaptable technical support for the real-time and accurate operation scheduling of urban public transportation.

[0035] Please refer to Figure 2 , Figure 2 This is a flowchart of a scene feature parameter extraction method provided in one embodiment of the present invention; the scene feature parameter extraction method includes: Step S20: Based on the calibrated location, extract the inherent scene characteristics and real-time passenger flow status of the location.

[0036] In step S20 above, based on the precise physical location of the alarm event obtained by multi-source correlation data calibration, the inherent scene characteristics and real-time passenger flow status corresponding to that location are specifically extracted from the multi-source correlation data. Among them, the inherent scene characteristics are the spatial attribute features of the location itself that affect the distribution and evolution of passenger flow, which can be further decomposed into passenger flow evacuation-related attributes of evacuation difficulty coefficient, surrounding related scene attributes of correlation scene coefficient, and passenger flow aggregation-related attributes of aggregation pattern coefficient. The real-time passenger flow status is the dynamic change feature of passenger flow at that location when the alarm event occurs, which is the core basis for quantification into fluctuation amplitude coefficient. The extracted inherent scene characteristics and real-time passenger flow status completely cover all the basic information for subsequent conversion into various feature coefficients and calculation of the comprehensive risk level value of the scene, laying the data foundation for subsequent conversion into structured scene feature parameters through preset quantification rules.

[0037] Step S21: By using preset quantization rules, the inherent scene characteristics and real-time passenger flow status are transformed into scene feature parameters.

[0038] In step S21 above, the extracted inherent scene characteristics and real-time passenger flow status can be standardized and integrated according to the preset quantification rules. The preset quantification rules are quantifiable standardized transformation and calculation rules customized for the characteristics of urban public transportation passenger flow scenarios. They cover two core rules: feature coefficient quantification rules and scenario comprehensive risk level weighted calculation rules. In the specific implementation process: First, according to the feature coefficient quantification rules, the inherent scene characteristics can be converted into the corresponding evacuation difficulty coefficient, associated scene coefficient and aggregation pattern coefficient respectively. At the same time, the real-time passenger flow status can be converted into the fluctuation amplitude coefficient. Each coefficient is quantified into a standardized value with a unified dimension. Based on the preset weighted calculation rules, the evacuation difficulty coefficient, the associated scenario coefficient, the clustering pattern coefficient, and the fluctuation amplitude coefficient are assigned weights appropriate to the risk assessment of urban public transportation passenger flow scenarios. The four types of coefficients are then weighted and summed to obtain a comprehensive scenario risk level value that can accurately characterize the risk level of the current passenger flow scenario. Finally, the evacuation difficulty coefficient, associated scenario coefficient, clustering pattern coefficient, fluctuation amplitude coefficient, and comprehensive scenario risk level value are structurally integrated to form a structured scenario feature parameter that is dimensionally complete, numerically standardized, and can be directly recognized and calculated by the model. This provides accurate and computable prior feature basis for the subsequent scenario-based online training of the passenger flow prediction model.

[0039] Among them, the characteristic coefficient quantification rules are hierarchical assignment / formula calculation rules constructed based on urban public transportation industry standards and historical passenger flow scenario data. For example, for the evacuation difficulty coefficient, based on the actual values ​​of indicators such as the width of the passage area, the number of entrances and exits, and the situation of obstacles, a standardized coefficient in the range of 0-1 is obtained by calculation through a preset formula or hierarchical assignment. The higher the value, the greater the difficulty of passenger flow evacuation. For the correlation scenario coefficient, based on whether the surrounding area is close to large venues, office buildings, schools, and other passenger flow gathering and dispersing scenarios, and the degree of impact of such scenarios on the station's passenger flow, a standardized coefficient is obtained by hierarchical assignment. For the aggregation pattern coefficient, based on the aggregation pattern of historical passenger flow data and the actual value of real-time passenger flow growth rate, a standardized coefficient is obtained by calculation through a preset statistical formula. For the fluctuation amplitude coefficient, based on the real-time change rate and fluctuation range of indicators such as passenger flow density and instantaneous flow velocity during the alarm event period, a standardized coefficient is obtained by calculation through a preset variance / change rate formula, which intuitively represents the intensity of passenger flow dynamic fluctuations.

[0040] The weighted calculation rule for the comprehensive risk level of a scenario is a weighted summation rule constructed based on the degree of influence of various coefficients on the risk of passenger flow scenarios. First, it is trained using historical risk data of urban public transportation passenger flow scenarios to determine the weight ratio of evacuation difficulty coefficient, associated scenario coefficient, clustering pattern coefficient, and fluctuation amplitude coefficient in risk assessment (e.g., evacuation difficulty coefficient weight 0.35, clustering pattern coefficient weight 0.3, fluctuation amplitude coefficient weight 0.2, associated scenario coefficient weight 0.15). Then, it is calculated according to the preset formula: Comprehensive Risk Level Value of Scenario = Evacuation Difficulty Coefficient × Corresponding Weight + Associated Scenario Coefficient × Corresponding Weight + Clustering Pattern Coefficient × Corresponding Weight + Fluctuation Amplitude Coefficient × Corresponding Weight. The final comprehensive risk level value of the scenario is a standardized value that can be directly used for the classification and assessment of the risk level of passenger flow scenarios.

[0041] Therefore, the scene feature parameter extraction method provided by the embodiments of the present invention extracts inherent scene characteristics and real-time passenger flow status through precise location association after calibration, and then transforms them into standardized and computable scene feature parameters through preset quantization rules. This not only provides high-quality prior feature inputs that fit the actual scene for online training of passenger flow prediction models, but also replaces the autonomous learning process of basic scene knowledge, greatly reducing the feature learning and convergence cost of the model, effectively improving the relevance and timeliness of online training of the model, ensuring that the optimized model can adapt to the instantaneous changes in urban public transportation passenger flow scenarios in a timely manner, thereby improving the prediction accuracy in sudden passenger flow scenarios, and helping to achieve accurate matching of passenger flow prediction results with the real-time and accurate requirements of urban public transportation operation scheduling.

[0042] Please refer to Figure 3 , Figure 3 A flowchart of a passenger flow prediction model training method provided for an embodiment of the present invention; the passenger flow prediction model training method includes: Step S30: Based on scene feature parameters, filter the incremental passenger flow training samples to form a high-quality training sample set that is suitable for the current scene.

[0043] In step S30 above, a precise mapping relationship between scene types and sample simplification rules can be constructed based on structured scene feature parameters. The quantified evacuation difficulty coefficient, associated scene coefficient, clustering regularity coefficient, fluctuation amplitude coefficient, and scene comprehensive risk level value are transformed into programmable sample screening logic, clarifying the core dimensions and threshold conditions for sample screening. Based on this screening logic, the original incremental passenger flow training samples are finely screened. First, irrelevant samples outside the alarm area, redundant samples with no obvious fluctuation in passenger flow, and invalid samples with abnormal data collection are removed. Then, focusing on the current scene characteristics, samples of high / low evacuation difficulty scenes that match the evacuation difficulty coefficient, passenger flow clustering feature samples that match the clustering regularity coefficient, passenger flow dynamic fluctuation samples that correspond to the fluctuation amplitude coefficient, and peak data points and key time segments in high / medium / low risk scenes defined by the scene comprehensive risk level value are retained. Finally, the effective samples after screening are standardized and preprocessed (such as unifying data dimensions, filling in missing values, and removing outliers) to form a high-quality training sample set with sample dimensions that are highly consistent with the current passenger flow scene and controllable data quality.

[0044] Step S31: Adjust the passenger flow prediction model training strategy and algorithm training parameters based on scene feature parameters to form a scene-based training configuration scheme.

[0045] In step S31 above, firstly, based on the passenger flow scenario characteristics corresponding to the structured scenario feature parameters, and combined with the network architecture characteristics of the pre-completed basic offline training ST-GNN+multi-scale Transformer model, the training strategy of the passenger flow prediction model is specifically adjusted. Then, the core training parameters of the online gradient descent algorithm (SGD) are adapted and adjusted. The two are combined to form a scenario-based training configuration scheme adapted to the current passenger flow scenario. For scenarios with high overall risk levels, large evacuation difficulty coefficients, and high fluctuation amplitude coefficients, a training strategy focusing on core prediction objectives is adopted. The core layer of ST-GNN capturing spatial passenger flow distribution and the attention layer of the multi-scale Transformer mining minute-level short-term passenger flow fluctuations are optimized to improve the model's ability to predict passenger flow peaks and congestion points. Simultaneously, the SGD parameters are adjusted. The D algorithm parameters were optimized to increase the learning rate and reasonably reduce the number of training iterations for non-core model parameters from their default values, thus shortening training time while maintaining training accuracy. For scenarios with low overall risk levels, low evacuation difficulty coefficients, and low fluctuation amplitude coefficients, a lightweight training strategy was adopted to retain the training accuracy of core model parameters while appropriately reducing the overall training intensity. The learning rate of the SGD algorithm was lowered and the number of training iterations was further reduced to decrease computational resource consumption. For characteristic passenger flow scenarios with correlation coefficients and clustering pattern coefficients, the model's learning weights for the spatiotemporal evolution of passenger flow in such scenarios were specifically strengthened to make the model training more closely match the passenger flow characteristics of the current scenario. At the same time, all adjustments to training strategies and algorithm parameters took into account the computational constraints of edge devices at urban public transportation stations, ensuring that the configuration scheme was adapted to the training and operation environment at the edge.

[0046] Step S32: Using a high-quality training sample set and a scenario-based training configuration scheme, the passenger flow prediction model is trained online in a targeted manner to obtain an optimized model that is adapted to the current scenario.

[0047] In step S32 above, a high-quality training sample set is input into the ST-GNN+ multi-scale Transformer model that has completed basic offline training. Targeted online training is performed according to the scenario-based training configuration scheme. The model weight parameters are adjusted in real time through the SGD algorithm. The training is completed entirely by relying on the computing power of edge devices. The training process focuses on the core features of the current passenger flow, taking into account both training efficiency and prediction accuracy. After training is completed, an optimized model that can accurately adapt to the current passenger flow scenario is obtained, realizing real-time synchronization between training and scenario.

[0048] Therefore, the passenger flow prediction model training method provided by the embodiments of the present invention not only removes redundant and invalid data by filtering samples, which greatly reduces the amount of training data processing and computation time, but also relies on scenario-based training configuration to enable the passenger flow prediction model training to focus on the core features and changing patterns of the current passenger flow scenario, avoiding indiscriminate full-parameter iteration, improving the targeting and efficiency of training, effectively solving the pain point of insufficient timeliness of traditional online training, and ensuring that the optimized model can adapt to the instantaneous changes in urban public transportation passenger flow scenarios in a timely manner.

[0049] Please refer to Figure 4 , Figure 4 A flowchart illustrating a training frequency setting method provided in an embodiment of the present invention; the training frequency setting method includes: Step S40: Extract the scene comprehensive risk level value from the scene feature parameters to determine the final risk level of the alarm event.

[0050] In step S40 above, the weighted comprehensive risk level value of the scenario is accurately extracted from the structured scenario feature parameters. This value is a quantitative representation of the risk level of the passenger flow scenario. Subsequently, based on the risk management and control standards for urban public transportation passenger flow scenarios, multiple risk level classification thresholds are preset. The extracted comprehensive risk level value of the scenario is matched and compared with the thresholds, and it is classified into high, medium, and low levels according to the numerical range. This determines the final risk level of the alarm event and provides a clear and quantitative basis for the differentiated configuration of subsequent training frequencies.

[0051] Step S41: Establish a mapping relationship between online training frequencies based on the final risk level, and configure differentiated initial online training frequencies for alarm events of different risk levels.

[0052] In step S41 above, a one-to-one mapping relationship between risk levels and online training frequencies can be established based on the high, medium, and low final risk levels defined by the alarm events. Differentiated initial training frequencies are then adapted according to the degree of risk. The highest frequency is configured for high-risk levels, the regular frequency for medium-risk levels, and the lowest frequency for low-risk levels. This ensures that the training frequency matches the risk and changing characteristics of the passenger flow scenario, balancing rapid model iteration with energy consumption control, and avoiding resource waste caused by indiscriminate training.

[0053] Therefore, the training frequency setting method provided by the embodiments of the present invention, by providing objective basis based on the risk level of the scenario, allows the training frequency to be adapted to the actual risk level and change characteristics of the passenger flow scenario, avoiding the waste of computing resources and increased energy consumption caused by indiscriminate training, and matching a higher initial training frequency for high-risk sudden passenger flow scenarios, ensuring rapid iteration and optimization of the model, accurately capturing the instantaneous change pattern of passenger flow, effectively improving the response speed and timeliness of online training in high-risk scenarios, and solving the pain points of insufficient timeliness and low resource utilization efficiency of traditional online training.

[0054] Please refer to Figure 5 , Figure 5 A flowchart of a multi-source associated data acquisition method provided for an embodiment of the present invention; the multi-source associated data acquisition method includes: Step S50: Collect passenger flow data at stations and in passage areas through real-time passenger flow monitoring.

[0055] In step S50 above, real-time data collection across the entire area can be carried out by relying on passenger flow monitoring equipment deployed at urban public transportation stations and passage areas. The dynamics of passenger flow at each monitoring point are continuously captured. The collected passenger flow operation data covers at least one of passenger flow density, passage area occupancy rate, and instantaneous passenger flow velocity. All data are updated in real time at a uniform time granularity, fully reflecting the actual operation status of passenger flow at stations and passage areas, and providing objective and quantifiable basic data support for subsequent alarm event triggering.

[0056] Step S51: Set a preset passenger flow threshold, compare the passenger flow operation data with the preset passenger flow threshold, and automatically trigger an alarm event if the passenger flow operation data exceeds the preset passenger flow threshold.

[0057] In step S51 above, firstly, based on the urban public transportation station operation specifications and historical passenger flow congestion data, corresponding preset passenger flow thresholds can be set for various passenger flow operation indicators such as passenger flow density and passage area occupancy rate, forming a multi-dimensional threshold judgment system; then, the real-time collected passenger flow operation data is compared with the corresponding preset thresholds one by one. If any indicator value exceeds the preset threshold, the system will automatically trigger an alarm event to promptly capture abnormal passenger flow status and provide trigger signals for subsequent passenger flow scenario handling and model optimization.

[0058] Step S52: Record the basic alarm information and obtain multi-source correlation data that is spatiotemporally related to the alarm event.

[0059] In step S52 above, basic alarm information of alarm events can be automatically recorded, including the event trigger time, initial location, and the type and value of passenger flow operation indicators exceeding the threshold. At the same time, based on the spatiotemporal dimension of this basic information, multi-source associated data related to spatiotemporal aspects can be retrieved from the urban public transportation operation data platform, covering GPS positioning, station geographic information, surveillance video, equipment distribution, station topology, and historical passenger flow data of the same period, forming a complete multi-source associated dataset of alarm events, providing comprehensive data support for subsequent alarm location calibration and scene feature parameter extraction.

[0060] Therefore, the multi-source associated data acquisition method provided by the embodiments of the present invention uses quantifiable passenger flow operation indicators as the basis for alarm judgment, so that the alarm triggering of abnormal scenarios such as sudden passenger flow gathering and congestion has an objective standard, effectively avoiding false alarms and missed alarms, and ensuring that abnormal dynamics of urban public transportation passenger flow can be captured in a timely manner. Furthermore, by collecting multi-source associated data that is spatiotemporally bound to alarm events and eliminating irrelevant data interference, it provides comprehensive, realistic and actual abnormal scenario raw data support for subsequent accurate alarm location calibration and scene feature parameter extraction, ensuring the accuracy, pertinence and effectiveness of the entire process of online self-optimization of passenger flow prediction model from the data source.

[0061] Please refer to Figure 6 , Figure 6 A flowchart of an iterative optimization method for a passenger flow prediction model provided in an embodiment of the present invention; the iterative optimization method for a passenger flow prediction model includes: Step S60: Collect on-site business processing data of alarm events, and integrate the on-site business processing data, scene feature parameters and high-quality training sample set into the incremental training sample library to update the sample library.

[0062] In step S60 above, on-site business processing data during the handling of alarm events can be comprehensively collected, covering actual operational measures such as capacity allocation and passenger flow management, as well as corresponding passenger flow change feedback data. Subsequently, this on-site business processing data is integrated with the structured scene feature parameters extracted from this alarm scenario and the high-quality training sample set used for training, and uniformly incorporated into the incremental training sample library to complete the real-time update of the sample library, enrich the relevant data of sudden passenger flow scenarios in the sample library, and provide real data support that fits actual operation for the subsequent iterative optimization of the model.

[0063] Step S61: In the event of a new alarm event, the optimization of the passenger flow prediction model is repeated based on the updated sample library.

[0064] In step S61 above, when a new passenger flow alarm event is triggered, the updated incremental training sample library can be retrieved. Based on this data, the entire process of optimization, from multi-source associated data acquisition and scene feature parameter extraction to scenario-based online training of the passenger flow prediction model, can be repeatedly executed. This allows the passenger flow prediction model to iterate in a continuously rich array of real-world scene samples, thereby continuously improving its adaptability and prediction accuracy for various sudden passenger flow scenarios and achieving self-optimization and upgrading of the passenger flow prediction model.

[0065] Therefore, the passenger flow prediction model iterative optimization method provided by the embodiments of the present invention enables the passenger flow prediction model to continuously learn the evolution law of different sudden passenger flow scenarios and the actual operation and handling logic through continuous updating of the sample library and cyclical execution of model optimization, thereby achieving continuous iterative upgrade of prediction capabilities and completely breaking through the rigid limitations of traditional models that are "offline trained and statically applied".

[0066] Please refer to Figure 7 , Figure 7 This is a schematic diagram of an online self-optimization system for a passenger flow prediction model provided in an embodiment of the present invention. The system includes: a data acquisition module 10, a parameter extraction module 20, a model training module 30, a passenger flow prediction module 40, and a model iteration optimization module 50. Specifically, the data acquisition module 10 obtains relevant multi-source correlation data upon receiving an alarm event; the parameter extraction module 20 calibrates the location of the alarm event based on the multi-source correlation data and extracts scene feature parameters based on the calibrated location; the model training module 30 performs scenario-based online training on the passenger flow prediction model based on the scene feature parameters to obtain an optimized model adapted to the current scenario; the passenger flow prediction module 40 obtains the passenger flow prediction result for the current scenario based on the optimized passenger flow prediction model; and the model iteration optimization module 50 collects on-site business processing data related to alarm events, integrates the on-site business processing data, scene feature parameters, and a high-quality training sample set into an incremental training sample library for updating the sample library; and, upon triggering a new alarm event, repeatedly optimizes the passenger flow prediction model based on the updated sample library.

[0067] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The system implementations described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some communication interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0068] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0069] Furthermore, in the various embodiments of the present invention, the functional modules can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0070] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0071] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An online self-optimization method for a passenger flow prediction model, characterized in that, The method includes: Upon receiving an alarm event, obtain relevant multi-source correlation data; The location of the alarm event is calibrated based on the multi-source correlation data; scene feature parameters are extracted based on the calibrated location. The passenger flow prediction model is trained online in a scenario-specific manner based on the scenario feature parameters to obtain an optimized model that is adapted to the current scenario. Based on the optimized passenger flow prediction model, the passenger flow prediction result for the current scenario is obtained.

2. The method according to claim 1, characterized in that, The step of extracting scene feature parameters based on the calibrated location includes: Based on the calibrated location, the inherent scene characteristics and real-time passenger flow status of the location are extracted; By using preset quantization rules, the inherent scene characteristics and real-time passenger flow status are transformed into scene feature parameters.

3. The method according to claim 2, characterized in that, The process of converting the inherent scene characteristics and real-time passenger flow status into scene feature parameters includes: The inherent scene characteristics are transformed into evacuation difficulty coefficient, related scene coefficient, and clustering pattern coefficient, and the real-time passenger flow status is transformed into a coefficient of fluctuation amplitude. The overall risk level of the scenario is obtained by weighting the evacuation difficulty coefficient, the associated scenario coefficient, the aggregation pattern coefficient, and the fluctuation amplitude coefficient. The evacuation difficulty coefficient, associated scenario coefficient, clustering pattern coefficient, fluctuation amplitude coefficient, and overall scenario risk level value are integrated to form a structured scenario characteristic parameter.

4. The method according to claim 1, characterized in that, The step of performing scenario-based online training on the passenger flow prediction model based on the scenario feature parameters to obtain an optimized model adapted to the current scenario includes: The incremental passenger flow training samples are filtered based on the scene feature parameters to form a high-quality training sample set that is suitable for the current scene. The training strategy and algorithm training parameters of the passenger flow prediction model are adjusted based on the scene feature parameters to form a scene-based training configuration scheme. Using the high-quality training sample set and scenario-based training configuration scheme, the passenger flow prediction model is trained online in a targeted manner to obtain an optimized model adapted to the current scenario.

5. The method according to claim 1, characterized in that, The method further includes: Extract the overall risk level value of the scene from the scene feature parameters to determine the final risk level of the alarm event; A mapping relationship for online training frequencies is established based on the final risk level, and differentiated initial online training frequencies are configured for alarm events of different risk levels.

6. The method according to claim 1, characterized in that, The process of obtaining relevant multi-source correlation data upon receiving an alarm event includes: Passenger flow data is collected at stations and passageways through real-time passenger flow monitoring. A preset passenger flow threshold is set, and the passenger flow operation data is compared with the preset passenger flow threshold. If the passenger flow operation data exceeds the preset passenger flow threshold, an alarm event is automatically triggered. Record the basic alarm information to obtain multi-source correlation data that is spatiotemporally related to the alarm event.

7. The method according to claim 6, characterized in that, in, The multi-source associated data includes GPS positioning, site geographic information, surveillance video, equipment distribution, site topology, and historical passenger flow data for the same period.

8. The method according to claim 1, characterized in that, The method further includes: Collect on-site business processing data of alarm events, and integrate the on-site business processing data, scene feature parameters and high-quality training sample sets into the incremental training sample library to update the sample library. In the event of a new alarm event, the optimization of the passenger flow prediction model is repeated based on the updated sample library.

9. An online self-optimization system for a passenger flow prediction model, characterized in that, include: The data acquisition module is used to obtain relevant multi-source correlation data when an alarm event is received; The parameter extraction module is used to calibrate the location of the alarm event based on the multi-source correlation data; and to extract scene feature parameters based on the calibrated location. The model training module is used to perform scenario-based online training on the passenger flow prediction model based on the scenario feature parameters, so as to obtain an optimized model that adapts to the current scenario. The passenger flow prediction module is used to obtain the passenger flow prediction results for the current scenario based on the optimized passenger flow prediction model.

10. The system according to claim 9, characterized in that, The system also includes a model iteration and optimization module, which collects on-site business processing data of alarm events, integrates the on-site business processing data with scene feature parameters and high-quality training sample sets into an incremental training sample library, and updates the sample library; when a new alarm event is triggered, the optimization of the passenger flow prediction model is repeatedly executed based on the updated sample library.

Images