A dynamic optimization method and system for rail transit station connection resources

Abstract

The application discloses a kind of dynamic optimization method and system of rail transit station connection resource, the method includes: obtaining the mobile trajectory data and public transport travel record of user;According to mobile trajectory data, construct the heterogeneous graph that describes the association between individual, geographical unit and time sequence segment of user;The behavior representation vector of user node in heterogeneous graph is extracted using graph neural network, and the behavior representation vector is mapped as the group attribute label of user;According to group attribute label and public transport travel record, the running efficiency value of each station and global service accessibility difference degree are determined;Simulate to cancel each station and recalculate service accessibility difference degree, according to difference degree change amount determine the marginal influence coefficient of each station to global service balance;Each station is configured as decision-making agent, to generate connection resource adjustment instruction;Adjustment instruction is issued to field execution terminal, to drive connection facility to execute the change of transport capacity, capacity or rate, significantly improve precision.

Description

Technical Field

[0001] This invention relates to the field of transportation connection technology, specifically to a dynamic optimization method and system for rail transit station connection resources. Background Technology

[0002] Connectivity resources for urban rail transit stations include bus routes, shared bicycle parking spaces, parking lots, and dynamic pricing. The scientific allocation of these resources requires ensuring overall passenger flow efficiency while also considering service balance across different areas.

[0003] However, in existing technologies, socioeconomic attributes such as residents' income that reflect the travel characteristics of different groups mainly rely on population censuses or special traffic surveys. Such surveys have long update cycles and high implementation costs, and cannot reflect the actual changes in the spatial distribution of residents in daily life. The fairness assessment indicators calculated based on the above static data are also limited to offline macro analysis and are difficult to dynamically match with daily fluctuating passenger flow data. Moreover, existing resource optimization methods are mostly offline preset scenario solutions, and when the actual passenger flow demand changes, the generated configuration scheme cannot respond and adjust in real time.

[0004] Furthermore, the allocation of urban rail transit connection resources is essentially a multi-objective dynamic optimization problem, in which there is an inherent tension between "operational efficiency" and "service balance": maximizing operational efficiency tends to concentrate public transport capacity and parking resources at high-demand, high-passenger-flow stations, while maximizing service balance requires tilting resources towards low-coverage, peripheral areas. Mathematically, these two constitute a multi-objective conflict. Existing methods lack modeling and quantification tools for this conflict, typically using weighted summation or constraint transformation to simplify the multi-objective problem into a single objective. However, once the weight coefficients are fixed, they cannot be dynamically adjusted with changes in the spatiotemporal distribution of passenger flow, making it impossible to achieve a Pareto optimal dynamic solution for efficiency and fairness under different operating scenarios.

[0005] The cumulative effect of the above problems makes it difficult for existing technologies to achieve synergistic optimization of connection resource operation efficiency and service balance in a dynamically changing environment. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention provides a dynamic optimization method and system for rail transit station connection resources to solve the problems in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A dynamic optimization method for rail transit station connection resources includes: Obtain users' movement trajectory data and public transportation travel records; Based on the movement trajectory data, a heterogeneous map is constructed, which is used to describe the spatiotemporal relationships between individual users, geographical units, and time series segments. A graph neural network is used to extract the behavioral representation vectors of user nodes in the heterogeneous graph, and the behavioral representation vectors are mapped to the group attribute labels of users. Based on the group attribute tags and the public transportation travel records, determine the operational efficiency value of each station and the global service accessibility difference. The simulation removes each station from the rail transit network one by one, and recalculates the service accessibility difference after removal. The increment of the service accessibility difference before and after removal is used as the marginal impact coefficient of each station on the global service equilibrium. Each site is configured as a decision agent, and the marginal impact coefficient corresponding to each site is used as the weight factor of the fairness sub-item in the reinforcement learning reward function of the decision agent. The strategy is iterated with the goal of co-optimizing the running efficiency value and the service reachability difference, and the connection resource adjustment actions of each site are generated. The adjustment action is converted into an adjustment command and sent to the on-site execution terminal.

[0008] In one embodiment, the heterogeneous spectrum includes: User nodes and geographic unit nodes are connected by access relationship edges formed based on the movement trajectory data; geographic unit nodes are connected by adjacency edges formed based on spatial topological adjacency relationships. The initial characteristics of the user node include at least travel frequency and nighttime dwell stability, and the initial characteristics of the geographic unit node include at least urban functional density.

[0009] In one embodiment, the graph neural network extracts features in the following manner: Aggregate the features of user nodes sharing the same geographic unit within the same spatiotemporal slice, and output the spatial distribution density of the group; Aggregate the features of the same user node across adjacent time-series segments and output the temporal evolution features of individual behavior.

[0010] In one embodiment, the behavior representation vector is mapped to group attribute labels using multi-class supervised fine-tuning; the loss function of the fine-tuning includes a class weight factor, which is negatively correlated with the proportion of each class in the labeled samples.

[0011] In one embodiment, the global service reachability dissimilarity is determined in the following way: Based on the road network topology and public transportation capacity timetable, determine the number of stations that can be reached within the preset travel time threshold for each geographical unit, and use this number as the service coverage of that geographical unit. Residents within a geographic unit are sorted according to the group attribute labels, and a corresponding curve between the cumulative percentage of residents and the cumulative percentage of service coverage is constructed. Calculate the area of ​​the corresponding curve that deviates from the equilibrium reference line, and use the area to characterize the global service reachability difference.

[0012] In one embodiment, the marginal impact coefficient corresponding to each station is used as a weighting factor for the fairness sub-item in the reinforcement learning reward function of the decision-making agent, including: The reward function includes an efficiency sub-item, a fairness sub-item, and a cost sub-item; the weight coefficient of the fairness sub-item is positively correlated with the marginal influence coefficient of the corresponding station of the decision agent, so that the station with the larger marginal influence coefficient bears a higher responsibility for fairness optimization in the strategy iteration.

[0013] In one embodiment, the decision-making agent employs a centralized training and distributed execution architecture, including: During the training phase, the actions and global state changes of each decision-making agent are collected centrally to update the global evaluation network; During the execution phase, each decision-making agent independently outputs actions based on local observation information; the local observation information includes the efficiency value of the local station, the efficiency values ​​of neighboring stations, the density of surrounding populations, and the current connection resource configuration parameters.

[0014] In one embodiment, the connection resource adjustment action includes adjusting the bus capacity, adjusting the shared bicycle capacity, and adjusting the parking fee rate. When each decision-making agent outputs the action, it is constrained by the total global resources. The constraint is achieved by superimposing a penalty term on the reward function. The penalty term is positively correlated with the degree to which the total resource adjustment of each site exceeds the preset upper limit.

[0015] In one embodiment, before converting the adjustment action into an adjustment command and issuing it, the adjustment action within a continuous time window is subjected to a first-order low-pass filter; An execution status monitoring loop is set up so that if any station fails to respond effectively to the adjustment command within a preset number of consecutive cycles, the station is switched to a preset backup operation mode.

[0016] A dynamic optimization system for public transportation station connection resources includes: The data access unit is used to receive mobile signaling data streams and public transportation transaction data streams; The map management unit is used to construct and maintain heterogeneous maps between individual users, geospatial units, and time series segments based on the data from the data access unit. The attribute-aware unit has a built-in pre-trained and fine-tuned graph neural network model, which is used to output the group attribute classification of users based on the heterogeneous graph. The equilibrium measurement unit is used to calculate the global service equilibrium degree based on the distribution of the group attribute classification on the geographic unit and the service coverage of each site, and to recalculate the global service equilibrium degree after simulating the removal of each site to determine the marginal contribution value of each site. The collaborative decision-making unit has a built-in multi-agent policy network with each site as an independent decision-making entity. It uses the marginal contribution value of each site as the weight factor of the balance sub-item in the reward function of the agent, and performs policy iteration with global operating efficiency and global service balance as the collaborative optimization objectives to output a connection resource adjustment scheme. The conversion unit is used to convert the adjustment scheme into a drive signal that is compatible with the bus dispatch terminal, the shared bicycle control terminal, and the parking fee terminal.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention utilizes heterogeneous graphs to fuse users' spatiotemporal behavioral association information and achieves information sharing among similar groups through the message passing mechanism of graph autoencoders. Under the condition of extremely small labeled sample size (less than 1% of the total number of users), it improves the accuracy of income level classification and can capture the temporal changes of group attributes, significantly improving the accuracy of group attribute perception.

[0018] This invention uses mobile signaling data-driven group attribute distribution to replace traditional static population census data, and combines road network and public transportation capacity timetables to calculate service accessibility differences in real time, so that fairness assessment indicators such as the Gini coefficient are upgraded from annual static updates to dynamic updates at the 15-minute level, and fairness indicators are dynamically quantified.

[0019] This invention decomposes the global fairness index into the marginal contribution coefficient of each station through counterfactual reasoning, and uses this coefficient as a correction term for the multi-agent reinforcement learning reward function, so that the global fairness goal is accurately transmitted to the individual decision of each station, achieving Pareto improvement in both efficiency and fairness. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a dynamic optimization method for rail transit station connection resources provided in one embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of a user-region-time heterogeneous graph structure for a dynamic optimization method of rail transit station connection resources provided in one embodiment of the present invention.

[0022] Figure 3 This is a graph autoencoder pre-training and revenue classification fine-tuning network structure diagram for a dynamic optimization method of rail transit station connection resources provided in one embodiment of the present invention.

[0023] Figure 4 This is a multi-agent reinforcement learning framework diagram (centralized training, distributed execution) for a dynamic optimization method of rail transit station connection resources provided in one embodiment of the present invention.

[0024] Figure 5 This is a distribution map of the overall evaluation scores of Guangzhou rail transit stations, which is provided in one embodiment of the present invention as a dynamic optimization method for rail transit station connection resources.

[0025] Figure 6 This is a distribution map of the overall evaluation scores of rail transit stations in different concentric circles, which is provided in one embodiment of the present invention as a dynamic optimization method for rail transit station connection resources.

[0026] Figure 7 This is a heat map showing the correlation of comprehensive evaluation indicators for a dynamic optimization method for rail transit station connection resources provided in one embodiment of the present invention.

[0027] Figure 8 The image shows the average reward variation during the MAPPO training process of a dynamic optimization method for rail transit station connection resources provided in one embodiment of the present invention.

[0028] Figure 9 This is an overall flowchart of a dynamic optimization method for rail transit station connection resources provided in one embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention first fuses users' spatiotemporal behavioral features through heterogeneous graphs and extracts the spatial distribution of group attributes using graph neural networks; second, it quantifies the contribution of each station to the global service equilibrium through counterfactual reasoning; finally, it uses the contribution value of each station as a differentiated guiding signal and embeds it into a multi-agent reinforcement learning framework, so that each station, while pursuing its own operational efficiency, assumes optimization responsibility commensurate with its contribution to the global equilibrium.

[0031] like Figures 1 to 9 As shown, the present invention provides a dynamic optimization method for rail transit station connection resources, comprising the following steps: S100: Obtain user's mobile trajectory data and public transportation travel records; Specifically, this invention collects multi-source data in real time, including mobile phone signaling data, subway AFC card swiping data, point of interest data, and road network data, providing a data foundation for subsequent heterogeneous map construction and group attribute inference.

[0032] In one embodiment, the mobile signaling data includes mobile data from ground base stations and mobile location data from the subway's proprietary cellular network.

[0033] Mobile phone data from ground base stations includes user ID, base station ID, timestamp, event type, longitude, latitude and longitude, and other tag attributes, which are used to determine the user's handover and movement trajectory between ground base stations.

[0034] The mobile phone location data from the subway's dedicated cellular network includes tag attributes such as user ID, base station ID, corresponding subway station, timestamp, event type, longitude, and latitude, which are used to determine the routes and stations a user takes when traveling by subway. The field information of this data is shown in Table 1.

[0035] Table 1. Data Field Information Table for Dedicated Cellular Network in Subway After the two data sources are integrated, the user's movement trajectory in the urban space throughout the day can be fully depicted, including the user's geographical location at different times, the length of stay at each geographical location, and information on entering, exiting and passing through stations in the subway system.

[0036] This invention also collects link table information for subway lines, including subway lines, subway stations, up / down information, transfer station identifiers, and other subway attribute information. The field information of this data is shown in Table 2.

[0037] Table 2 Link Field Information Table for Metro Lines The subway link data provides a public transportation capacity topology basis for subsequent calculations of the number of stations reachable by each traffic zone within a preset travel time threshold.

[0038] The above data is sliced ​​according to a preset time window. In one embodiment, the 24 hours of a day are divided into 15-minute time windows, resulting in 96 time-series segments. The location information of each user within each time window is aggregated into a spatiotemporal record of "the user appearing within a certain geographic grid," providing basic data for the subsequent construction of heterogeneous maps.

[0039] During the data preprocessing stage, the following cleaning operations are performed on the original signaling data: (1) Remove abnormal location records caused by signal drift, that is, remove records whose latitude and longitude exceed the city administrative boundary or whose displacement speed exceeds a reasonable threshold (such as speed exceeding 120 km / h); (2) Deduplicate consecutive records of the same user within the same base station coverage area in adjacent timestamps and merge them into one dwell event; (3) For records with missing fields, if the missing field is a secondary field (such as event type), fill in the default value; if the missing field is a core field (such as user number, timestamp), directly remove the record.

[0040] S200. Based on the movement trajectory data, construct a heterogeneous map describing the relationship between individual users, geographical units, and time series segments; Specifically, the heterogeneous graph G=(V,E) constructed by this invention contains three types of nodes and two types of edges, such as... Figure 2 As shown.

[0041] The node types and initial characteristics include: User node U represents each observed mobile terminal user. The initial feature vector for each user node... It includes the following statistical characteristics: Trip frequency: The number of cross-base station moves that occur within a day, reflecting the user's daily travel activity. The statistical method is as follows: For each user, all their signaling records for the entire day are traversed, and each trip is counted whenever the base station number changes. The cumulative value for the entire day is used as the trip frequency.

[0042] Nighttime dwell time stability: This is determined by statistically analyzing the base station locations where users appear most frequently during the early morning hours (e.g., 0:00 to 5:00), and is characterized by the proportion of times the same user appears at the same base station during this period to the total number of nighttime records. The higher this ratio, the more stable the user's nighttime dwell time.

[0043] Travel time flexibility: This is characterized by the degree to which a user's travel time deviates from the city's average commute time. Specifically, the departure time of all the user's travel events is first counted, and the standard deviation of its time deviation relative to the city's morning peak (7:00-9:00) and evening peak (17:00-19:00) is calculated. The smaller the standard deviation, the more fixed the user's travel time (inflexible travel), and the larger the standard deviation, the more flexible the travel time (flexible travel).

[0044] Geographic unit node (regional node) Z represents a geographic grid after dividing the urban space according to a preset granularity. Grid division can be carried out in one of the following ways: (a) by administrative boundaries, such as street or community boundaries; (b) by regular grids, such as 500m×500m or 1km×1km square grids; (c) by Traffic Analysis Zone (TAZ), which takes into account factors such as road network structure, natural geographic boundaries and administrative divisions.

[0045] Initial feature vector of each region node Includes the following features: Point of Interest (POI) density: The ratio of the number of POIs of various types within the grid to the grid area, calculated separately for each POI category, including dining, shopping, business / office, residential, and public service POIs. POI density comprehensively reflects the level of urban service provision and functional mix in the area.

[0046] Urban function density: The degree of agglomeration of various urban functions within this grid can be quantified by the kernel density estimate of POI.

[0047] Median house price in a region: As an indirect indicator of the region's socioeconomic status, it can be obtained from real estate transaction platforms or publicly available government data.

[0048] Time segmentation node (time node) T: Represents a time segment after a day has been divided according to a preset interval. In one embodiment, the 24 hours of a day are divided into 96 time nodes at 15-minute intervals. The initial feature vector of each time node... The time period encoding vector can be generated in one of the following ways: (a) one-hot encoding, which encodes each of the 96 time periods into a 96-dimensional one-hot vector; (b) periodic encoding, which maps the time period number to sin and cos function values ​​to reflect the periodic continuity of time; (c) semantic time period encoding, which divides a day into semantic time periods such as morning peak, off-peak, evening peak, and nighttime, and encodes them into different category values.

[0049] The edge types include: User-Region Access Edge: If a user appears in region z during time period t, an access relationship edge is established between the user node and the region node. The existence of this edge represents the spatiotemporal affiliation between the user and the geographic space, and is the main path for subsequent message passing in the graph neural network. The weight of the edge can be set as the frequency or duration of the user's appearance in the region during that time period.

[0050] Region-to-Region Spatial Adjacency Edge: If two geographic grids are spatially adjacent (sharing a boundary or vertex), a spatial adjacency edge is established between the corresponding region nodes. This edge allows feature information to be transferred between adjacent geographic units, achieving spatial smoothing and mutual enhancement of regional features. The edge weight can be set to the length of the common boundary between the two grids or a normalized value of the shared boundary.

[0051] In one embodiment, the heterogeneous graph has approximately 2 million user nodes, 5,000 regional grid nodes, 96 time nodes (one every 15 minutes for a total of 24 hours), and approximately 80 million edges.

[0052] As needed, this invention constructs a heterogeneous graph containing spatiotemporal correlation information, providing data support for graph neural networks to improve the accuracy of group attribute inference by utilizing graph structure information, and solving the problem that traditional methods use only isolated statistical features, resulting in poor inference accuracy in sparse data regions.

[0053] This invention employs a graph autoencoder (GAE) to embed and extract the aforementioned heterogeneous graphs, such as... Figure 3 As shown in the diagram, the encoder consists of two parts: an encoder and a decoder.

[0054] Encoder structure: A heterogeneous graph attention network (HAN) or a relational graph convolutional network (RGCN) is used as the encoder to map each node to a low-dimensional embedding vector. ,in Represents a node The set of neighboring nodes. The core operation of the encoder is message passing and feature aggregation on the multi-type edges of the heterogeneous graph.

[0055] In one embodiment, a relational graph convolutional network is used as the encoder, which contains two graph convolutional layers, each with an output dimension of 128 dimensions.

[0056] For a given user node, its embedding vector is updated by aggregating the feature information of the regional nodes visited by the user within the same time segment, as well as the feature information of other user nodes with similar spatiotemporal behavioral patterns. Through this aggregation, the representation vectors of user nodes with similar spatiotemporal behavioral patterns in the latent space tend to converge, allowing for inference even if the data for these users in the original feature space is sparse, by leveraging information from similar groups.

[0057] For a given regional node, its embedding vector is updated by aggregating the feature information of all user nodes served by that region within the same time segment, as well as the feature information of other regional nodes spatially adjacent to that region. Through this aggregation, the regional node not only carries its own urban functional characteristics but also the behavioral characteristics of its internal user group, achieving a deep fusion of regional attributes and population attributes.

[0058] Decoder: A bilinear layer is used as the decoder to reconstruct the embedded vectors into node attributes. Probability of the existence of edges .

[0059] Pre-training loss function: An unsupervised loss function is used during the pre-training phase to optimize two objectives simultaneously: The first term is the node attribute reconstruction loss, which uses mean squared error (MSE) to measure the decoder's ability to recover the initial features of nodes from the embedded vectors. The second term is the edge prediction loss, which uses binary cross-entropy (BCE) to measure the accuracy of the model in predicting whether there is an edge connection between any two nodes in a heterogeneous graph. λ is the weighting coefficient between the two losses.

[0060] The Adam optimizer was used for pre-training with a learning rate of 0.001 and a batch size of 1024 node pairs, for a total of 100 training epochs. During training, early stopping was used to monitor the edge prediction accuracy on the validation set, and training was stopped when the validation set loss no longer decreased for 10 consecutive epochs.

[0061] Through joint optimization of these two reconstruction tasks, the embedding vectors fully preserve the node attribute and structural information of the original graph. This step, through pre-training of a graph autoencoder, learns a general behavioral representation on massive amounts of unlabeled spatiotemporal behavioral data, providing a high-quality initial embedding for subsequent supervised fine-tuning with a small number of labeled samples.

[0062] S300. Use a graph neural network to extract the behavioral representation vectors of user nodes in the heterogeneous graph, and map the behavioral representation vectors to the user's group attribute labels; After obtaining the embedding vector of the user node, this invention uses a small amount of labeled sample data to perform supervised fine-tuning of the embedding vector, mapping it to the user's group attribute label.

[0063] In one embodiment, the group attribute label is the residents' income level, using a three-category setting: low income, middle income, and high income. The labeled sample data comes from micro-data from the population census (such as sampling data from the latest population census of each city), with a sample size of approximately 5,000 records.

[0064] During the supervised fine-tuning stage, a classification layer is added to the encoder output. The input to the classification layer is the embedding vector of the user node (128-dimensional), and the output is the probability distribution of three preset categories. The classification layer structure is: fully connected layer (128-dimensional → 64-dimensional) + ReLU activation + fully connected layer (64-dimensional → 3-dimensional) + Softmax activation.

[0065] The loss function for fine-tuning training is: in, The focus loss weights are used to address class imbalance. Based on the standard cross-entropy loss, the focus loss automatically reduces the weights of easily classified samples with high classification confidence and automatically increases the weights of difficult-to-classify samples with low classification confidence. This makes the training process pay more attention to the classification boundaries of minority class samples, effectively improving the recognition accuracy of the minority class.

[0066] The necessity of the above design lies in the fact that the distribution of the number of samples in each category is naturally unbalanced (usually the middle-income group has the highest proportion, while the high-income and low-income groups have relatively lower proportions). If this is not addressed, the classification model will tend to predict all samples as the majority class to obtain overall accuracy. Focus loss effectively mitigates the adverse effects of class imbalance on the classification boundary by reducing the weight of easily classified samples.

[0067] After fine-tuning, the probability distribution of each user node obtained through the classification layer is... By aggregating data according to their geographical units, we can obtain the population density and its temporal changes for different groups within each region: in, Indicates site The preset geographic buffer.

[0068] To verify the reliability of the aggregation results, the aggregation results can be compared with the district-level census statistics data. When the difference in distribution between the two exceeds a preset threshold, the judgment threshold of the classification layer is adjusted to ensure the consistency and stability of the inversion results at the regional statistical level.

[0069] In one embodiment, fine-tuning was performed using 5000 user labels, and focus loss was employed to address class imbalance. The final revenue inference accuracy reached 82%, a 14 percentage point improvement over the traditional XGBoost method's 68%. Comparison results of different methods are shown in Table 3.

[0070] Table 3 Comparison of different methods This step employs a two-stage transfer learning strategy of "unsupervised pre-training + fine-tuning with a small number of labeled samples". First, it uses massive amounts of unlabeled spatiotemporal behavioral data to learn general behavioral representations, and then uses a small number of labeled samples for domain adaptation, achieving high-precision group attribute inference with extremely low labeling costs.

[0071] It should be noted that income level inference in this step is not simply a machine learning classification task. Its core value lies in coupling the inference results with subsequent spatial accessibility calculations. Specifically, it uses the inferred population spatial distribution density as a weighting factor in the Gini coefficient calculation, thus elevating fairness quantification from traditional "regional mean statistics" to "individual-group distribution weighted statistics." This coupling relationship is the essential feature that distinguishes this invention from general income inference methods. Even if income inference itself has certain classification errors, the subsequent calculations in this invention are more sensitive to the distribution proportions of each group within the same geographical unit (rather than the absolute accuracy of individual-level labels), making the overall fairness assessment robust to inference errors.

[0072] S400. Based on the group attribute tags and the public transportation travel records, determine the operational efficiency value of each station and the global service accessibility difference. This invention calculates the operational efficiency value of each station within a preset time window based on AFC public transport travel records.

[0073] In one embodiment, the dynamic competitiveness score of site i within time window t Defined as the percentage of motorized rail travel trips at this station during the target time period: The numerator represents the number of rail transit trips entering the station, obtained through statistics of card-swipe entry data recorded by the subway AFC system. Specifically, all card-swipe records for this station within time period t are read from the AFC system, and the number of entries is accumulated after deduplication by user.

[0074] The denominator is the total motorized travel volume in the area surrounding the station (e.g., within an 800-meter radius buffer zone around the station), which includes the sum of various motorized travel modes such as rail transit, ground public transport, and private cars. The specific calculation method is as follows: (1) Rail transit travel volume: the total number of passengers entering each station within the buffer zone is calculated through the subway AFC system; (2) Ground public transport travel volume: the total number of passengers boarding each bus station within the buffer zone is calculated through bus IC card swiping data or bus GPS trajectory data; (3) Private car travel volume: estimated through the driving mode recognition results in traffic flow monitoring data or mobile phone signaling data. The higher this ratio, the stronger the backbone role of the station in the regional transportation system and the better its operational efficiency.

[0075] The operational efficiency value is calculated at the station level and dynamically updated in 15-minute time windows, which can reflect the efficiency changes of each station at different times due to passenger flow fluctuations.

[0076] In one embodiment, the present invention calculates the service accessibility differences between different groups based on road network topology and public transportation capacity timetables, thereby quantifying the fairness of resource allocation.

[0077] Step 1: Determine the service coverage of each geographic unit.

[0078] The city is divided into M traffic analysis zones (TAZs). For each zone m, based on the real-time road network topology and public transportation (including rail transit and surface buses) capacity and timetables, the number of stations or jobs reachable from that zone within a preset travel time threshold is calculated. This value is defined as the rail service accessibility of that zone. in, Let I{·} be the actual travel time from community m to station j within time period t, and let I{·} be the indicator function. A preset travel time threshold is set (in one embodiment, the value is 30 minutes).

[0079] The above calculations need to comprehensively consider factors such as actual road traffic conditions, public transportation transfer times, and station entry and exit times. In one embodiment, Dijkstra's shortest path algorithm in graph theory is used, with intersections in the road network as nodes, road segments as edges, and road segment travel time as weights, combined with the stop time information of public transportation lines, to calculate the shortest actual travel time from the centroid of the community to each station.

[0080] Specifically, constructing a multimodal transportation network ,in, These are road network nodes (including intersections and stations). This is a set of road segments. The weight of each segment is the actual travel time through that segment, calculated by dividing the segment length by the average speed of that segment during that time period. The average speed of the segment can be obtained in real-time from floating car GPS data or traffic monitoring data. For subway travel segments, the weight is the train travel time plus the arrival and departure times, obtained from the subway timetable. All stations with a travel time less than or equal to T_max are included in the reachable count.

[0081] Step 2: Establish the correspondence between resident group ranking and accessibility.

[0082] Residents within each traffic-adjacent area are sorted according to group attribute labels (such as income level from low to high) to form a cumulative resident percentage sequence. For each cumulative percentage position, calculate the cumulative percentage of service accessibility actually enjoyed by the corresponding resident group. Specifically, the residents are ranked and then added up sequentially. For every certain percentage (e.g., 1%) of residents added up, the cumulative percentage of the service accessibility value of the community where that group of residents resides is recorded relative to the total service accessibility value of all communities.

[0083] Step 3: Calculate the service reachability difference.

[0084] According to the above ( , A concentration distribution curve is plotted based on the sequence. The horizontal axis of the curve represents the cumulative percentage of residents (from low income to high income), and the vertical axis represents the cumulative percentage of service accessibility. If resource allocation is perfectly balanced, residents in each income quantile enjoy a proportional share of service accessibility, and the concentration distribution curve coincides perfectly with the 45° diagonal (equilibrium reference line). If the proportion of service accessibility enjoyed by high-income groups is higher than their population proportion, the curve lies below the diagonal; if low-income groups enjoy more services, the curve lies above the diagonal.

[0085] Spatiotemporal Gini coefficient The calculation formula is: The service accessibility disparity, or Gini coefficient, ranges from 0 to 1. A value of 0 indicates perfect equilibrium, while a higher value indicates greater disparities in service accessibility between different groups and poorer fairness in resource allocation.

[0086] Define the site fairness index The higher the index, the better the overall service equilibrium.

[0087] In one embodiment, Guangzhou City was divided into 3,989 traffic analysis zones. The number of accessible rail stations within a 30-minute isochronous circle (calculated using the road network and subway timetable) was used to calculate the initial Gini coefficient for the morning peak hours as 0.42, indicating poor fairness.

[0088] This step enables dynamic perception of resident group attributes by introducing mobile signaling data, and calculates accessibility by combining it with real-time road network status. This allows the service accessibility difference to be updated in real time with changes in passenger flow and road conditions, providing real-time perception capabilities for subsequent dynamic decision-making.

[0089] S500. Simulate removing each station from the rail transit network one by one, recalculate the service accessibility difference after removal, and use the increment of the service accessibility difference before and after removal as the marginal impact coefficient of each station on the global service equilibrium. It should be noted that when simulating station removal, for stations with transfer relationships, the redistribution of transfer passenger flow due to the removal of the station needs to be addressed simultaneously. Specifically, transfer passenger flow that passed through the removed station in the original planned route is allocated to nearby alternative stations according to the shortest path (recalculated using Dijkstra's algorithm in the remaining station network). The transfer passenger flow received by the alternative station is included in its service coverage recalculation. For isolated routes where transfers can only be completed through the removed station, the rail transit service for the corresponding passenger flow is considered unreachable and included in the unserved travel demand, thus being incorporated into the recalculation of global service accessibility differences. This redistribution mechanism ensures the logical consistency of passenger flow distribution during counterfactual reasoning and avoids underestimation or overestimation of service coverage due to simply removing stations.

[0090] This invention uses counterfactual reasoning to quantify the contribution of each site to the overall service balance.

[0091] Based on global service reachability As a baseline value. For any target site. The simulation involves removing the station from the rail transit network, specifically including: removing the station from the station list, recalculating the service coverage of all traffic cells after the removal of the station (travel routes that were originally accessible to the station are now considered inaccessible), and then recalculating the global service reachability difference using the same method as described above. .

[0092] The change in the degree of difference before and after the removal of the target site is defined as the marginal contribution of that site: Where F(t) is the fairness index when all sites are included. To remove the site The fairness index was then recalculated. Since the fairness index F = 1 - G, this marginal contribution is equivalent to... .

[0093] This value is an incremental value, representing the extent to which the global unfairness worsens after removing the site. The larger the marginal contribution, the more irreplaceable the site's role in maintaining global service equilibrium, and the higher its contribution to overall fairness.

[0094] The value is an increment, measuring the extent to which global unfairness worsens after removing the site. Marginal contribution. The larger the value, the more irreplaceable the site's role in maintaining global service balance, and the higher its contribution to overall fairness. If... A negative value indicates that removing the site actually improves fairness (i.e., the site may have exacerbated unfairness), and the optimization guidelines for such sites should be different.

[0095] It should be noted that the aforementioned "simulated cancellation" involves logically deleting and recalculating the set of sites within the computational model, rather than physically canceling the operation of the site. This counterfactual reasoning process can be completed quickly within a computer system without actually altering the operational state of the physical network.

[0096] This step uses counterfactual reasoning to break down the abstract global fairness index into individual contribution values ​​for each site, providing a quantitative basis for subsequent differentiated decision-making.

[0097] S600. Configure each site as a decision agent, and use the marginal impact coefficient corresponding to each site as the weight factor of the fairness sub-item in the reinforcement learning reward function of the decision agent. Perform strategy iteration with the goal of co-optimizing the running efficiency value and the service reachability difference, and generate connection resource adjustment actions for each site. This invention configures each subway station as an independent decision-making agent, employing a multi-agent reinforcement learning framework that combines centralized training with distributed execution, such as... Figure 4 As shown.

[0098] The state space design is as follows: The observation state of each agent i at time t Includes the following dimensions: Site competitiveness: The operational efficiency value of site i in the current time period. .

[0099] Partial status of surrounding stations: the competitiveness value of stations adjacent to this station (such as adjacent by line or geographical proximity).

[0100] Density of different income groups in the surrounding area: Population density of each income group (low-income, middle-income, high-income) within the buffer zone of this site. .

[0101] Current connection resource allocation: including connecting bus services around this station. Shared bicycle parking spaces Parking prices .

[0102] In one embodiment, each agent observes its own state and the states of three surrounding stations, resulting in a 64-dimensional state space. During execution, each agent can only observe its own local state information and cannot know the global state.

[0103] The motion space design is as follows: The action output of each agent i includes adjustments in three dimensions: in: The adjustment amount of connecting bus services is discretized into five levels: {-2, -1, 0, +1, +2}. The adjustment amount for the number of shared bicycle parking spaces is discretized into five levels: {-5, -3, 0, +3, +5}. The adjustment ratio for dynamic pricing of parking lots is discretized into five levels: {-10%, -5%, 0, +5%, +10%}.

[0104] The combination of the three dimensions results in 125 possible discrete action options. The actual adjustment range of each dimension can be preset with different base step sizes based on city size and station type.

[0105] The above-mentioned discrete tiers are based on the following engineering constraints: (1) The bus frequency adjustment tier of ±2 bus trips / hour corresponds to the minimum adjustment granularity (15 minutes) of the departure interval of regular routes in the urban bus dispatching specifications. Exceeding this adjustment range will break the coordination range of synchronous dispatching of other stations on the route; (2) The shared bicycle parking space adjustment tier of ±5 parking spaces corresponds to the minimum space management unit of the electronic fence (usually arranged in a grid of 5 parking spaces). Adjustments below this granularity cannot be effectively marked on the physical ground; (3) The parking fee rate adjustment ratio of ±10% corresponds to the minimum threshold of the price lever having a significant impact on the demand elasticity. Price adjustments below this level are unlikely to have an observable regulatory effect on parking demand.

[0106] The reward function is designed as follows: The global reward function is designed as a weighted combination of three sub-objectives: The first item is the arithmetic mean of the operational efficiency values ​​of all stations, reflecting the overall operational efficiency. The second term is the global fairness index F(t) = 1 - G(t), which reflects the degree of global service balance. The third item is the total cost incurred by all site resource adjustment actions, specifically: in, These are the unit costs for adjusting bus schedules, shared bicycle parking spaces, and parking prices, respectively.

[0107] , , These are the weighting coefficients for the three sub-objectives. In one embodiment, =0.5、 =0.4、 =0.1.

[0108] In one embodiment, the marginal contribution coefficient of each station is used as a correction term in the corresponding agent's reward function. Specifically, for stations with higher marginal contribution coefficients (i.e., those contributing more to the global equilibrium), each action that improves fairness receives an additional positive reward bonus, making them more inclined to assume responsibility for maintaining the global equilibrium when making decisions. Conversely, for stations with lower marginal contribution coefficients, their fairness-oriented actions receive weaker additional incentives, allowing them to focus more on optimizing their own efficiency while ensuring basic fairness constraints.

[0109] Through the aforementioned differentiated guidance, the goal of global fairness is precisely transmitted to the individual decision-making at each site.

[0110] Resource Constraint: The actions of each agent are limited by the upper limit of the total global resources. This constraint is implemented using the Lagrange relaxation method. in, The original objective function for policy optimization is... Adjust the total amount of resources for all sites. To preset the global resource limit, For Lagrange multipliers. When the total resource adjustment exceeds the preset upper limit, the second term generates a positive penalty value, reducing the global optimization objective and forcing the agents to coordinate the control of resource increments in subsequent decisions.

[0111] The centralized training and distributed execution architecture is as follows: This invention employs a multi-agent reinforcement learning framework with centralized training and distributed execution.

[0112] Training phase: Policy network for each agent With a centralized value network Simultaneous training, among which The joint state of all agents. This represents the joint action of all agents. The advantage function is estimated as follows: in, As a discount factor, This is the output of a centralized value network.

[0113] The strategy update adopts a tailored PPO objective: in, This represents the probability ratio between the current strategy and the old strategy. These are the trimming parameters.

[0114] Through centralized training, the agents develop a tacit understanding and learn to coordinate their respective action outputs when the global state changes.

[0115] Execution Phase: After training, only the policy networks of each agent are deployed to edge computing nodes. Each agent independently generates actions based on its own observed local state information. No communication or cooperation is required between agents, and the decision latency is controlled within milliseconds.

[0116] In one embodiment, the number of agents is 30 key stations (other stations use a global average strategy), the MAPPO algorithm is used, trained for 3 million steps, with a discount factor of 0.99 and a GAE parameter of 0.95. Figure 8 As shown, the initial average reward during training was approximately 28.5, and the average reward after convergence was approximately 37.6, representing an improvement of approximately 31.9%.

[0117] This step greatly reduces the complexity of the decision-making problem by setting each site as an independent agent, with each agent only handling its own low-dimensional state and action space. At the same time, it achieves collaboration between agents through centralized training and ensures real-time performance through distributed execution, realizing real-time collaborative optimization of a large-scale site cluster.

[0118] S700, The adjustment action is converted into an adjustment command and sent to the field execution terminal; specifically, Before issuing connection resource adjustment instructions to the field execution terminal, this invention performs smoothing filtering on the instruction sequence generated within a continuous time window: Specifically, the action value output by each agent is processed using a first-order low-pass filter: in, The smoothed instruction actually issued at the current moment. The original action output by the agent at the current moment. The instructions actually issued at the previous moment. This is a smoothing coefficient, with a value between 0.2 and 0.4.

[0119] The necessity of the aforementioned smoothing process lies in the fact that the distribution of group attribute labels may generate statistical noise due to instantaneous fluctuations in mobile signaling data. If the unsmoothed raw instructions are directly issued to the execution terminal, it may cause frequent jumps in bus departure frequencies between two adjacent time windows. This not only increases the scheduling costs of bus companies and the workload of drivers and conductors, but also affects the travel experience of passengers. Through smoothing filtering, the changes in the instruction sequence are controlled within a reasonable range, avoiding oscillations at the execution terminal caused by data noise.

[0120] This invention sets safe operating constraints for adjusting the connection resources of each site, ensuring that the adjusted resource configuration parameters always remain within the safe operating range: Bus service frequency shall not be less than the minimum guaranteed frequency: ;in, To ensure that the minimum guaranteed number of bus trips (e.g., 4 trips / hour) is set, the bus connection service at any station is not lower than the basic service level. The number of shared bicycle parking spaces must be no less than zero. Ensure that there is no negative parking space allocation; Parking fees are within the preset range: ;in, and These are the minimum and maximum fees permitted by law or policy, respectively.

[0121] When the action output by the agent exceeds the above safety constraint boundary after smoothing, a pruning operation is performed: the action value that exceeds the boundary is truncated to the nearest legal boundary value.

[0122] This invention sets up an instruction execution status monitoring loop (watchdog mechanism) at each site. Specifically, after receiving an adjustment instruction and completing the execution action, the field execution terminal must return an execution success confirmation signal to the edge computing node. If an adjustment instruction at a certain site fails to receive a valid response confirmation from the execution terminal multiple times consecutively (e.g., three times consecutively, corresponding to 45 minutes), the decision-making agent at that site is automatically switched to a backup operation mode.

[0123] In the minimum guarantee operation mode, the configuration of the site's connection resources is no longer decided by the multi-agent policy network, but falls back to the scheduling strategy based on a fixed timetable (such as running at the preset minimum guarantee frequency) until the monitoring loop confirms that the execution terminal has restored normal communication and action capabilities, and then switches back to the agent decision mode.

[0124] The aforementioned fallback mechanism ensures that the system can maintain basic services under any abnormal circumstances (such as communication interruption, execution terminal failure, data loss, etc.), preventing the entire site connection service from being paralyzed due to a single point of failure in the intelligent decision-making module.

[0125] This invention also provides a dynamic optimization system for rail transit station connection resources, comprising: The data access unit receives mobile signaling data streams and public transportation transaction data streams. This unit receives in real-time mobile signaling data (including terrestrial base station signaling and subway-dedicated cellular network signaling) pushed from the data gateway of the telecommunications operator, card-swiping transaction data pushed from the subway AFC system, and bus operation data pushed from the bus dispatch system. The data access unit performs format unification, time alignment, and quality verification on the received multi-source data, and pushes the verified data to the map management unit.

[0126] The graph management unit is used to construct and maintain heterogeneous graphs between individual users, geospatial units, and time-series segments based on the data from the data access unit. This unit dynamically updates node features and edge relationships in the heterogeneous graph using 15-minute time slices. Specifically, this unit maintains a graph database, supports CRUD operations on nodes and edges, and triggers a graph update process periodically (e.g., every 15 minutes).

[0127] The attribute-aware unit, with a built-in pre-trained and fine-tuned graph neural network model, is used to output user group attribute classifications based on the heterogeneous graph. This unit first loads the pre-trained graph autoencoder model parameters, then loads the fine-tuned classification layer parameters at the encoder output, performs forward inference on the input user node embedding vectors, and outputs the probability distribution of each user belonging to each group attribute label.

[0128] The equilibrium measurement unit is used to calculate the global service equilibrium degree based on the distribution of the group attribute classification on the geographic unit and the service coverage of each site, and to recalculate the equilibrium degree after simulating the removal of each site to determine the marginal contribution value of each site.

[0129] Furthermore, the equilibrium measurement unit includes: The coverage calculation subunit is used to calculate the isochronous coverage circle of each geographical unit based on real-time traffic conditions and rail operation plans. This subunit loads real-time traffic data (segment speed) and subway timetables, performs multi-mode shortest path calculations, and outputs the set of stations reachable by each traffic zone within a preset travel time threshold.

[0130] The deviation quantification sub-unit is used to construct a concentrated distribution curve of service coverage based on group attributes and to quantify the deviation between this curve and the equilibrium curve. This sub-unit sorts residents according to group attribute labels, constructs a Lorenz curve, and calculates the Gini coefficient.

[0131] The attribution analysis subunit generates a balanced contribution ranking for each site by systematically removing sites and recording changes in deviation values. This subunit performs counterfactual reasoning, logically deleting each site and re-invoking the coverage measurement subunit and deviation quantification subunit for calculation.

[0132] The collaborative decision-making unit incorporates a multi-agent policy network with each site acting as an independent decision-maker. It aims to achieve collaborative optimization of global operational efficiency and global service balance. Based on the marginal contribution value of each site, it differentiates the optimization direction of each agent and outputs a connection resource adjustment scheme. This unit deploys the trained agent policy networks. Each agent independently performs forward inference based on local observation information, outputting action values, which are then smoothed and filtered to form the adjustment scheme.

[0133] The execution conversion unit is used to convert the adjustment scheme into drive signals compatible with the bus dispatch terminal, shared bicycle control terminal, and parking fee collection terminal. This unit has a built-in protocol conversion module that converts standardized adjustment commands into communication protocol formats accepted by each execution terminal (such as RESTful API calls for the bus dispatch system, gRPC interface calls for the shared bicycle platform, and Modbus commands for the parking fee collection system), and distributes them to each field execution terminal via a secure network channel.

[0134] This invention compresses trained graph neural network models using knowledge distillation techniques. Specifically, a pre-trained large-scale graph neural network (such as a 2-layer, 128-dimensional relational graph convolutional network) is used as the teacher network, and a simpler, lightweight network (such as a 2-layer, 32-dimensional graph convolutional network) is used as the student network. The soft labels output by the teacher network guide the training of the student network. The compressed model has approximately 20% fewer parameters than the original model, and the inference speed is improved by more than 5 times.

[0135] The compressed graph neural network model and the multi-agent policy network are jointly deployed on edge computing nodes in the station domain. These edge computing nodes are located in roadside units or station servers (such as NVIDIA Jetson AGX Orin), with each edge node responsible for processing real-time data streams in the area surrounding its station. Inference latency is controlled within 50 milliseconds to meet the response time requirements for real-time decision-making.

[0136] Regarding data security, when aggregating user movement trajectory data, edge computing nodes only upload grid-level statistical results to the central cloud platform, without transmitting any individual-level trajectory data. In one embodiment, differential privacy protection is applied to the aggregated statistics: in, This represents the population density statistics for group label c within geographic grid g. For random noise that follows a Laplace distribution, For the sensitivity of the query function, This is a privacy budget parameter. This processing ensures that even if an attacker obtains the aggregated statistics uploaded by edge nodes, they cannot deduce any individual user's trajectory information or group attribute tags from it.

[0137] Figures 5 to 8 The implementation effect of the present invention in Guangzhou is shown.

[0138] In one embodiment, Guangzhou City delineates 3989 traffic analysis zones and selects 30 key rail transit stations as decision-making agents (the remaining stations adopt a global average strategy). The system operation flow is as follows: The data access unit continuously receives signaling data from ground base stations of telecommunications operators, data from the subway's dedicated cellular network, and data streams from subway AFC card swiping. After format unification and time alignment, the data is transmitted to the map management unit. The map management unit uses 15-minute time slices to build and update a heterogeneous map in real time, containing approximately 2 million user nodes, 5,000 geographic unit nodes, and 96 time-series segment nodes.

[0139] The graph neural network model in the attribute-aware unit processes approximately 80 million edge relationships (including user-region access edges and region-region adjacency edges) during the pre-training phase, which lasts for 100 epochs with a learning rate of 0.001. In the fine-tuning phase, 5,000 census micro-data labels are used, and the classification layer is trained using focus loss. Ultimately, the attribute-aware unit achieves an overall classification accuracy of 82% for user income levels, a 14 percentage point improvement compared to the 68% accuracy of the traditional XGBoost method.

[0140] The coverage calculation subunit in the equilibrium measurement unit calculates the 30-minute isochronous circle range of each traffic zone based on real-time traffic information and subway timetable; the deviation quantification subunit constructs a centralized distribution curve according to income ranking and calculates the global spatiotemporal Gini coefficient, with the initial Gini coefficient for the morning peak being 0.42; the attribution analysis subunit generates an equilibrium contribution ranking table for each station by eliminating counterfactual inferences from 30 key stations one by one, and this ranking table is fed back to the collaborative decision-making unit in real time.

[0141] The multi-agent policy network in the collaborative decision-making unit uses 30 key stations as independent decision-makers. Each agent has a 64-dimensional state space and an action space containing 3 dimensions with 5 discrete options each. It was trained for 3 million steps using a multi-agent proximal policy optimization algorithm with a discount factor of 0.99 and a GAE parameter of 0.95. Figure 8As shown, the average reward during training converged from approximately 28.5 initially to approximately 37.6, representing an improvement of approximately 31.9%.

[0142] The execution conversion unit converts the resource adjustment plan output by the collaborative decision-making unit into standardized control instructions: the bus capacity adjustment amount is pushed to the bus intelligent dispatching system through the RESTful API, and the dispatching system automatically generates the departure timetable adjustment plan; the shared bicycle capacity adjustment amount is sent to each shared bicycle operation platform, and the platform adjusts the electronic fence capacity configuration and dispatching plan accordingly; the parking fee rate adjustment ratio is sent to the parking fee management system, and the management system automatically updates the gate fee standard.

[0143] Table 4 Comparison of Key Indicators Before and After Optimization As shown in Table 4, under the static fixed configuration before optimization, the average competitiveness of the 30 sites was 0.48, the global fairness index (1-Gini coefficient) was 0.58, and the Gini coefficient was 0.42. After dynamic optimization and adjustment according to this invention, the average competitiveness increased to 0.53 (improvement +10.4%), the fairness index increased to 0.72 (improvement +24.1%), and the Gini coefficient decreased to 0.28 (improvement -33.3%). This achieved synergistic optimization of operational efficiency and service fairness, verifying the effectiveness and engineering applicability of the proposed solution.

[0144] To verify the independent contribution of each core module of this invention, the following ablation experiments were set up: (1) Ablation scheme A: the counterfactual marginal influence coefficient was removed, and all agents adopted a unified fairness weight factor; (2) Ablation scheme B: the graph neural network group attribute inference module was removed, and static population distribution data from the census was used instead; (3) Ablation scheme C: single agent centralized decision-making was used to replace the multi-agent distributed architecture. The experiment was run for 3 million steps under the same experimental conditions, and the comparison results are shown in Table 5. Table 5 Comparison of Key Indicators in Ablation Experiments The results show that counterfactual differentiation weights contribute the most to improving fairness (by about 0.11), graph neural network dynamic perception significantly improves both efficiency and fairness compared to static census data, and multi-agent distributed architecture reduces decision latency by an order of magnitude while ensuring optimization effect, meeting real-time control requirements.

[0145] Figure 5 This shows a map illustrating the distribution of the overall scores for the comprehensive evaluation of Guangzhou's rail transit stations. Figure 6 The distribution of the overall evaluation scores for rail transit stations in different concentric circles is shown. Figure 7A heatmap showing the correlation of comprehensive evaluation indicators is presented. The above figures illustrate the comprehensive evaluation results of the sites after the implementation of this invention in Guangzhou from different dimensions.

[0146] This invention effectively solves the technical problems in the prior art where service efficiency and fairness are difficult to optimize in a coordinated manner due to the lag in revenue data, static fairness assessment, and lack of dynamic closed-loop resource allocation.

[0147] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method described in any of the preceding claims.

[0148] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the preceding claims.

[0149] The foregoing description of specific exemplary embodiments of the present invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it is obvious that many changes and variations can be made based on the above teachings. Although embodiments of the invention have been shown and described, these specific embodiments are merely explanations of the invention and are not intended to limit it. The specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. The purpose of selecting and describing exemplary embodiments is to explain the specific principles of the invention and its practical application, so that those skilled in the art, after reading this specification, can make modifications, substitutions, variations, and various choices and changes to the embodiments as needed without departing from the principles and spirit of the invention, provided that such modifications, substitutions, variations, and choices and changes are within the scope of the claims of the invention and are protected by patent law.

Claims

1. A method for dynamic optimization of rail transit station connection resources, characterized in that, include: Obtain users' movement trajectory data and public transportation travel records; Based on the movement trajectory data, a heterogeneous map is constructed, which is used to describe the spatiotemporal relationships between individual users, geographical units, and time series segments. A graph neural network is used to extract the behavioral representation vectors of user nodes in the heterogeneous graph, and the behavioral representation vectors are mapped to the group attribute labels of users. Based on the group attribute tags and the public transportation travel records, determine the operational efficiency value of each station and the global service accessibility difference. The simulation removes each station from the rail transit network one by one, and recalculates the service accessibility difference after removal. The increment of the service accessibility difference before and after removal is used as the marginal impact coefficient of each station on the global service equilibrium. Each site is configured as a decision agent, and the marginal impact coefficient corresponding to each site is used as the weight factor of the fairness sub-item in the reinforcement learning reward function of the decision agent. The strategy is iterated with the goal of co-optimizing the running efficiency value and the service reachability difference, and the connection resource adjustment actions of each site are generated. The adjustment action is converted into an adjustment command and sent to the on-site execution terminal.

2. The method for dynamic optimization of rail transit station connection resources according to claim 1, characterized in that, In the heterogeneous spectrum: User nodes and geographic unit nodes are connected by access relationship edges formed based on the movement trajectory data; geographic unit nodes are connected by adjacency edges formed based on spatial topological adjacency relationships. The initial characteristics of the user node include at least travel frequency and nighttime dwell stability, and the initial characteristics of the geographic unit node include at least urban functional density.

3. The method for dynamic optimization of rail transit station connection resources according to claim 2, characterized in that, The graph neural network extracts features in the following way: Aggregate the features of user nodes sharing the same geographic unit within the same spatiotemporal slice, and output the spatial distribution density of the group; Aggregate the features of the same user node across adjacent time-series segments and output the temporal evolution features of individual behavior.

4. The method for dynamic optimization of rail transit station connection resources according to claim 1, characterized in that, The behavior representation vector is mapped to group attribute labels and fine-tuned using multi-class supervision; the loss function of the fine-tuning includes a class weight factor, which is negatively correlated with the proportion of each class in the labeled samples.

5. The method for dynamic optimization of rail transit station connection resources according to claim 1, characterized in that, The global service reachability disparity is determined in the following way: Based on the road network topology and public transportation capacity timetable, determine the number of stations that can be reached within the preset travel time threshold for each geographical unit, and use this number as the service coverage of that geographical unit. Residents within a geographic unit are sorted according to the group attribute labels, and a corresponding curve between the cumulative percentage of residents and the cumulative percentage of service coverage is constructed. Calculate the area of ​​the corresponding curve that deviates from the equilibrium reference line, and use the area to characterize the global service reachability difference.

6. The method for dynamic optimization of rail transit station connection resources according to claim 1, characterized in that, The marginal impact coefficients corresponding to each site are used as weighting factors for the fairness sub-item in the reinforcement learning reward function of the decision-making agent, including: The reward function includes an efficiency sub-item, a fairness sub-item, and a cost sub-item; the weight coefficient of the fairness sub-item is positively correlated with the marginal influence coefficient of the corresponding station of the decision agent, so that the station with the larger marginal influence coefficient bears a higher responsibility for fairness optimization in the strategy iteration.

7. The method for dynamic optimization of rail transit station connection resources according to claim 1, characterized in that, The decision-making agent adopts a centralized training and distributed execution architecture, including: During the training phase, the actions and global state changes of each decision-making agent are collected centrally to update the global evaluation network; During the execution phase, each decision-making agent independently outputs actions based on local observation information; the local observation information includes the efficiency value of the local station, the efficiency values ​​of neighboring stations, the density of surrounding populations, and the current connection resource configuration parameters.

8. The method for dynamic optimization of rail transit station connection resources according to claim 7, characterized in that, The adjustments to the connecting resources include adjustments to public transport capacity, shared bicycle capacity, and parking fee rates. When each decision-making agent outputs the action, it is constrained by the total global resources. The constraint is achieved by superimposing a penalty term on the reward function. The penalty term is positively correlated with the degree to which the total resource adjustment of each site exceeds the preset upper limit.

9. The method for dynamic optimization of rail transit station connection resources according to claim 1, characterized in that, Before converting the adjustment action into an adjustment command and issuing it, the adjustment action within the continuous time window is subjected to a first-order low-pass filter. An execution status monitoring loop is set up so that if any station fails to respond effectively to the adjustment command within a preset number of consecutive cycles, the station is switched to a preset backup operation mode.

10. A dynamic optimization system for public transportation station connection resources, characterized in that, include: The data access unit is used to receive mobile signaling data streams and public transportation transaction data streams; The map management unit is used to construct and maintain heterogeneous maps between individual users, geospatial units, and time series segments based on the data from the data access unit. The attribute-aware unit has a built-in pre-trained and fine-tuned graph neural network model, which is used to output the group attribute classification of users based on the heterogeneous graph. The equilibrium measurement unit is used to calculate the global service equilibrium degree based on the distribution of the group attribute classification on the geographic unit and the service coverage of each site, and to recalculate the global service equilibrium degree after simulating the removal of each site to determine the marginal contribution value of each site. The collaborative decision-making unit has a built-in multi-agent policy network with each site as an independent decision-making entity. It uses the marginal contribution value of each site as the weight factor of the balance sub-item in the reward function of the agent, and performs policy iteration with global operating efficiency and global service balance as the collaborative optimization objectives to output a connection resource adjustment scheme. The conversion unit is used to convert the adjustment scheme into a drive signal that is compatible with the bus dispatch terminal, the shared bicycle control terminal, and the parking fee terminal.

Images