A parking space supply and demand prediction scheduling method based on a space-time graph neural network

Abstract

The application discloses a berth supply and demand prediction scheduling method based on a space-time graph neural network, and relates to the technical field of intelligent parking management and space-time data prediction scheduling. The method reconstructs a trusted berth state by acquiring multi-source parking data in a road parking scene and forming time slice data of berth nodes, extracts berth supply and demand time sequence mode features, and characterizes future transferable parking demand as transferable demand shares. A scheduling response dynamic graph is constructed to generate candidate scheduling actions, and a benchmark comparison graph and an intervention comparison graph are constructed. Through shared coding, differential coding and action-aware space-time graph neural network models, benchmark prediction trajectories and intervention prediction trajectories are obtained synchronously. A berth scheduling strategy is determined based on counterfactual scheduling benefits, and the scheduling response dynamic graph is calibrated through low-risk exploration scheduling feedback. The application can solve the problem of inaccurate scheduling impact evaluation in road berth supply and demand prediction, and improve the reliability of parking guidance and berth scheduling.

Description

Technical Field

[0001] This invention relates to the field of intelligent parking management and spatiotemporal data prediction and scheduling technology, and in particular to a parking space supply and demand prediction and scheduling method based on spatiotemporal graph neural networks. Background Technology

[0002] With the development of urban on-street parking management platforms, data such as parking space occupancy status, order records, turnover rate, road environment, and holiday activities are increasingly being used for parking supply and demand forecasting and parking guidance. Existing technologies typically predict parking space availability or regional supply and demand pressure within a certain period based on historical occupancy rates, entry and exit statistics, spatial distance, or road connectivity, and then provide parking guidance, inspection arrangements, or operational scheduling suggestions to managers or drivers accordingly.

[0003] However, the supply and demand situation in on-street parking scenarios is highly dynamic and regionally interconnected. When parking pressure increases in a certain area, the management platform may alter the spatial distribution of parking demand through methods such as guidance, inspections, and the opening of shared parking spaces. Existing solutions often separate the prediction and scheduling processes; that is, they first predict the supply and demand situation and then perform empirical scheduling based on the prediction results. Because prediction models typically do not adequately characterize the impact of scheduling behavior on subsequent supply and demand changes, it is difficult to determine before scheduling whether a particular scheduling suggestion can truly alleviate high-pressure areas, and it is also difficult to determine whether the receiving area will experience new congestion due to increased demand.

[0004] Furthermore, on-street parking data comes from complex sources, including geomagnetic sensors, high-position video feeds, manual inspections, and order systems, which may suffer from latency, incompleteness, or conflicts. Supply and demand patterns, road accessibility, and historical capacity utilization also vary across different areas. Relying solely on static spatial relationships or historical averages for prediction and scheduling can easily lead to misinterpreting natural fluctuations as scheduling effectiveness or ignoring inconsistencies in scheduling response capabilities across different areas. This can result in inaccurate parking guidance, overloaded service areas, unreasonable allocation of inspection resources, and difficulty in continuously optimizing scheduling strategies based on actual operational feedback. Therefore, a technical solution is urgently needed that can more accurately assess the impact of scheduling during on-street parking supply and demand forecasting and utilize feedback to correct scheduling decisions. Summary of the Invention

[0005] To address the problem of inaccurate assessment of scheduling impact in on-street berth supply and demand forecasting, this invention proposes a berth supply and demand forecasting and scheduling method based on a spatiotemporal graph neural network.

[0006] The present invention achieves the above objectives through the following technical solutions:

[0007] A berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural networks, the method comprising:

[0008] Acquire multi-source parking data in on-street parking scenarios and form time-slice data of parking space nodes according to preset time granularity;

[0009] Based on the time slice data, the reliable berth status is reconstructed, the berth supply and demand time sequence pattern features are extracted, and the future transferable parking demand is characterized as the transferable demand share.

[0010] Based on the trusted berth status, berth supply and demand time sequence pattern characteristics, and historical scheduling feedback data in the time slice data, a scheduling response dynamic graph is constructed. Candidate scheduling actions are generated according to the source node, receiving node, scheduling response edge weight and transferable demand share in the scheduling response dynamic graph. Paired benchmark comparison graphs and intervention comparison graphs are constructed for the candidate scheduling actions.

[0011] Shared encoding and differential encoding are performed on the baseline control map and the intervention control map to obtain common state representation and action effect representation. Together with the graph action variables corresponding to the candidate scheduling actions, they are used as prediction conditions input into the action perception spatiotemporal graph neural network model, so that the graph action variables participate in the graph propagation, time update and prediction output process. Under the conservation transfer constraint of the transferable demand share, the baseline prediction trajectory and the intervention prediction trajectory are obtained synchronously.

[0012] Based on the counterfactual scheduling benefits of the baseline predicted trajectory and the intervention predicted trajectory, a berth scheduling strategy is determined, and the scheduling response dynamic diagram is calibrated through low-risk trial scheduling feedback.

[0013] A further improvement of the present invention is that the time slice data includes at least occupancy observation data, order statistics data, operational statistics data, external context data, and historical scheduling feedback data;

[0014] Based on the occupancy observation data from at least two types of occupancy observation sources at the same berth node in the same time slice, and the dynamic reliability of the data source determined by the online status, observation delay, inter-source consistency and historical error feedback of the occupancy observation sources, the state confidence of the credible berth state is reconstructed and the state confidence of the credible berth state is generated.

[0015] Based on the order statistics, operational statistics, external context data and historical scheduling feedback data, berth supply and demand time series pattern features, including supply and demand phase relationship, peak and valley migration, turnover fluctuation, context response elasticity and historical scheduling response, are extracted.

[0016] Based on the source berth node, estimated arrival time slice, acceptable transfer cost, supply and demand time sequence pattern label, acceptance constraint satisfaction, state confidence, and historical scheduling response, a transferability score for the future predicted parking demand is calculated. The future predicted parking demand is then adaptively divided into multiple transferable demand shares according to the transferability score. The sum of multiple transferable demand shares belonging to the same source berth node and the same estimated arrival time slice does not exceed the future predicted parking demand of that source berth node in that estimated arrival time slice.

[0017] A further improvement of the present invention is that the construction of the scheduling response dynamic graph includes:

[0018] Basic berth edges are generated based on the geographical distance and road connectivity between berth nodes, and time-series pattern similar edges are generated based on the similarity of berth supply and demand time-series pattern features between berth nodes.

[0019] Within the acceptance screening range defined by the basic berth edge and the time sequence pattern similarity edge, berth nodes whose supply and demand pressure exceeds the preset pressure threshold are identified as source nodes, and berth nodes with supply margin and whose state confidence meets the preset confidence conditions are identified as acceptance nodes.

[0020] Based on the historical scheduling response rate, the magnitude of pressure reduction after historical scheduling, and the magnitude of congestion after historical scheduling within the same historical scheduling feedback window, an asymmetric scheduling response edge is generated from the source node to the receiving node. Based on the historical scheduling response rate, the magnitude of pressure reduction after historical scheduling, the magnitude of congestion after historical scheduling, the state confidence of the receiving node, the acceptable transfer cost of the transferable demand share, and the transferability score, the scheduling response edge weight of the asymmetric scheduling response edge is determined.

[0021] Configure response differential labels for the asymmetric scheduling response edge, including the expected pressure decrease of the source node, the expected pressure increase of the receiving node, the upper limit of the demand share transfer, and the demand share transfer direction;

[0022] The basic berth edge, the time-series pattern similar edge, and the asymmetric scheduling response edge are merged into the scheduling response dynamic graph.

[0023] A further improvement of the present invention is that candidate scheduling actions are generated, and pairs of baseline comparison maps and intervention comparison maps are constructed for the candidate scheduling actions, including:

[0024] In the asymmetric scheduling response edge, based on the scheduling response edge weight, response differential label and transferable demand share, candidate transfer node pairs and candidate transfer ratios of transferable demand shares between candidate transfer node pairs are determined, as well as the transferable demand share transfer amount determined by the candidate transfer ratios and transferable demand shares.

[0025] If the transferable demand share transfer amount does not exceed the upper limit of the demand share transfer amount, the candidate transfer node pair, scheduling response edge weight, candidate transfer ratio, transferable demand share and response differential label are combined into a candidate scheduling action and encoded as the graph action variable;

[0026] The scheduling response dynamic diagram without applying candidate scheduling actions is determined as the baseline comparison diagram, and the scheduling response dynamic diagram formed by transferring the share of transferable demand along the asymmetric scheduling response edge after applying candidate scheduling actions is determined as the intervention comparison diagram.

[0027] When constructing the baseline comparison map and the intervention comparison map, the graph structure and node characteristics that are unrelated to the candidate scheduling actions are kept consistent, and the distribution of transferable demand share, scheduling response edge weight, initial value of supply and demand pressure of source node and initial value of supply and demand pressure of receiving node are changed according to the response differential label.

[0028] A contrast mask is generated based on the response differential label to limit the scope of differential coding.

[0029] A further improvement of the present invention is that the baseline control map and the intervention control map are subjected to shared encoding and differential encoding, including:

[0030] The baseline control map and the intervention control map are encoded using a parameter-sharing graph encoder to obtain the baseline map code and the intervention map code, respectively.

[0031] The common-state representation is obtained by sharing pooling the coding regions in the baseline and intervention maps that are not marked by the control difference mask; the action effect representation is obtained by differential coding the coding regions in the baseline and intervention maps that are marked by the control difference mask.

[0032] The graph action variables are used to perform gating correction on the action effect representation so that the action effect representation retains only the transferable demand share transfer effect, scheduling response edge weight change effect, and initial supply and demand pressure change effect related to the candidate scheduling action.

[0033] The common-state representation and the gated action effect representation are used as prediction conditions for the action-aware spatiotemporal graph neural network model.

[0034] A further improvement of the present invention is that the graph action variables participate in the graph propagation, time update, and prediction output processes, including:

[0035] Gating analysis is performed on the graph action variables to obtain the edge propagation gate, node update gate, and trajectory output gate;

[0036] During graph propagation, the baseline comparison graph and the intervention comparison graph are processed using shared propagation parameters respectively, and the edge propagation gate is used to control the action message transmission along the asymmetric scheduling response edge in the intervention comparison graph, so that the transferable demand share participates in the graph propagation from the source node to the receiving node according to the demand share transfer direction;

[0037] During the time update process, the node update gate is used to write the expected pressure decrease of the source node and the expected pressure increase of the receiving node as a pair of state changes into the future prediction window, so that the candidate scheduling action participates in the time state recursion of the source node and the receiving node.

[0038] During the prediction output process, the common-state representation and the gated-corrected action effect representation are controlled by the trajectory output gate to enter the baseline output branch and the intervention output branch, respectively. The baseline output branch suppresses the action effect representation, and the intervention output branch introduces the action effect representation within the range defined by the control difference mask, thereby obtaining the initial baseline prediction trajectory and the initial intervention prediction trajectory.

[0039] A further improvement of the present invention lies in simultaneously obtaining the baseline prediction trajectory and the intervention prediction trajectory under the conservation transfer constraint of the transferable demand share, including:

[0040] Within the same future prediction window, based on the response differential label and the comparison difference mask, a local conservation relationship is established for the reduction in source node demand, the increase in receiving node demand, and the transferable demand share caused by the candidate scheduling action.

[0041] When the initial intervention prediction trajectory does not satisfy the local conservation relationship, the source node demand prediction value and the receiving node demand prediction value are corrected in pairs only within the range of the source node, receiving node and corresponding asymmetric scheduling response edge defined by the contrast difference mask, through the conservation redistribution mechanism. The pair correction is constrained by the upper limit of the demand share transfer amount and by the expected pressure decrease of the source node and the expected pressure increase of the receiving node, so that the demand decrease of the source node and the demand increase of the receiving node correspond to each other within the future prediction window.

[0042] The local conservation relationship is constrained by the following formula:

[0043] ;

[0044] In the formula, Indicates berth node In the future The baseline transferable demand share for each time slice, Indicates the execution of candidate scheduling actions. Rear berth node In the future Intervention in a single time slice can transfer a share of demand; Indicates the execution of candidate scheduling actions Later in the future Within a time slice, the berth nodes Transfer to berth node The transferable share of demand; Indicates the execution of candidate scheduling actions Later in the future Within a time slice, the berth nodes Transfer to berth node The transferable share of demand; Indicates the current time slice The scheduling response dynamic diagram is composed of berth nodes. Pointing to berth node Scheduling response edge weight;

[0045] The intervention prediction trajectory is generated based on the paired modified share of transferable demand for the intervention, and the baseline prediction trajectory is generated based on the baseline share of transferable demand when the candidate scheduling action is not applied.

[0046] A further improvement of the present invention is that the berth scheduling strategy is determined based on the counterfactual scheduling benefits of the baseline predicted trajectory and the intervention predicted trajectory, including:

[0047] Based on the baseline prediction trajectory and the intervention prediction trajectory, the baseline conservative supply and demand pressure and the intervention conservative supply and demand pressure are calculated respectively.

[0048] Within the scope of the differential mask, the scheduling benefits within the mask for candidate scheduling actions are determined based on the conservative decrease in supply and demand pressure of the source node, the conservative increase in supply and demand pressure of the receiving node, the transfer cost of the transferable demand share, and the state confidence of the receiving node.

[0049] Within the berth node range where the difference mask is not marked, the mask leakage penalty for candidate scheduling actions is determined based on the supply and demand pressure deviation between the baseline predicted trajectory and the intervention predicted trajectory.

[0050] The counterfactual scheduling benefit is determined based on the scheduling benefit within the mask, the leakage penalty outside the mask, the outbound penalty for receiving nodes exceeding the receiving pressure threshold, and the conservation correction penalty generated by pairwise correction.

[0051] The candidate scheduling action that satisfies the preset benefit conditions and whose intervention conservative supply and demand pressure at the receiving node does not exceed the receiving pressure threshold is selected to generate the parking space scheduling strategy. The parking space scheduling strategy includes at least one of the following: high supply and demand pressure area early warning, receiving area recommendation, parking guidance recommendation, inspection priority suggestion, and shared parking space opening suggestion. The parking space scheduling strategy is output to aggregated parking demand.

[0052] A further improvement of the present invention is that the action-aware spatiotemporal graph neural network model is obtained through joint training of counterfactual scheduling gain and constraint loss, wherein the joint training includes:

[0053] Based on the deviation between the baseline predicted trajectory and the actual supply and demand trajectory when the candidate scheduling action was not applied, and the deviation between the intervention predicted trajectory and the actual supply and demand change trajectory corresponding to historical scheduling feedback data, the prediction loss is obtained. Based on the deviation of the local conservation relationship, the conservation constraint loss is obtained. Based on the supply and demand pressure deviation between the baseline predicted trajectory and the intervention predicted trajectory within the berth node range not marked by the control difference mask, the leakage loss outside the mask is obtained. ;

[0054] Based on the amount of supply and demand pressure exceeding the threshold of the receiving node's intervention, determine the boundary violation penalty item for the receiving node, and determine the transfer cost penalty item based on the candidate transfer ratio and transfer cost corresponding to the candidate scheduling action.

[0055] The predicted loss Conservation constraint loss Losses due to mask leakage The model training objective is formed by the combination of the node out-of-bounds penalty, the transfer cost penalty, and the counterfactual scheduling benefit. , represented as:

[0056] ;

[0057] In the formula, Indicates candidate scheduling actions The set of receiving nodes involved, Indicates candidate scheduling actions The set of source nodes involved; This represents the total number of time slices within the future forecast window; Indicates the execution of candidate scheduling actions. Subsequent Node In the future Intervention in a specific timeframe to mitigate supply and demand pressures Indicates the receiving node The corresponding bearing pressure threshold; Represents the positive part function; Indicates candidate scheduling actions Source node To the receiving node The proportion of candidate transfers, Indicates that it is from the source node To the receiving node The transfer cost; Indicates candidate scheduling actions Counterfactual scheduling benefits; , , , and This represents the weighting parameter.

[0058] A further improvement of the present invention is that the scheduling response dynamic graph is calibrated through low-risk trial scheduling feedback, including:

[0059] When the low-risk probing conditions are met, a low-proportion probing scheduling action is generated along the asymmetric scheduling response edge to be calibrated. The low-risk probing conditions include: the source node is not a node with extremely high supply and demand pressure, the lower bound of the supply prediction of the receiving node is greater than the preset safety margin, the probing transfer ratio is lower than the preset probing ratio threshold, and the expected intervention conservative supply and demand pressure of the receiving node after the low-proportion probing scheduling action is executed does not exceed the receiving pressure threshold.

[0060] After executing the low-proportion trial scheduling action, the actual pressure change of the source node, the actual berth status change of the receiving node, and the actual amount of transferable demand share are obtained in the feedback window. The actual amount of transferable demand share is obtained based on the aggregated entry increment, the change in idle berths, and the historical natural fluctuation baseline differential estimation of the receiving node in the feedback window.

[0061] The trial response residual is obtained based on the deviation between the actual pressure change, the actual berth status change, the actual load, and the intervention prediction trajectory.

[0062] Based on the trial response residual, calibrate the historical scheduling response rate, historical pressure drop after scheduling, historical congestion absorption after scheduling, scheduling response edge weight, and response differential label corresponding to the asymmetric scheduling response edge.

[0063] The calibrated scheduling response edge weights and response difference labels are used for the construction of the scheduling response dynamic graph, generation of candidate scheduling actions, construction of benchmark comparison graphs and intervention comparison graphs, and determination of berth scheduling strategies in the next prediction cycle.

[0064] The beneficial effects of this invention are: it combines on-street parking supply and demand forecasting with scheduling impact assessment, allowing for comparison of future supply and demand changes under different scheduling scenarios before generating scheduling strategies, thereby reducing the uncertainty caused by relying solely on historical trends or empirical rules for scheduling. By reconstructing the reliable state of multi-source parking data and combining the patterns of parking space supply and demand changes with historical scheduling feedback to determine the scheduling basis, the accuracy of identifying high-pressure areas and available areas is improved, reducing the risk of erroneous scheduling caused by data conflicts, equipment errors, or static distance judgments. It can distinguish between the actual impact of scheduling behavior and the natural fluctuations of parking demand itself in the judgment of scheduling benefits, avoiding the incorrect inclusion of supply and demand changes in irrelevant areas in the scheduling effect; simultaneously, by constraining transferable parking demand and continuously calibrating the scheduling basis using low-risk trial scheduling feedback, the probability of overload in available areas can be reduced, improving the reliability of parking guidance recommendations, high supply and demand pressure area early warnings, available area recommendations, inspection priority suggestions, and shared parking space opening suggestions, enabling on-street parking scheduling strategies to be continuously optimized according to actual operating conditions. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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 based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0067] This invention addresses the difficulty in accurately determining "where parking shortages are imminent, which areas can absorb the load, and whether scheduling is truly effective" in on-street parking management. It proposes a parking space supply and demand prediction and scheduling method based on a spatiotemporal graph neural network. The core idea is to first construct a dynamic scheduling response graph based on the historical scheduling response relationships between each parking space node. Then, for each candidate scheduling action, a baseline comparison graph (without scheduling) and an intervention comparison graph (with scheduling) are constructed. Shared and differential coding are used to identify the supply and demand changes caused by the scheduling action itself, and the differences between the two predicted trajectories are compared under the constraint of conserving transferable demand share. This allows for the determination of whether a scheduling action can truly reduce the supply and demand pressure at the source node, while avoiding transferring pressure to already near-saturated receiving nodes. Furthermore, low-risk trial scheduling feedback is used to continuously calibrate the dynamic scheduling response graph, enabling subsequent predictions and scheduling strategies to be dynamically adjusted based on actual operational conditions. This achieves early prediction, robust diversion, and closed-loop optimization of areas with tight on-street parking supply and demand.

[0068] like Figure 1 The diagram illustrates an embodiment of the present invention. This embodiment's parking space supply and demand forecasting and scheduling method is primarily applied to on-street parking scenarios. The system can be deployed on the backend of a city-wide smart parking management platform, an edge-cloud collaborative parking management platform, or a parking operation management platform. The input is aggregated data from on-street parking operations, and the output includes future parking space supply and demand forecasts, parking space scheduling strategies, and a calibrated dynamic scheduling response diagram.

[0069] The vehicle entry / exit data, order statistics, and user behavior-related data involved in this invention are all calculated in an aggregated statistical form and are not intended to identify specific vehicles, users, or individuals. Any personal information or vehicle identification information must be used with the user's consent and in compliance with laws and regulations, and must be anonymized, desensitized, or aggregated before model training or inference. The algorithm rules of this invention do not contain discriminatory judgment conditions targeting specific groups.

[0070] This method mainly includes the following steps:

[0071] S1: Acquire multi-source parking data in on-street parking scenarios and form time slice data of parking space nodes according to preset time granularity.

[0072] Parking space nodes can be individual on-street parking spaces, parking stations, street parking areas, or administrative district parking areas. The specific granularity can be determined based on the management precision of the parking management platform. Preset time granularity can be 5 minutes, 10 minutes, 15 minutes, or 30 minutes. For example, if the time granularity is 15 minutes, the occupancy status, number of occupants, number of occupants, parking space utilization rate, turnover rate, and external context of each parking space node within 15 minutes will be aggregated into a single time slice of data.

[0073] Time-slice data includes at least occupancy observation data, order statistics, operational statistics, external context data, and historical dispatch feedback data. Occupancy observation data can come from at least two types of occupancy observation sources, such as geomagnetic equipment, high-position video, PDA registration, and manual inspection. Order statistics can include aggregated statistics such as historical entry numbers, historical exit numbers, temporary parking orders, and monthly subscription orders. Operational statistics can include historical berth utilization rate, historical berth turnover rate, number of berths in use, and number of berths available. External context data can include weather, holidays, large-scale events, road construction, and temporary traffic control. Historical dispatch feedback data can include historical dispatch response rate, the extent of pressure reduction after historical dispatch, and the extent of congestion handling after historical dispatch.

[0074] S2: Reconstruct the reliable berth status based on time slice data, extract the time series pattern features of berth supply and demand, and characterize future transferable parking demand as transferable demand share.

[0075] Specifically, for occupancy observation data from at least two types of occupancy observation sources at the same berth node within the same time slice, the reliable berth state is reconstructed based on the dynamic reliability of each occupancy observation source. This reconstructed state represents the reliable occupancy status of the berth node within the corresponding time slice. The dynamic reliability of the data source is determined by the online status of the occupancy observation source, observation latency, inter-source consistency, and historical error feedback. A data source with high dynamic reliability is characterized by being online, having short observation latency, being consistent with other occupancy observation sources, and having low historical errors.

[0076] In one implementation, each occupancy observation is converted into an occupancy observation value between 0 and 1, where 0 represents idle, 1 represents occupied, and values ​​between 0 and 1 represent probabilistic occupancy. The dynamic reliability of the data source can be expressed as: ; Indicates berth node In time slice Next Dynamic reliability of data sources for each occupancy observation source; Indicates online status; 1 indicates online status, and 0 indicates offline status. Indicates the normalized observation delay; This indicates the consistency between the occupied observation source and other occupied observation sources; Indicates historical error feedback; to These are weight parameters that can be adjusted based on the validation set. This is the Sigmoid function, used to map the result to a range between 0 and 1.

[0077] The reliable berth status can be calculated as follows: ; Indicates berth node In time slice The credible berth status, Indicates the first Occupied observation values ​​of each occupancy observation source. Indicates the number of observation sources occupied. To prevent constants with a denominator of zero, we can take... . and All are dimensionless occupancy probabilities.

[0078] Furthermore, a state confidence level is generated based on the deviations of multiple occupancy observation sources relative to the credible berth state. (Values ​​range from 0 to 1): The more consistent the multiple occupied observation sources are, the higher the state confidence; the greater the difference in occupied observation sources, the lower the state confidence.

[0079] The berth supply and demand time-series characteristics include supply and demand phase relationship, peak-valley migration, turnover fluctuation, contextual response resilience, and historical scheduling response. These characteristics are extracted from order statistics, operational statistics, external contextual data, and historical scheduling feedback data, and are used to represent the supply and demand variation patterns of berth nodes. Specifically, the supply and demand phase relationship represents the time difference between peak arrival and peak departure; peak-valley migration represents the time offset of peaks or troughs under different dates, weather, or event conditions; turnover fluctuation represents the magnitude of change in berth turnover rate over different time periods; contextual response resilience represents the impact of weather, holidays, or large-scale events on berth demand; and historical scheduling response represents the changes in pressure and capacity of berth nodes after historical scheduling.

[0080] Based on the source parking space nodes, estimated arrival time slices, acceptable transfer costs, supply and demand time series pattern labels, constraint satisfaction, state confidence, and historical scheduling responses, a transferability score for the future predicted parking demand is calculated. The transferability score can be a normalized value between 0 and 1; a higher value indicates that the portion of future parking demand is more suitable for transfer to other parking space nodes. In one implementation, the transferability score can be expressed as: ; Indicates the source berth node In the future Transferability score for each time slice; Indicates the source berth node In the future The degree of constraint satisfaction corresponding to each time slice; Indicates historical scheduling response, This represents the normalized value of acceptable transfer costs. This represents the transition probability corresponding to the supply and demand time series pattern label. to These are the weight parameters.

[0081] Constraint Satisfaction Used to indicate the source berth node In the future The predicted parking demand for each time slice is the degree to which it can be accommodated within the candidate capacity range, with a value ranging from 0 to 1. It can be determined based on the lower bound of the predicted supply of parking spaces within the candidate capacity range, the preset safety margin, the baseline conservative supply and demand pressure, the capacity pressure threshold, and the state confidence level. When there are no parking spaces within the candidate capacity range that meet the conditions of supply margin, safety margin, or capacity pressure threshold, ... The corresponding future predicted parking demand is not classified as a transferable demand share, or the corresponding transferable demand share is set to 0; when there are parking space nodes that meet the acceptance constraints, Based on a satisfaction level value between 0 and 1, future predicted parking demand is adaptively divided into multiple transferable demand shares according to the transferability score. Each transferable demand share is a unit for calculating aggregated parking demand and does not correspond to a single vehicle or user. The sum of multiple transferable demand shares belonging to the same source parking space node and the same expected arrival time slice does not exceed the future predicted parking demand of that source parking space node in that expected arrival time slice, thus avoiding the generation of transfer objects exceeding the predicted demand.

[0082] S3: Construct a scheduling response dynamic graph based on the trusted berth status, berth supply and demand time sequence pattern characteristics, and historical scheduling feedback data in time slice data. Generate candidate scheduling actions based on the source node, receiving node, scheduling response edge weight, and transferable demand share in the scheduling response dynamic graph, and construct paired benchmark comparison graphs and intervention comparison graphs for the candidate scheduling actions.

[0083] Specifically, basic berth edges are generated based on the geographical distance and road connectivity between berth nodes to represent the spatial reachability between them. For example, a basic berth edge is generated if the walking distance between two parking stations is less than a preset distance threshold, or if they can be reached within a preset time via the road network. Temporal pattern similarity edges are generated based on the similarity of berth supply and demand time-series characteristics between berth nodes. Similarity can be calculated using cosine similarity, dynamic time-warped distance, or correlation coefficient. If two berth nodes have similar entry peaks, exit peaks, turnover fluctuations, or contextual response elasticity, a temporal pattern similarity edge is generated between them.

[0084] Within the selection range defined by the basic berth edge and the temporal pattern similarity edge, berth nodes whose supply and demand pressure exceeds a preset pressure threshold are identified as source nodes, and berth nodes with supply margin and whose status confidence meets preset confidence conditions are identified as receiving nodes. Supply and demand pressure can be determined by the ratio of predicted demand to predicted supply; supply margin can be determined by the number of predicted available berths or the lower bound of supply prediction; preset confidence conditions are used to avoid identifying berth nodes with unreliable status as receiving nodes.

[0085] For both the source node and the receiving node, an asymmetric scheduling response edge is generated from the source node to the receiving node based on the historical scheduling response rate, the magnitude of pressure reduction after historical scheduling, and the magnitude of congestion after historical scheduling within the same historical scheduling feedback window. The historical scheduling feedback window refers to the time range used to observe changes in the source node and the receiving node after a historical scheduling action, such as 30 minutes, 45 minutes, or 60 minutes. Since the transfer capability from the source node to the receiving node is not necessarily the same as the reverse transfer capability, the scheduling response edge is set as an asymmetric edge.

[0086] The scheduling response edge weight is used to represent the feasibility of transferring a share of transferable demand from the source node to the receiving node. In one implementation, the scheduling response edge weight can be expressed as: ; Indicates the current time slice The next berth node Pointing to berth node Scheduling response edge weight, Indicates the historical scheduling response rate. This indicates the extent to which pressure has decreased after historical adjustments. This indicates the level of congestion following historical scheduling. Indicates the receiving node State confidence, The normalized value representing the acceptable transfer cost or actual transfer cost of the share of transferable demand. Indicates the transferability score. to The weights are parameters. All the above inputs can be normalized to between 0 and 1. Therefore, the scheduling response edge weights are dimensionless values, ranging from 0 to 1.

[0087] Configure response differential labels for asymmetric scheduling response edges, including the expected pressure decrease of the source node, the expected pressure increase of the receiving node, the upper limit of demand share transfer, and the direction of demand share transfer. Response differential labels are used to subsequently generate candidate scheduling actions, construct baseline and intervention comparison maps, and limit the scope of differential coding and conservation redistribution.

[0088] The basic berth edges, temporally similar edges, and asymmetric scheduling response edges are merged into a scheduling response dynamic graph, which is used to express the spatial reachability, supply and demand temporal similarity, and historical scheduling response relationships between berth nodes. The scheduling response dynamic graph is updated with each time slice, and the node state, edge weights, and response difference labels can be recalculated in each time slice.

[0089] The baseline comparison graph represents the graph state before any candidate scheduling action is applied, while the intervention comparison graph represents the graph state after any candidate scheduling action is applied. One implementation process is as follows:

[0090] In an asymmetric scheduling response edge, based on the scheduling response edge weight, response differential label, and transferable demand share, candidate transfer node pairs and the candidate transfer ratio of the transferable demand share between candidate transfer node pairs are determined, along with the transferable demand share transfer amount determined by the candidate transfer ratio and the transferable demand share. A candidate transfer node pair consists of a source node and a receiving node; the candidate transfer ratio represents the proportion of the transferable demand share planned to be transferred along the asymmetric scheduling response edge in the current candidate scheduling action; the transferable demand share transfer amount is obtained by multiplying the candidate transfer ratio by the corresponding transferable demand share, and is used to represent the aggregated demand amount actually participating in the candidate scheduling action along the asymmetric scheduling response edge.

[0091] When the transferable demand share corresponding to the candidate transfer ratio does not exceed the upper limit of the demand share transfer amount in the response differential label, the candidate transfer node pair, scheduling response edge weight, candidate transfer ratio, transferable demand share, and response differential label are combined into a candidate scheduling action and encoded as a graph action variable. The graph action variable is used to enable the action-aware spatiotemporal graph neural network model to identify which source node transferred how much transferable demand share to which receiving node, as well as the expected pressure change corresponding to the transfer.

[0092] For the same candidate scheduling action, a baseline control diagram and an intervention control diagram are constructed. The scheduling response dynamic diagram without applying a candidate scheduling action is determined as the baseline control diagram; the scheduling response dynamic diagram formed by the transfer of transferable demand share along the asymmetric scheduling response edge after applying a candidate scheduling action is determined as the intervention control diagram.

[0093] When constructing the two comparison graphs, the graph structure and node characteristics unrelated to the candidate scheduling actions are kept consistent. The distribution of transferable demand share, scheduling response edge weights, initial values ​​of supply and demand pressure at source nodes, and initial values ​​of supply and demand pressure at receiving nodes are modified according to the response differential labels. This processing ensures that the difference between the baseline and intervention comparison graphs is solely caused by the candidate scheduling actions, facilitating subsequent counterfactual comparisons. A comparison difference mask is generated based on the response differential labels to mark the source nodes, receiving nodes, asymmetric scheduling response edges, and related node characteristics affected by the candidate scheduling actions. Subsequent differential encoding is performed only within the range defined by this mask, thus avoiding misinterpreting natural fluctuations unrelated to the candidate scheduling actions as the effects of the scheduling actions.

[0094] S4: Shared encoding and differential encoding are performed on the baseline control map and the intervention control map to obtain common-state representation and action effect representation. Together with the graph action variables corresponding to the candidate scheduling actions, they are used as prediction conditions input into the action perception spatiotemporal graph neural network model, so that the graph action variables participate in the graph propagation, time update and prediction output process. Under the conservation transfer constraint of transferable demand share, the baseline prediction trajectory and the intervention prediction trajectory are obtained synchronously.

[0095] Specifically, a parameter-sharing graph encoder is used to encode the baseline control map and the intervention control map separately, resulting in baseline map codes and intervention map codes. The graph encoder can employ graph convolutional networks, graph attention networks, spatiotemporal graph convolutional networks, or combinations thereof. Parameter sharing means that the two maps use the same set of encoding parameters, thus ensuring that the baseline map codes and intervention map codes can be directly compared.

[0096] For coded regions not marked by the control difference mask, the baseline map encoding and the intervention map encoding are shared-pooled to obtain a common-state representation, used to represent the common operating state of the berth system in the current time slice. Shared-pooling can employ average pooling, max pooling, attention pooling, or gated pooling. For coded regions marked by the control difference mask, the baseline map encoding and the intervention map encoding are differentially encoded to obtain a representation of the action effect. Differential encoding can employ code subtraction, code concatenation and mapping, attention differencing, or gated differencing. For example: ; This represents the action effect. Indicates the intervention diagram coding. Indicates the base map encoding. Indicates the difference mask. This indicates element-wise multiplication. If a node or edge is not marked with a contrast mask, the difference result at that location is suppressed.

[0097] Furthermore, the action effect representation is gated and corrected using graph action variables, ensuring that the action effect representation retains only the transferable demand share transfer effect, scheduling response edge weight change effect, and initial supply and demand pressure change effect related to candidate scheduling actions. The corrected action effect representation, together with the common-state representation, serves as the prediction condition for the action-aware spatiotemporal graph neural network model.

[0098] The action-aware spatiotemporal graph neural network model includes a graph propagation layer, a time update layer, and a prediction output layer.

[0099] First, consider the graph action variables. Perform gating analysis to obtain the side propagation gate. Node update gate and trajectory output gate Gating parsing can be implemented in the following ways: ; , , , , , These are trainable parameters. All three gating values ​​can take values ​​between 0 and 1, and are used to control the strength of the candidate scheduling action in the graph propagation, time update, and output prediction processes.

[0100] During graph propagation, shared propagation parameters are used to process the baseline control graph and the intervention control graph separately. For the intervention control graph, edge propagation gates control the transmission of action messages along asymmetric scheduling response edges, enabling transferable demand shares to participate in graph propagation from the source node to the receiving node according to the demand share transfer direction. For the baseline control graph, no action messages corresponding to candidate scheduling actions are applied to maintain the natural evolution under the unscheduled condition.

[0101] During the time update process, the node update gate writes the expected pressure decrease of the source node and the expected pressure increase of the receiving node as paired state changes into the future prediction window. For the source node, the time update layer reduces its future demand state or supply-demand pressure state; for the receiving node, the time update layer increases its future demand state or supply-demand pressure state. In this way, candidate scheduling actions participate in the time state recursion, rather than just serving as ordinary input features.

[0102] During the prediction output process, the trajectory output gated common-state representation and the gated corrected motion effect representation are respectively fed into the baseline output branch and the intervention output branch. The baseline output branch suppresses the motion effect representation and outputs the initial baseline predicted trajectory; the intervention output branch introduces the motion effect representation within the range defined by the control difference mask and outputs the initial intervention predicted trajectory.

[0103] In the actual model, the graph propagation layer can use 2 to 4 layers of graph convolutional or graph attention layers, the time update layer can use a temporal convolutional network, a gated recurrent unit, or a Transformer temporal encoder, and the prediction output layer can use a multilayer perceptron. The hidden layer dimension can be set to 64, 128, or 256, and the future prediction window can be set to 4 to 24 time slices.

[0104] In this embodiment, both the baseline prediction trajectory and the intervention prediction trajectory are prediction sequences arranged by time slices within the future prediction window. Each prediction trajectory includes at least the predicted parking demand, predicted available berth supply, predicted berth utilization rate, predicted supply and demand pressure, and distribution of transferable demand share for each berth node in each future time slice. The baseline prediction trajectory represents the natural evolution of the berth system without the application of candidate scheduling actions, while the intervention prediction trajectory represents the evolution of the berth system after the application of candidate scheduling actions.

[0105] Within the same future forecast window, based on response differential labels and a comparison difference mask, a local conservation relationship is established for the reduction in demand at source nodes, the increase in demand at receiving nodes, and the transfer of transferable demand share caused by candidate scheduling actions. The meaning of this local conservation relationship is: the reduced transferable demand share at source nodes should correspond to the increased transferable demand share at receiving nodes; berth nodes not marked by the comparison difference mask do not participate in the conservation redistribution caused by the candidate scheduling action.

[0106] When the initial intervention prediction trajectory does not satisfy the local conservation relationship, the source node demand prediction value and the receiving node demand prediction value are pairedly corrected only within the range of the source node, receiving node, and corresponding asymmetric scheduling response edge defined by the contrast mask, through a conservation redistribution mechanism. The paired correction is constrained by the upper limit of the demand share transfer amount, and the correction direction is constrained by the expected pressure decrease of the source node and the expected pressure increase of the receiving node.

[0107] Local conservation relations are constrained by the following formula:

[0108] ;

[0109] In the formula, Indicates berth node In the future The baseline transferable demand share for each time slice, Indicates the execution of candidate scheduling actions. Rear berth node In the future Intervention in a single time slice can transfer a share of demand; Indicates the execution of candidate scheduling actions Later in the future Within a time slice, the berth nodes Transfer to berth node The transferable share of demand; Indicates the execution of candidate scheduling actions Later in the future Within a time slice, the berth nodes Transfer to berth node The transferable share of demand; Indicates the current time slice The scheduling response dynamic diagram is composed of berth nodes. Pointing to berth node The scheduling response edge weight. Among them, , , and Both refer to the quantity of transferable demand share or the normalized demand share weight. The boundary weights are dimensionless and range from 0 to 1.

[0110] After pairwise corrections are completed, an intervention prediction trajectory is generated based on the pairwise corrected share of transferable demand, and a baseline prediction trajectory is generated based on the baseline share of transferable demand when no candidate scheduling action is applied. Since the correction process is performed only within the limits of the control difference mask, it does not erroneously alter the berth node prediction results unrelated to candidate scheduling actions.

[0111] S5: Determine berth scheduling strategies based on counterfactual scheduling benefits of baseline and intervention prediction trajectories, and calibrate the scheduling response dynamic diagram through low-risk trial scheduling feedback, so that the scheduling response edge weights and response differential labels in the next prediction cycle are closer to the actual operation.

[0112] Specifically, based on the baseline forecast trajectory and the intervention forecast trajectory, baseline conservative supply and demand pressure and intervention conservative supply and demand pressure are calculated respectively. The conservative supply and demand pressure is used to represent the degree of supply and demand tension at berth nodes within future time slices. Upper bound of demand forecast. and lower bound of supply forecast It can be obtained through historical prediction residual statistics: ; ; Indicates berth node In the future Predicted parking demand for each time slot This indicates the predicted supply of available berths. This represents the preset quantile value of the demand forecast residual. This represents the preset quantile value of the supply forecast residual. and It can be obtained from the statistical analysis of historical prediction residuals in the validation set, for example, by using the 90th, 95th, or 99th percentile values.

[0113] Conservative supply and demand pressures can be expressed as: ; Indicates berth node In the future The conservative supply and demand pressure of a given time frame To prevent constants with a denominator of zero, both demand and supply are expressed as the quantity of berth demand or the normalized weighted berth demand, thus the ratio is a dimensionless pressure value.

[0114] Within the scope of the differential mask, the in-mask scheduling benefit of candidate scheduling actions is determined based on the conservative decrease in supply and demand pressure at the source node, the conservative increase in supply and demand pressure at the receiving node, the transfer cost of the transferable demand share, and the state confidence of the receiving node. Within the berth nodes not marked by the differential mask, the out-of-mask leakage penalty of candidate scheduling actions is determined based on the supply and demand pressure deviation between the baseline predicted trajectory and the intervention predicted trajectory.

[0115] The counterfactual scheduling benefit can be expressed as: ; Indicates candidate scheduling actions Counterfactual scheduling benefits; The scheduling benefit within the mask limit range of the comparison difference mask can be determined based on the conservative decrease in supply and demand pressure of the source node and the conservative increase in supply and demand pressure of the receiving node. The penalty for leakage outside the mask is determined based on the supply and demand pressure deviation of the unmarked nodes in the comparison mask; This indicates the penalty for exceeding the boundary of the receiving node, which is determined based on the extent to which the receiving node exceeds the receiving pressure threshold; The transfer cost representing the share of transferable demand; This represents the conservation correction penalty, determined based on the magnitude of the paired correction. , , , These are the weight parameters. , , , It can be normalized to a dimensionless value, which facilitates weighted calculation.

[0116] Candidate scheduling actions that meet preset benefit conditions and whose intervention conservative supply and demand pressure at the receiving node does not exceed the receiving pressure threshold are selected to generate a parking space scheduling strategy. The parking space scheduling strategy includes at least one of the following: high supply and demand pressure area early warning, available area recommendation, parking guidance recommendation, inspection priority suggestion, and shared parking space opening suggestion. The parking space scheduling strategy outputs aggregated parking demand and does not output individualized judgment results for specific vehicles or users.

[0117] When the low-risk probing conditions are met, a low-proportion probing scheduling action is generated along the edge of the asymmetric scheduling response to be calibrated. The low-risk probing conditions include: the source node is not a node with extremely high supply and demand pressure; the lower bound of the supply forecast for the receiving node is greater than a preset safety margin; the probing transfer ratio is lower than a preset probing ratio threshold; and the expected intervention conservative supply and demand pressure of the receiving node after executing the low-proportion probing scheduling action does not exceed a receiving pressure threshold. The preset probing ratio threshold can be set to 5% to 15% to avoid excessive disturbance caused by the probing scheduling.

[0118] After executing a low-proportion trial dispatch action, the actual pressure change at the source node, the actual berth status change at the receiving node, and the actual acceptance of the transferable demand share are obtained within a feedback window. The feedback window can be one or more time slices, such as 15 minutes, 30 minutes, or 60 minutes. The actual acceptance of the transferable demand share is estimated through aggregated data, not by identifying individual vehicles. In one implementation, the actual acceptance is estimated based on the aggregated entry increment, berth availability change, and historical natural fluctuation baseline differential of the receiving node within the feedback window. The historical natural fluctuation baseline can be determined by the average entry volume or average berth availability change of the same receiving node under the same week type, similar time period, similar weather conditions, and without trial dispatch. For example: ; Indicates from the source node To the receiving node The estimated actual volume of transactions. Indicates the receiving node Aggregated entry increments within the feedback window Indicates the receiving node The baseline for natural entry under similar historical conditions. All quantities mentioned above refer to berth demand or aggregated entry quantities, and are of consistent dimensions.

[0119] The trial response residual is obtained based on the deviation between the actual pressure change, the actual berth status change, the actual throughput, and the intervention prediction trajectory. If the actual throughput is lower than expected by the intervention prediction trajectory, or if the actual pressure rise at the throughput node exceeds expectations, the trial response residual will be larger; if the actual feedback is close to the intervention prediction trajectory, the trial response residual will be smaller.

[0120] The historical scheduling response rate, historical post-scheduling pressure decline magnitude, historical post-scheduling congestion magnitude, scheduling response edge weight, and response differential label are calibrated based on the trial response residual. In one implementation, the scheduling response edge weight can be updated as follows: ; Indicates from the source node Point to the receiving node The calibrated scheduling response edge weights, Indicates the edge weights of the scheduling response before calibration; This indicates that the edge weights of the observed responses obtained from the low-proportion trial scheduling feedback can be determined by the source node. Actual pressure drop, receiving node The actual berth status change, actual acceptance volume, and trial response residuals are jointly determined, if the source node The actual pressure has decreased significantly, and the supporting nodes If there is no congestion and the actual capacity is high, then Higher, if the receiving node If congestion occurs or the actual capacity is low, then Lower; This represents the update rate, which can range from 0.05 to 0.3. This means that the result will be limited to between 0 and 1.

[0121] The response differential label can be updated based on the actual change within the same feedback window. For example, the expected pressure drop at the source node can be aligned with the actual pressure drop, and the expected pressure increase at the receiving node can be aligned with the actual pressure increase. The upper limit of the demand share transfer can also be adjusted based on the actual receiving volume. The calibrated scheduling response edge weights and response differential labels are used for constructing the scheduling response dynamic graph, generating candidate scheduling actions, constructing benchmark and intervention comparison graphs, and determining berth scheduling strategies for the next forecast period. Through this feedback calibration mechanism, the system can continuously correct the scheduling response dynamic graph using aggregated berth status without relying on vehicle-level tracking, gradually bringing berth supply and demand forecasts and scheduling strategies closer to the actual operating state.

[0122] In one preferred embodiment, the action-aware spatiotemporal graph neural network model is obtained by jointly training counterfactual scheduling gain and constraint loss. The training samples include historical samples in which no candidate scheduling actions have been applied and historical scheduling feedback samples in which historical scheduling actions have been applied.

[0123] The joint training includes:

[0124] The prediction loss is obtained based on the deviation between the baseline predicted trajectory and the actual supply and demand trajectory when the candidate scheduling action was not applied, and the deviation between the intervention predicted trajectory and the actual supply and demand change trajectory corresponding to historical scheduling feedback data. Loss can be predicted using mean absolute error, mean squared error, or a weighted combination of both.

[0125] The deviation based on the local conservation relationship yields the conservation constraint loss. Based on the supply and demand pressure deviation between the baseline predicted trajectory and the intervention predicted trajectory within the range of berth nodes not marked by the control difference mask, the leakage loss outside the mask is obtained. ;

[0126] The boundary violation penalty item for the receiving node is determined based on the amount of conservative supply and demand pressure exceeding the receiving pressure threshold. The transfer cost penalty item is determined based on the candidate transfer ratio and transfer cost corresponding to the candidate scheduling action.

[0127] Predicting losses Conservation constraint loss Losses due to mask leakage The model training objective is formed by the combination of the node out-of-bounds penalty, the transfer cost penalty, and the counterfactual scheduling benefit. , represented as:

[0128] ;

[0129] In the formula, Indicates candidate scheduling actions The set of receiving nodes involved, Indicates candidate scheduling actions The set of source nodes involved; This represents the total number of time slices within the future forecast window; Indicates the execution of candidate scheduling actions. Subsequent Node In the future Intervention in a specific timeframe to mitigate supply and demand pressures Indicates the receiving node The corresponding bearing pressure threshold; Represents the positive part function; Indicates candidate scheduling actions Source node To the receiving node The proportion of candidate transfers, Indicates that it is from the source node To the receiving node The transfer cost; Indicates candidate scheduling actions Counterfactual scheduling benefits; , , , and This represents the weighting parameter.

[0130] Among the above training objectives, predicting loss. Conservation constraint loss Losses due to mask leakage The out-of-bounds penalty for receiving nodes, the transfer cost penalty, and the counterfactual scheduling benefit can all be normalized before training to ensure that all items have the same or comparable dimensionless scale. The positive part function indicates that a penalty is only incurred when the intervention conservative supply and demand pressure of the receiving node exceeds the receiving pressure threshold. Training can use the Adam or AdamW optimizer, with a learning rate of 0.0001 to 0.001, a batch size of 16 to 128, and 50 to 300 training epochs. The training, validation, and test sets are divided chronologically to avoid future data leakage.

[0131] , , , and These are all non-negative, dimensionless weight parameters used to adjust the relative impact of conservation constraint loss, mask leakage loss, out-of-bounds penalty for receiving nodes, transfer cost penalty, and counterfactual scheduling benefit on the model training objective. Since the original dimensions or numerical scales of each loss and benefit term may differ, normalization is required before determining the weight parameters.

[0132] In one implementation, historical means or quantiles of each item are calculated on the training set, and each item is converted into a dimensionless normalized quantity. For example, the prediction loss, conservation constraint loss, mask leakage loss, out-of-bounds penalty for receiving nodes, transfer cost penalty, and counterfactual scheduling benefit are divided by the corresponding training set mean, standard deviation, or preset quantile, respectively, so that each item is on a comparable numerical scale. After normalization, all items in the model training objective are dimensionless, and the weight parameters are also dimensionless, with consistent dimensions on both sides of the formula.

[0133] In one implementation, the predicted loss can be... The weight is fixed at 1, and the remaining weight parameters are determined by a comprehensive evaluation metric on the validation set. The comprehensive evaluation metric may include prediction error, local conservation violation rate, mask leakage magnitude, out-of-bounds rate of receiving nodes, average transfer cost, and the number of nodes with high supply and demand pressure after scheduling. Weight parameters can be searched within a preset candidate range, for example: ;in, The candidate range is higher than some other weights in order to prioritize avoiding the receiving node from exceeding the receiving pressure threshold after scheduling. The candidate range is relatively low to avoid excessive suppression of effective scheduling actions by transfer costs. The above value range can be adjusted according to different cities, road densities, and parking management strategies.

[0134] In another implementation, grid search, random search, Bayesian optimization, or early stopping on the validation set can also be used to determine the weight parameters. Specifically, the model is trained on the training set, and a comprehensive evaluation index is calculated on the validation set. The set of weight parameters that optimizes the comprehensive evaluation index on the validation set is selected as the final weight parameters. If it is necessary for the weight parameters to automatically adjust during training, each weight parameter can be set as a trainable parameter, and its non-negative value can be guaranteed by the following constraints: ; , This represents the weight parameter corresponding to the loss or gain item. and These represent the preset lower limit and upper limit, respectively. This represents the trainable parameters. In this way, the weight parameters are always limited to a preset range, avoiding negative or excessive weights that could lead to model training instability.

[0135] In actual training, empirical initial values ​​can be used first, for example... , , , and Then adjust using the validation set. If a large number of out-of-bounds nodes appear in the validation set, increase the limit. If several predicted trajectories exhibit significant perturbations at unlabeled nodes in the control difference mask, then the accuracy will improve. If the model output frequently violates local conservation relationships, then improve... If the model tends to select candidate scheduling actions with excessively high transfer costs, then improve... If the model is too conservative and fails to generate effective scheduling benefits, then the efficiency should be appropriately increased. .

[0136] In summary, this invention can simultaneously assess the impact of scheduling actions on future supply and demand changes during the on-street parking supply and demand forecasting process, reduce the interference of natural supply and demand fluctuations on the judgment of scheduling benefits, improve the accuracy of high-pressure area identification, available area selection and parking guidance recommendation, and continuously revise the scheduling basis through low-risk trial scheduling feedback, reduce the risk of overload in available areas, and improve the stability and reliability of on-street parking resource scheduling.

[0137] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural networks, characterized in that, The method includes: Acquire multi-source parking data in on-street parking scenarios and form time-slice data of parking space nodes according to preset time granularity; Based on the time slice data, the reliable berth status is reconstructed, the berth supply and demand time sequence pattern features are extracted, and the future transferable parking demand is characterized as the transferable demand share. Based on the trusted berth status, berth supply and demand time sequence pattern characteristics, and historical scheduling feedback data in the time slice data, a scheduling response dynamic graph is constructed. Candidate scheduling actions are generated according to the source node, receiving node, scheduling response edge weight and transferable demand share in the scheduling response dynamic graph. Paired benchmark comparison graphs and intervention comparison graphs are constructed for the candidate scheduling actions. Shared encoding and differential encoding are performed on the baseline control map and the intervention control map to obtain common state representation and action effect representation. Together with the graph action variables corresponding to the candidate scheduling actions, they are used as prediction conditions input into the action perception spatiotemporal graph neural network model, so that the graph action variables participate in the graph propagation, time update and prediction output process. Under the conservation transfer constraint of the transferable demand share, the baseline prediction trajectory and the intervention prediction trajectory are obtained synchronously. Based on the counterfactual scheduling benefits of the baseline predicted trajectory and the intervention predicted trajectory, a berth scheduling strategy is determined, and the scheduling response dynamic diagram is calibrated through low-risk trial scheduling feedback.

2. The berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural network according to claim 1, characterized in that, The time slice data includes at least occupancy observation data, order statistics data, operational statistics data, external context data, and historical scheduling feedback data; Based on the occupancy observation data from at least two types of occupancy observation sources at the same berth node in the same time slice, and the dynamic reliability of the data source determined by the online status, observation delay, inter-source consistency and historical error feedback of the occupancy observation sources, the state confidence of the credible berth state is reconstructed and the state confidence of the credible berth state is generated. Based on the order statistics, operational statistics, external context data and historical scheduling feedback data, berth supply and demand time series pattern features, including supply and demand phase relationship, peak and valley migration, turnover fluctuation, context response elasticity and historical scheduling response, are extracted. Based on the source berth node, estimated arrival time slice, acceptable transfer cost, supply and demand time sequence pattern label, acceptance constraint satisfaction, state confidence, and historical scheduling response, a transferability score for the future predicted parking demand is calculated. The future predicted parking demand is then adaptively divided into multiple transferable demand shares according to the transferability score. The sum of multiple transferable demand shares belonging to the same source berth node and the same estimated arrival time slice does not exceed the future predicted parking demand of that source berth node in that estimated arrival time slice.

3. The berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural network according to claim 2, characterized in that, The construction of the scheduling response dynamic graph includes: Basic berth edges are generated based on the geographical distance and road connectivity between berth nodes, and time-series pattern similar edges are generated based on the similarity of berth supply and demand time-series pattern features between berth nodes. Within the acceptance screening range defined by the basic berth edge and the time sequence pattern similarity edge, berth nodes whose supply and demand pressure exceeds the preset pressure threshold are identified as source nodes, and berth nodes with supply margin and whose state confidence meets the preset confidence conditions are identified as acceptance nodes. Based on the historical scheduling response rate, the magnitude of pressure reduction after historical scheduling, and the magnitude of congestion after historical scheduling within the same historical scheduling feedback window, an asymmetric scheduling response edge is generated from the source node to the receiving node. Based on the historical scheduling response rate, the magnitude of pressure reduction after historical scheduling, the magnitude of congestion after historical scheduling, the state confidence of the receiving node, the acceptable transfer cost of the transferable demand share, and the transferability score, the scheduling response edge weight of the asymmetric scheduling response edge is determined. Configure response differential labels for the asymmetric scheduling response edge, including the expected pressure decrease of the source node, the expected pressure increase of the receiving node, the upper limit of the demand share transfer, and the demand share transfer direction; The basic berth edge, the time-series pattern similar edge, and the asymmetric scheduling response edge are merged into the scheduling response dynamic graph.

4. The berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural network according to claim 3, characterized in that, Generate candidate scheduling actions, and construct paired baseline and intervention comparison maps for the candidate scheduling actions, including: In the asymmetric scheduling response edge, based on the scheduling response edge weight, response differential label and transferable demand share, candidate transfer node pairs and candidate transfer ratios of transferable demand shares between candidate transfer node pairs are determined, as well as the transferable demand share transfer amount determined by the candidate transfer ratios and transferable demand shares. If the transferable demand share transfer amount does not exceed the upper limit of the demand share transfer amount, the candidate transfer node pair, scheduling response edge weight, candidate transfer ratio, transferable demand share and response differential label are combined into a candidate scheduling action and encoded as the graph action variable; The scheduling response dynamic diagram without applying candidate scheduling actions is determined as the baseline comparison diagram, and the scheduling response dynamic diagram formed by transferring the share of transferable demand along the asymmetric scheduling response edge after applying candidate scheduling actions is determined as the intervention comparison diagram. When constructing the baseline comparison map and the intervention comparison map, the graph structure and node characteristics that are unrelated to the candidate scheduling actions are kept consistent, and the distribution of transferable demand share, scheduling response edge weight, initial value of supply and demand pressure of source node and initial value of supply and demand pressure of receiving node are changed according to the response differential label. A contrast mask is generated based on the response differential label to limit the scope of differential coding.

5. The berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural network according to claim 4, characterized in that, The baseline and intervention control maps are subjected to shared and differential coding, including: The baseline control map and the intervention control map are encoded using a parameter-sharing graph encoder to obtain the baseline map code and the intervention map code, respectively. The common-state representation is obtained by sharing pooling the coding regions in the baseline and intervention maps that are not marked by the control difference mask; the action effect representation is obtained by differential coding the coding regions in the baseline and intervention maps that are marked by the control difference mask. The graph action variables are used to perform gating correction on the action effect representation so that the action effect representation retains only the transferable demand share transfer effect, scheduling response edge weight change effect, and initial supply and demand pressure change effect related to the candidate scheduling action. The common-state representation and the gated action effect representation are used as prediction conditions for the action-aware spatiotemporal graph neural network model.

6. The berth supply and demand forecasting and scheduling method based on spatiotemporal graph neural network according to claim 5, characterized in that, The graph action variables are involved in the graph propagation, time update, and prediction output processes, including: Gating analysis is performed on the graph action variables to obtain the edge propagation gate, node update gate, and trajectory output gate; During graph propagation, the baseline comparison graph and the intervention comparison graph are processed using shared propagation parameters respectively, and the edge propagation gate is used to control the action message transmission along the asymmetric scheduling response edge in the intervention comparison graph, so that the transferable demand share participates in the graph propagation from the source node to the receiving node according to the demand share transfer direction; During the time update process, the node update gate is used to write the expected pressure decrease of the source node and the expected pressure increase of the receiving node as a pair of state changes into the future prediction window, so that the candidate scheduling action participates in the time state recursion of the source node and the receiving node. During the prediction output process, the common-state representation and the gated-corrected action effect representation are controlled by the trajectory output gate to enter the baseline output branch and the intervention output branch, respectively. The baseline output branch suppresses the action effect representation, and the intervention output branch introduces the action effect representation within the range defined by the control difference mask, thereby obtaining the initial baseline prediction trajectory and the initial intervention prediction trajectory.

7. A berth supply and demand forecasting and scheduling method based on a spatiotemporal graph neural network according to claim 6, characterized in that, Under the conservation transfer constraint of the transferable demand share, the baseline prediction trajectory and the intervention prediction trajectory are obtained simultaneously, including: Within the same future prediction window, based on the response differential label and the comparison difference mask, a local conservation relationship is established for the reduction in source node demand, the increase in receiving node demand, and the transferable demand share caused by the candidate scheduling action. When the initial intervention prediction trajectory does not satisfy the local conservation relationship, the source node demand prediction value and the receiving node demand prediction value are corrected in pairs only within the range of the source node, receiving node and corresponding asymmetric scheduling response edge defined by the contrast difference mask, through the conservation redistribution mechanism. The pair correction is constrained by the upper limit of the demand share transfer amount and by the expected pressure decrease of the source node and the expected pressure increase of the receiving node, so that the demand decrease of the source node and the demand increase of the receiving node correspond to each other within the future prediction window. The local conservation relationship is constrained by the following formula: ； In the formula, Indicates berth node In the future The baseline transferable demand share for each time slice, Indicates the execution of candidate scheduling actions. Rear berth node In the future Intervention in a single time slice can transfer a share of demand; Indicates the execution of candidate scheduling actions Later in the future Within a time slice, the berth nodes Transfer to berth node The transferable share of demand; Indicates the execution of candidate scheduling actions Later in the future Within a time slice, the berth nodes Transfer to berth node The transferable share of demand; Indicates the current time slice The scheduling response dynamic diagram is composed of berth nodes. Pointing to berth node Scheduling response edge weight; The intervention prediction trajectory is generated based on the paired modified share of transferable demand for the intervention, and the baseline prediction trajectory is generated based on the baseline share of transferable demand when the candidate scheduling action is not applied.

8. A berth supply and demand forecasting and scheduling method based on a spatiotemporal graph neural network according to claim 7, characterized in that, Determining berth scheduling strategies based on the counterfactual scheduling gains of the baseline predicted trajectory and the intervention predicted trajectory includes: Based on the baseline prediction trajectory and the intervention prediction trajectory, the baseline conservative supply and demand pressure and the intervention conservative supply and demand pressure are calculated respectively. Within the scope of the differential mask, the scheduling benefits within the mask for candidate scheduling actions are determined based on the conservative decrease in supply and demand pressure of the source node, the conservative increase in supply and demand pressure of the receiving node, the transfer cost of the transferable demand share, and the state confidence of the receiving node. Within the berth node range where the difference mask is not marked, the mask leakage penalty for candidate scheduling actions is determined based on the supply and demand pressure deviation between the baseline predicted trajectory and the intervention predicted trajectory. The counterfactual scheduling benefit is determined based on the scheduling benefit within the mask, the leakage penalty outside the mask, the outbound penalty for receiving nodes exceeding the receiving pressure threshold, and the conservation correction penalty generated by pairwise correction. The candidate scheduling action that satisfies the preset benefit conditions and whose intervention conservative supply and demand pressure at the receiving node does not exceed the receiving pressure threshold is selected to generate the parking space scheduling strategy. The parking space scheduling strategy includes at least one of the following: high supply and demand pressure area early warning, receiving area recommendation, parking guidance recommendation, inspection priority suggestion, and shared parking space opening suggestion. The parking space scheduling strategy is output to aggregated parking demand.

9. A berth supply and demand forecasting and scheduling method based on a spatiotemporal graph neural network according to claim 8, characterized in that, The action-aware spatiotemporal graph neural network model is obtained through joint training of counterfactual scheduling reward and constraint loss, wherein the joint training includes: Based on the deviation between the baseline predicted trajectory and the actual supply and demand trajectory when the candidate scheduling action was not applied, and the deviation between the intervention predicted trajectory and the actual supply and demand change trajectory corresponding to historical scheduling feedback data, the prediction loss is obtained. Based on the deviation of the local conservation relationship, the conservation constraint loss is obtained. Based on the supply and demand pressure deviation between the baseline predicted trajectory and the intervention predicted trajectory within the berth node range not marked by the control difference mask, the leakage loss outside the mask is obtained. ; Based on the amount of supply and demand pressure exceeding the threshold of the receiving node's intervention, determine the boundary violation penalty item for the receiving node, and determine the transfer cost penalty item based on the candidate transfer ratio and transfer cost corresponding to the candidate scheduling action. The predicted loss Conservation constraint loss Losses due to mask leakage The model training objective is formed by the combination of the node out-of-bounds penalty, the transfer cost penalty, and the counterfactual scheduling benefit. , represented as: ； In the formula, Indicates candidate scheduling actions The set of receiving nodes involved, Indicates candidate scheduling actions The set of source nodes involved; This represents the total number of time slices within the future forecast window; Indicates the execution of candidate scheduling actions. Subsequent Node In the future Intervention in a specific timeframe to mitigate supply and demand pressures Indicates the receiving node The corresponding bearing pressure threshold; Represents the positive part function; Indicates candidate scheduling actions Source node To the receiving node The proportion of candidate transfers, Indicates that it is from the source node To the receiving node The transfer cost; Indicates candidate scheduling actions Counterfactual scheduling benefits; , , , and This represents the weighting parameter.

10. A berth supply and demand forecasting and scheduling method based on a spatiotemporal graph neural network according to claim 8, characterized in that, The scheduling response dynamic graph is calibrated through low-risk trial scheduling feedback, including: When the low-risk probing conditions are met, a low-proportion probing scheduling action is generated along the asymmetric scheduling response edge to be calibrated. The low-risk probing conditions include: the source node is not a node with extremely high supply and demand pressure, the lower bound of the supply prediction of the receiving node is greater than the preset safety margin, the probing transfer ratio is lower than the preset probing ratio threshold, and the expected intervention conservative supply and demand pressure of the receiving node after the low-proportion probing scheduling action is executed does not exceed the receiving pressure threshold. After executing the low-proportion trial scheduling action, the actual pressure change of the source node, the actual berth status change of the receiving node, and the actual amount of transferable demand share are obtained in the feedback window. The actual amount of transferable demand share is obtained based on the aggregated entry increment, the change in idle berths, and the historical natural fluctuation baseline differential estimation of the receiving node in the feedback window. The trial response residual is obtained based on the deviation between the actual pressure change, the actual berth status change, the actual load, and the intervention prediction trajectory. Based on the trial response residual, calibrate the historical scheduling response rate, historical pressure drop after scheduling, historical congestion absorption after scheduling, scheduling response edge weight, and response differential label corresponding to the asymmetric scheduling response edge. The calibrated scheduling response edge weights and response difference labels are used for the construction of the scheduling response dynamic graph, generation of candidate scheduling actions, construction of benchmark comparison graphs and intervention comparison graphs, and determination of berth scheduling strategies in the next prediction cycle.

Images