Intelligent network car-hailing dispatching system

Abstract

The application relates to the technical field of operation management, in particular to an intelligent network car-hailing dispatching system which comprises an aggregation deviation identification module, a dispatching anomaly detection module, a path density extraction module, an optimal path generation module and a dispatching path balancing module. According to the application, the spatial and temporal joint evaluation of the aggregation position changes of multiple time points in the passenger call request can realize the accurate characterization of the human flow dynamic characteristics, effectively distinguish the normal human flow density and the sudden aggregation behavior, and improve the sensing sensitivity of the abnormal travel demand. The feature vector construction of the regional order load state and the crowd aggregation deviation behavior is combined with the multi-dimensional feature fusion judgment of the isolation forest algorithm, so that the discrimination accuracy of the regional dispatching pressure is significantly enhanced, and the misjudgment risk caused by the single index fluctuation is avoided. On this basis, the structured extraction of the trajectory data in the dispatching abnormal area is combined with the direction density and the space node density judgment, so that the real high-frequency path can be quickly located.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of operation management, and particularly relates to an intelligent network car-hailing dispatching system. BACKGROUND

[0002] The operation management technical field to which the intelligent network car-hailing dispatching system belongs contains relevant technical contents of planning, coordinating and controlling operation resources in traffic travel services. The core content of the technical field is to realize dynamic optimization and management of resources such as vehicles, manpower, time and routes in travel services through informatization means.

[0003] Among them, the intelligent network car-hailing dispatching system refers to a dispatching mechanism for establishing a service connection between a user initiating a travel request through a mobile Internet platform and a vehicle providing travel services. The technical matters targeted by the system cover order receiving and classification, vehicle availability identification, passenger location determination, distance calculation between vehicles and passengers, dispatching priority sorting, real-time path comparison and travel time evaluation and the like.

[0004] The existing intelligent network car-hailing dispatching system is mainly constructed around traditional processes such as order receiving, vehicle matching, distance calculation and path estimation, and its identification of passenger gathering behavior relies on the static distribution of request locations, lacks the ability to analyze the gathering dynamics under time series, and cannot timely reflect the passenger flow change signals in sudden scenarios. For example, a large number of passengers request in a short time due to sudden activities in a certain place, and the traditional system cannot perceive the spatial shift characteristics behind it, resulting in response lag. The identification of dispatching abnormalities also highly depends on single-factor judgment of order quantity and vehicle distribution, which is easy to cause misidentification due to temporary index fluctuation, and then causes uneven resource scheduling. In terms of path selection, the conventional method is mostly based on simple frequency aggregation of static path or historical trajectory on the map, and cannot finely extract and judge the path distribution characteristics of abnormal areas, which is easy to mix in redundant routes and affect the scheduling efficiency. The path structure generation also does not fully consider the node connectivity and direction consistency, which may cause the generated path to have problems such as turning back and multi-node jitter, affecting the execution stability. More importantly, the existing scheme does not establish a linkage adjustment mechanism between path load and real-time vehicle distribution, and it is difficult to implement dynamic guidance and adjustment of high-pressure paths, which is easy to cause dispatching pressure aggregation in peak periods, limiting the service scalability of the overall dispatching system. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art, and to propose an intelligent network car-hailing dispatching system.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: an intelligent network car-hailing dispatching system, the system comprising:

[0007] The aggregation offset identification module collects passenger call request data to obtain a current aggregation position sequence of the passenger, performs spatial and time sequence change analysis on the aggregation position sequence at multiple times, and calculates an aggregation position offset value of the sequence;

[0008] The scheduling exception detection module obtains order load information of a current scheduling area, and forms a region state vector after normalizing the aggregation position offset value, calculates an abnormal factor of the region state vector, and outputs a corresponding scheduling exception state instruction according to the abnormal factor;

[0009] The path density extraction module filters abnormal region trajectory segments in vehicle trajectory data records through the scheduling exception state instruction, and identifies a path density candidate set through direction density and node density of the trajectory segments;

[0010] The optimal path generation module performs scheduling path filtering through the trajectory segments in the path density candidate set, and generates a network car scheduling path adjustment result.

[0011] The application improves that the aggregation position offset value is specifically a position change amplitude, a coincident aggregation density and a time sequence fluctuation intensity, the scheduling exception state instruction includes a region burst level, a trigger time node and a scheduling control label, the path density candidate set is specifically a node spatial density segment, a direction angle change segment and a path return aggregation area, and the network car scheduling path adjustment result includes a feasible path set, a key node sequence and a path reachable relationship.

[0012] The application improves that the aggregation offset identification module includes:

[0013] The request parameter extraction submodule collects initial positioning coordinates of a passenger call request, response time delay of a candidate vehicle and map click frequency within 1 minute before the request, calculates a Euclidean distance change rate between the request positioning coordinates, a total amount of click frequency and a sample median of the response time delay, and obtains a behavior trajectory basic parameter set;

[0014] The trajectory variation identification submodule identifies a mutation point section of the Euclidean distance change rate, the total amount of click frequency and the response time delay in the behavior trajectory basic parameter set, extracts a mean difference on both sides of the mutation point section, and obtains a trajectory change difference index group;

[0015] The offset value calculation submodule calculates a joint vector module value of the mean difference of each parameter according to the mean difference of the parameters before and after the section in the trajectory change difference index group, and obtains the aggregation position offset value.

[0016] The application improves that the scheduling exception detection module includes:

[0017] The dispatch order load extraction submodule obtains three index parameters of a specified area order request frequency, a proportion of a number of dispatchable vehicles and an average response time delay in a current dispatch area, performs numerical normalization on the three indexes, and obtains standardized order load information of the dispatch area.

[0018] The state vector construction submodule combines the standardized order load information of the dispatch area and the normalized aggregated position offset value into a feature vector structure, and obtains a region state vector.

[0019] The abnormality level determination submodule inputs the region state vector into an isolation forest algorithm model, calculates a region abnormality score of a path of each sample, performs interval judgment according to the abnormality score and a burst level threshold standard, and outputs a dispatch abnormality state instruction.

[0020] The present application improves that the path density extraction module comprises:

[0021] The trajectory data acquisition submodule acquires vehicle driving trajectory records of a specified time in a marked region in the dispatch abnormality state instruction, extracts a first and last boundary node and all continuous spatial points in each trajectory, calculates a sequence of adjacent distances of the continuous spatial points, a sequence of direction angle changes and a cumulative number of turn-back nodes, and obtains a group of trajectory structure feature parameters.

[0022] The trajectory segment screening submodule calls the group of trajectory structure feature parameters, a target region boundary and a region abnormality score in the dispatch abnormality state instruction, screens segments in which path nodes are continuously distributed in a specified region range in the trajectory data, and obtains a path segment set in the region.

[0023] The dense structure identification submodule calculates a mean value of direction angle changes and a trajectory node distribution density for each trajectory in the path segment set in the region according to a specified length, judges whether the two parameters simultaneously exceed a dense determination standard, and constructs a path dense candidate set.

[0024] The present application improves that the optimal path generation module comprises:

[0025] The trajectory structure analysis submodule calls a node sequence of each trajectory in the path dense candidate set, judges a spatial connectivity relationship between continuous nodes and a direction vector change trend under the premise of limiting path accessibility, and obtains a group of node structure change parameters.

[0026] The polyline node screening submodule screens a line segment interval in which a continuous direction change value is lower than a direction change threshold value according to the group of node structure change parameters, screens nodes in the target interval by using a Douglas-Pok polyline simplification algorithm, and generates a path simplified node sequence.

[0027] The path reconstruction checking sub-module performs trajectory graph structure reconstruction based on the path simplification node sequence, detects the connection relationship between nodes and the path connectivity integrity, and obtains the online car-hailing scheduling path adjustment result.

[0028] The application further comprises a scheduling path balancing module, which performs load balancing determination on the online car-hailing scheduling path adjustment result and real-time vehicle distribution state, optimizes scheduling pressure distribution through allocation adjustment instructions, and constructs a scheduling candidate path set.

[0029] The scheduling candidate path set specifically refers to a load evaluation path, a vehicle transfer path and a regional balancing path.

[0030] The application further comprises that the scheduling path balancing module comprises:

[0031] The vehicle state extraction sub-module calls the target area identifier and trajectory structure of each path in the online car-hailing scheduling path adjustment result, collects real-time positioning data of all idle vehicles in the current period and whether they are in a order-accepting state, and obtains a path corresponding vehicle distribution set.

[0032] The scheduling pressure calculation sub-module calculates the number of idle vehicles on a specified distance for each path trajectory structure according to the path corresponding vehicle distribution set, and generates a path scheduling pressure index group.

[0033] The path allocation adjustment sub-module identifies an abnormal path with a pressure value exceeding the standard deviation range of other paths based on the path scheduling pressure index group, determines the vehicle allocation direction and target path in combination with the idle vehicle transferability condition, and generates a scheduling candidate path set.

[0034] Compared with the prior art, the application has the following advantages and positive effects:

[0035] This invention achieves accurate characterization of dynamic pedestrian flow characteristics by jointly evaluating the spatial and temporal changes in the aggregation location of passenger ride-hailing requests at multiple time points. This effectively distinguishes between normal pedestrian density and sudden aggregation behavior, improving the sensitivity to abnormal travel demand. By constructing feature vectors from regional order load status and crowd aggregation shift behavior, and combining this with the isolated forest algorithm for multi-dimensional feature fusion judgment, the accuracy of regional scheduling pressure assessment is significantly enhanced, avoiding the risk of misjudgment caused by fluctuations in a single indicator. Furthermore, by structurally extracting trajectory data within abnormal scheduling areas and combining directional density and spatial node density judgments, the true high-frequency paths can be quickly located, eliminating interference from invalid and repetitive trajectories, thus improving the decision-making efficiency of path selection. Moreover, through continuous analysis of path direction change trends and simplified optimization of polyline nodes, a refined expression of the scheduling path structure can be achieved, compressing path complexity and reducing scheduling path computation costs while ensuring accessibility. In addition, by linking the results of scheduling route adjustments with the real-time vehicle distribution status, the system can dynamically determine the route load balance and guide idle vehicles to perform reasonable transfers and allocations, effectively alleviating the spatial concentration problem of scheduling resources and improving the overall balance of regional service capabilities. Attached Figure Description

[0036] Figure 1 This is a system module diagram proposed in this invention;

[0037] Figure 2 This is a system framework diagram proposed in this invention;

[0038] Figure 3 This is a schematic diagram of the aggregation offset recognition module of the present invention;

[0039] Figure 4 This is a schematic diagram of the scheduling anomaly detection module of the present invention;

[0040] Figure 5 This is a schematic diagram of the path density extraction module of the present invention;

[0041] Figure 6 This is a schematic diagram of the optimal path generation module of the present invention;

[0042] Figure 7 This is a schematic diagram of the scheduling path balancing module of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0045] Referring to Figure 1 The present application provides a technical solution: an intelligent online car dispatching system, the system comprising:

[0046] The aggregation offset identification module collects passenger call request data to obtain the current aggregation position sequence of the passenger, analyzes the spatial and temporal changes of the aggregation position sequence at multiple times, and calculates the aggregation position offset value of the sequence;

[0047] The dispatch anomaly detection module obtains the order load information of the current dispatch area, and forms a regional state vector after normalizing the aggregation position offset value, calculates the anomaly factor of the regional state vector, and outputs the corresponding dispatch anomaly state instruction according to the anomaly factor;

[0048] The path dense extraction module filters the abnormal area trajectory segment in the vehicle trajectory data record through the dispatch anomaly state instruction, and identifies the path dense candidate set through the direction density and node density of the trajectory segment;

[0049] The optimal path generation module filters the dispatch path through the trajectory segment in the path dense candidate set, and generates an online car dispatch path adjustment result;

[0050] The aggregation position offset value is specifically the position change amplitude, the coincident aggregation density, and the time sequence fluctuation intensity. The dispatch anomaly state instruction includes the regional burst level, the trigger time node, and the dispatch control label. The path dense candidate set is specifically the node spatial density segment, the direction angle change segment, and the path return aggregation area. The online car dispatch path adjustment result includes the feasible path set, the key node sequence, and the path accessibility relationship.

[0051] Referring to Figure 2 and Figure 3 The aggregation offset identification module comprises:

[0052] The request parameter extraction submodule collects the initial positioning coordinates of the passenger call request, the candidate vehicle response time delay, and the map click frequency within 1 minute before the request, calculates the Euclidean distance change rate between the request positioning coordinates, the total amount of click frequency, and the sample median of the response time delay, and obtains the behavior trajectory basic parameter set;

[0053] Based on the passenger call records continuously received in the background of the dispatching system, three types of original data are extracted at the time of each request, including the initial positioning coordinates uploaded by the passenger terminal, the response time delay records generated by the platform assigned candidate vehicles, and the click frequency of the passenger on the map area within 1 minute before the request. The positioning coordinates are processed by GPS sampling to remove drift, and are converted to latitude and longitude with six decimal places. The response time delay is uniformly in milliseconds, and the time difference between receiving the request and issuing the vehicle candidate list is recorded through the dispatch platform. The click frequency is counted by the UI operation log with a minute window, for example, a passenger zooms in on the map 6 times and clicks on the recommended pickup point 2 times within a minute, so the frequency is 8. Then, the positioning point time series data is constructed, the Euclidean distance between each two consecutive coordinates is calculated, and the average change rate of the entire sequence is calculated. If each segment is 40 meters, 60 meters, 90 meters, and 70 meters, the change rate is (40+60+90+70) / 4 = 65 meters. The click frequency is directly recorded as 8 times. The median of the response time delay sample is extracted from the sorted ascending order of the time delay values of all requests in the current dispatch area, for example, if the sample is 31, the 16th value after sorting is 1100 milliseconds, and the median is 1100. The complete set of behavior trajectory basic parameters is obtained.

[0054] The trajectory variation recognition submodule identifies the mutation point segments of the Euclidean distance change rate, the total amount of click frequency, and the response time delay in the behavior trajectory basic parameter set, and extracts the mean difference on both sides of the mutation point segment to obtain the trajectory change difference index group.

[0055] The trajectory basic parameter set is called, and time series change vectors are constructed for the three types of data. The time period is divided into several intervals using the equal interval sliding window method, for example, each hour of trajectory sampling is divided into 12 windows, each window is 5 minutes, the average value of the data in each window is calculated, the Euclidean distance change rate, the total amount of click frequency, and the response time delay are calculated, and the mean value change between adjacent windows is calculated. The mutation window segment is identified by combining the Bayesian variable point detection method, for example, if the Euclidean distance change rate of the 6th window is 0.45 and the 7th window jumps to 1.2, the jump is more than 0.75, which is determined as a mutation. If the response time delay changes from 950ms to 1800ms, the difference is 850ms, and the click frequency changes from 3 to 7, the difference is 4 times, these mutation segments are retained, and the mean difference of the two windows before and after the change is calculated to obtain the trajectory change difference index group. The numerical example is as follows: the Euclidean distance change rate difference is 0.75, the response time delay difference is 850 milliseconds, and the click frequency difference is 4.

[0056] The offset value calculation submodule calculates the joint vector module value of the mean difference of each parameter according to the mean difference of the parameters in the trajectory change difference index group, and obtains the aggregated position offset value.

[0057] According to the three mean difference parameters in the trajectory change difference index group, respectively set as Δ1, Δ2 and Δ3, wherein Δ1 represents the mean difference of the Euclidean distance change rate, the unit is kilometer per minute (km / min), Δ2 represents the mean difference of the response time delay, the unit is millisecond (ms), and Δ3 represents the mean difference of the click frequency, the unit is times per minute (times / min). In order to avoid the interference of non-uniformity to the result, normalization processing needs to be performed on each parameter, and the normalized value is calculated by using the minimum-maximum normalization method, and the normalization calculation formula is: Among them, Δ i represents the i-th original mean difference value, i=1, 2, 3 corresponds to the Euclidean distance change rate, the response time delay and the click frequency; min(Δ i ) represents the minimum value of the i-th in all samples, max(Δ i ) represents the maximum value; Δ′ i is the normalized value, which is uniformly mapped to the interval [0, 1]. If the actual difference values of the three indicators in a certain trajectory are: Δ1=1.1 km / min, Δ2=1200 ms, Δ3=6 times / min, and the maximum and minimum values in the sample are Δ1∈[0.1, 1.8], Δ2∈[300, 2500], Δ3∈[2, 10], then: The normalized three results are used for modulus value calculation of the degree of deviation, and the modulus value calculation formula is as follows: Among them, V represents the final aggregation position deviation value, which is a dimensionless value in the normalized scale; Δ′1, Δ′2, Δ′3 are the three normalized parameters, which represent the unified vector scale representation of the spatial deviation intensity of the aggregation behavior, the system response time delay fluctuation degree and the user interface interaction frequency change degree. The above normalized value is substituted into the calculation to obtain: The final aggregation position deviation value V=0.873 is obtained, and when the value exceeds the set deviation threshold V0=0.7 (the threshold is set by statistics on the 90th percentile of the deviation value of the historical stable area sample), the region is determined to exist the mutation state of passenger request aggregation behavior.

[0058] Please refer to Figure 2 and Figure 4 , the dispatching anomaly detection module comprises:

[0059] The dispatch order load extraction submodule obtains the specified area order request frequency, the number of dispatchable vehicles proportion and the average response time delay in the current dispatch area, performs numerical normalization processing on the three indicators, and obtains the standardized order load information of the dispatch area;

[0060] The running data in the specified area of the current area is obtained, first, the number of orders per unit time in each grid is extracted according to the system preset regional grid (such as 0.25 square kilometers per grid), the order request frequency per unit area is calculated, for example, the number of orders generated in a certain grid in 5 minutes is 30, the frequency is 6 orders / minute / square kilometer, at the same time, the number of idle schedulable vehicles in the system and the total number of vehicles in the region are extracted, the schedulable vehicle ratio is calculated, if the total number of vehicles is 100 and the number of idle vehicles is 40, the ratio is 40%, in addition, the response time delay of all response records in the current region in the last minute is counted, and the arithmetic mean is taken as the average response time delay, for example, a total of 20 responses, the total time delay is 24000 milliseconds, the average is 1200 milliseconds, then the three indicators are normalized, the normalization adopts the maximum and minimum scaling method, assuming that the order frequency range is [1, 10], the vehicle ratio range is [0.1, 0.9], and the response time delay range is [300, 3000], the corresponding sample value is 6, 0.4, and 1200, the normalized result is that the order frequency is 0.56, the vehicle ratio is 0.375, and the response time delay is 0.333, and finally the scheduling area standardized order load information is generated.

[0061] The state vector construction submodule combines the scheduling area standardized order load information and the normalized aggregated position offset value into a feature vector structure to obtain the region state vector.

[0062] Based on the scheduling area standardized order load information, the aggregated position offset value generated in the previous module is structured and spliced to form a multi-dimensional vector structure, and after the data form is unified, it is used as the region state vector. This process needs to ensure that the four dimension parameter units are consistent, and the format is standardized to a unified sample expression format. Assuming that the current standardized order frequency is 0.56, the vehicle ratio is 0.375, the response time delay is 0.333, and the aggregated position offset value is 0.873, the final state vector is [0.56, 0.375, 0.333, 0.873], and a one-dimensional array form is constructed according to the data format requirements of the scheduling system data management platform, which is used as input to the model structure recognition module to form the region state vector.

[0063] The abnormal level determination submodule inputs the region state vector into the isolation forest algorithm model, calculates the regional anomaly score of each sample path, performs interval judgment according to the anomaly score and the burst level threshold standard, and outputs the scheduling abnormal state instruction.

[0064] The region state vector is input into the isolation forest algorithm model, first, an isolation forest model composed of n=100 sub-trees is constructed from the trajectory data collected by the system and the scheduling index sample set, each sub-tree randomly selects m=256 state vectors from the total sample set without replacement to form a training subset Each tree randomly selects a split dimension from the input state vector dimension (originally 4 dimensions, including order request frequency, proportion of schedulable vehicles, average response time, and aggregated location offset value) during the construction process, generates a random split threshold between the minimum and maximum values in that dimension, and uses it as the basis for node division. The tree structure is recursively constructed until the number of samples under the current node is 1 or the tree depth reaches the maximum depth Each vector x to be detected is input into each sub-tree T i of the forest, and according to its characteristic value, it is selected to enter the left or right sub-tree according to the split threshold value. The path depth from the root node to the belonging leaf node is recorded as h i (x). To enhance the timeliness sensitivity of the model to sudden behavior, combined with the dynamic response characteristics of the aggregated location offset value and the scheduling index in the actual online taxi system, a sample generation timestamp dimension τ is introduced into the state vector, expanding the original four-dimensional state vector to a five-dimensional state vector where: x1: normalized unit area order request frequency (dimensionless, normalized to [0, 1]), x2: normalized proportion of schedulable vehicles (dimensionless, normalized to [0, 1]), x3: normalized average response time (dimensionless, normalized to [0, 1]), x4: normalized aggregated location offset value (dimensionless, normalized to [0, 1]), τ: the difference between the generation time of the state vector and the current scheduling time, in minutes, used to measure the recency of the state vector.

[0065] The calculation formula of the regional anomaly score is as follows:

[0066]

[0067] where s(x): the regional anomaly score, representing the average isolation degree of the state vector in the multi-tree structure, the larger the value, the more abnormal; n: the total number of trees in the forest, set to 100; h i (x): the path split depth of the state vector x in the i-th tree; c(m): a normalization factor to standardize the depth under different trees, defined as where H(k) represents the k-th harmonic number, which can be calculated by obtained; a: time enhancement coefficient, determines the influence strength of time term in overall anomaly score, set range is [0.5, 2.0], in dispatching system, a is set according to historical burst area reaction delay experience, if system average response delay sensitive period is 10 minutes, then recommended setting is 1.2-1.5; β: time attenuation rate, used to control the decreasing speed of sample timestamp influence on anomaly score, usually set to 0.05, 0.2], in practice, if sample influence significantly attenuates to 1 / 3 after 15 minutes, β can be set to 0.1; τ: interval of state sample from current dispatching time, such as sample record is 13:55, current time is 14:00, then τ=5. Assuming that the path depth of a certain state vector in 10 trees is recorded as [5, 6, 5, 5, 6, 6, 5, 7, 6, 5], the average path depth is 5.6, the sample number m=256, the normalization coefficient c(m)≈8.22, the time enhancement coefficient a=1.2, the time attenuation rate β=0.1, and the current sample time interval τ=5, then s(x)=(1+1.2·e -0.1·5 )·2 -5.6 / 8.22 ≈1.087, the final regional anomaly score is s(x)=1.087, combined with the dispatching system set level judgment standard: if s(x)>1.0, it is judged as "significant anomaly", if 0.7

[0068] Please refer to Figure 2 and Figure 5 , the path dense extraction module comprises:

[0069] The trajectory data acquisition submodule acquires the vehicle driving trajectory record of the marked area in the dispatching anomaly state instruction at the specified time, extracts the first and last boundary nodes and all continuous space points in each trajectory, calculates the adjacent distance sequence, direction angle change sequence and cumulative number of turn-back nodes of the continuous space points, and obtains the trajectory structure feature parameter group;

[0070] Based on the dispatch abnormal state instruction, the marker area number and the specified time window are extracted as the retrieval conditions, and the matching area and the historical trajectory record of all vehicles in the time period are retrieved from the trajectory database. The trajectory record is sampled at intervals of 5 seconds, and each trajectory is composed of a group of continuous spatial points, including vehicle number, timestamp, latitude and longitude, three basic information. The system extracts the structure of each trajectory data, selects the first time point corresponding to the coordinate as the first node, and the last time point corresponding to the coordinate as the tail node to form the trajectory boundary node, and extracts all continuous trajectory points between the two nodes to form the trajectory point sequence. Then, the spatial distance between each two adjacent nodes is calculated in time sequence, and recorded as the adjacent distance sequence, with the unit of meter. Then, the direction angle sequence is calculated based on the latitude and longitude change of the two nodes, and the direction angle is obtained by using arctan2(Δy, Δx) method, and is converted into radian value in the interval [0, 2π). At the same time, it is judged whether the direction angle mutation in the trajectory exceeds π radian to identify the turning point. Each time a mutation is found, it is counted as 1 turning. The number of all turning points in the whole trajectory is counted, and finally the trajectory structure feature parameter group corresponding to each trajectory is output, including: node number, average adjacent distance, average and standard deviation of direction angle change sequence, and three index data of turning node cumulative number. For example, trajectory X contains 90 nodes, the average adjacent distance is 11.4 meters, the direction angle standard deviation is 0.96 radian, and the turning node is 5.

[0071] The trajectory segment screening submodule calls the trajectory structure feature parameter group and the target area boundary and the area abnormal score in the dispatch abnormal state instruction, screens the segments in which the path nodes continuously distributed in the specified area range, and obtains the path segment set in the area.

[0072] The trajectory structure feature parameter group is called, and the target area boundary range and the area abnormal score (set to 1.087) provided in the dispatch abnormal state instruction are combined to perform path segment screening processing. First, a polygon coordinate range model is constructed according to the area boundary profile, and whether the node position coordinates in each trajectory are within the polygon range is judged in turn. If the proportion of the continuous nodes in the trajectory that fall within the range exceeds the set value (set to 60%), the continuous trajectory nodes in this segment are extracted as the path segment. Then, the level is confirmed according to the abnormal score field in the dispatch abnormal state instruction. The abnormal score value 1.087 exceeds the set significant abnormal threshold 1.0, so only the path segment set with the abnormal level of “significant” in the area is retained as the priority marked path. If the node numbers 25 to 48 in a trajectory continuously fall within the area range, and the total number of nodes in the trajectory is 35, the proportion of the path is 68.6%, which meets the retention condition. The system classifies it into the path segment set in the area, which is used as the input content of the next module to participate in the subsequent dense structure recognition processing.

[0073] The dense structure identification submodule calculates the average change of direction angle and the distribution density of trajectory nodes for each trajectory in the path segment set in the region according to the specified length, judges whether the two parameters exceed the dense determination standard at the same time, and constructs a path dense candidate set;

[0074] The structure density determination operation is performed on the path segment set in the region. First, each trajectory is divided according to the path length, the unit path segment length is set as L=100 meters, the continuous trajectory nodes in each length segment are extracted to form a node sequence, and the average change of direction angle and the distribution density of trajectory nodes are calculated for the node sequence. The average change of direction angle is based on the difference between every two continuous trajectory nodes, the direction angle is calculated by using the arctangent function, and is defined as θ q q =arctan2(y q+1 -y q ,x q+1 -x q ), where q=1, 2,..., k-1 represents the qth pair of trajectory points, the unit is radian, and the change of direction angle is δ q q =|θ q+1 -θ q |, and the average change of direction angle of all angle changes in the node segment is recorded as the average change of direction angle of the segment.

[0075] Suppose that a segment of trajectory contains 5 nodes, and the coordinate direction change angles are: [θ1=0.20, θ2=0.36, θ3=0.65, θ4=0.44], then the changes of direction angle are: δ1=|0.36-0.20|=0.16, δ2=|0.65-0.36|=0.29, δ3=|0.44-0.65|=0.21, and the average change of direction angle is: radian, and the node distribution density is calculated by using the two-dimensional kernel density estimation method (KDE) of the Gaussian kernel function, which is in the form of:

[0076]

[0077] wherein ρ(x, y): represents the estimated trajectory node density at position (x, y), the unit is piece / m 2 ; k: the total number of nodes in the path segment, the unit is piece; (x q , y q ): the qth trajectory node coordinate, the unit is meter, q=1,..., k; h: bandwidth, controls the influence range of each node, the unit is meter, and is set as h=5.0 m; (x, y): the position of the current kernel function action point, the unit is meter.

[0078] Let k = 3 nodes in the path segment, coordinates: P1 = (0, 0), P2 = (2, 1), P3 = (4, 0), select target point (x, y) = (2, 0.5), calculate the density estimate at this point: for q = 1: (x - x1) 2 +(y - y1) 2 =(2-0) 2 +(0.5-0) 2 =4+0.25=4.25, for q = 2: (x - x2) 2 +(y - y2) 2 =(2-2) 2 +(0.5-1) 2 =0+0.25=0.25, for q = 3: (x - x3) 2 +(y - y3) 2 =(2-4) 2 +(0.5-0) 2 =4+0.25=4.25, Substitute the above values:

[0079]

[0080] Then perform numerical integration on the entire trajectory minimum bounding rectangle region (e.g. 10m x 5m = 50m 2 ) with 1m step grid, let the total integral value be ∫∫ Ω ρ(x, y) dxdy ≈ 32.8, where, ρ represents the average node distribution density of this trajectory segment, unit: individual / m 2 ; Ω: represents the minimum containing rectangular region composed of all trajectory node point sets; A eff : is the area of the rectangular region, unit: m 2 , which can be calculated from the vertex coordinates; ∫∫ Ω ρ(x, y) dxdy: represents the integral total value of the kernel density function in this region, unit: individuals. Then the average node distribution density is: Let the dense determination criterion of the system be the average change threshold of the direction angle θ th = 0.28 radians, the node distribution density threshold ρ th = 0.65 individuals / m 2 , the average direction angle of this trajectory segment is 0.31 radians, and the node distribution density is 0.656 individuals / m 2 , both of which exceed the threshold, so the system determines this trajectory segment as a dense segment according to this, and finally all path segments that meet the dual determination conditions are included in the path dense candidate set.

[0081] Please refer to Figure 2 and Figure 6 The optimal path generation module comprises:

[0082] The trajectory structure analysis submodule calls each trajectory node sequence in the path dense candidate set, judges the spatial connectivity relationship and direction vector change trend between the continuous nodes under the premise of limiting path accessibility, and obtains a node structure change parameter group;

[0083] When obtaining the input data, first, rely on the dense segment result output by the previous recognition module, extract the trajectory number, node number, sampling time sequence and node latitude and longitude coordinates one by one, and construct a trajectory linked list sequence in timestamp order. Each trajectory is reconstructed into a directed point set chain structure in node order, and the constraint rule of path accessibility is set on this basis. The determination standard is: if the Euclidean distance between two consecutive nodes is less than 2 kilometers, and there is a passable road segment between the two points in the regional road topology map, it is considered as an accessible path. The 2-kilometer threshold is set according to the average block long-distance distance of urban trunk roads, and the distance distribution under the 95% confidence level of the historical trajectory data set is used as a reference sample to exclude abnormal long-jump segments and set boundaries. Then, the direction vector is constructed between each two consecutive nodes, the direction vector components are obtained through the difference between the node horizontal and vertical coordinates, and the angle between each three consecutive direction vectors is formed. The angle change value is obtained by the cosine angle calculation method, and is arranged in order as the direction change sequence. After the direction angle change sequence is extracted, the system simultaneously labels the position number of the current node in the trajectory, and generates the node structure change parameter group according to the trajectory ID, node number, direction change value and node straight line distance. The angle threshold is extracted from the typical feature statistics of the direction mutation in the past city-level network car turning behavior, and the significant turning standard is set to 30° as the direction mutation benchmark. Within 20° is considered as a linear stable segment, which is used as the data basis of the subsequent screening condition. Finally, the node sequence containing the structure trend label is obtained and the node structure change parameter group is output.

[0084] The polyline node screening submodule screens the line segment interval with continuous direction change value lower than the direction change threshold according to the node structure change parameter group, and uses the Douglas-Pok polyline simplification algorithm to screen the nodes in the target interval to generate a path simplified node sequence;

[0085] According to the node structure change parameter group, the line segment interval with a continuous direction change value lower than the direction change threshold value is screened out. In the execution process, first, the angle change sequence collected in the node structure change parameter group is read. For each group of angle values, it is identified in sequence whether it is less than the set direction change threshold value. The setting of the direction change threshold value is based on the statistical distribution of the vehicle turning behavior of the urban trunk road. Among 10,000 historical trajectory samples, 80% of the path segments maintain a straight line trend within an angle change of less than 30° in the turning angle distribution. Therefore, the direction change judgment criterion is set to 30°, which corresponds to π / 6 radians. When the continuous angle change in a certain segment of the trajectory is 12°, 18°, 24°, etc., it is marked as a gentle line segment interval, and all node sequences in the interval are output as candidate sections. For each candidate section, the Douglas-Peucker polyline simplification algorithm is called. The perpendicular distance from the middle node in the section to the connecting line of the start and end nodes is calculated, and the distance tolerance value ε is set to 8 meters. The tolerance value is referenced from the device sampling error and the upper limit of the cumulative tolerance of the path point error. At a sampling rate of 5 seconds, the average travel distance of the vehicle at medium-low speed is 8-10 meters. Therefore, 8 meters is used as the node screening criterion. If the distance from a certain node to the straight line is less than 8 meters, it is determined as a redundant node and is removed. Otherwise, it is retained as a key deformation node. After the node screening is completed, the system reassembles the screened and retained nodes and merges them according to the trajectory number.

[0086] The path reconstruction verification sub-module reconstructs the trajectory graph structure based on the simplified node sequence, detects the connection relationship between nodes and the integrity of the path connectivity, and obtains the network car dispatching path adjustment result.

[0087] The trajectory graph structure reconstruction is performed based on the simplified node sequence. In the initial step, the system first receives the simplified trajectory node sequence, and performs path connectivity verification on all adjacent node pairs in the node sequence. The method is to call the graph edge index in the preloaded road topology graph structure for any node pair (q, q+1), and query whether there is a legal road edge directly connecting the two nodes. If there is edge information and the distance between the start and end nodes is less than the set maximum distance threshold, it is considered to be connected. The maximum distance threshold set here is 500 meters according to the statistical connection distance of typical road sections of vehicles, which is based on the 95% quantile of the average road section length distribution of urban sub-trunk roads, excluding the influence of instantaneous long-distance jump sections caused by GPS drift. If a node pair does not meet the connection condition, the system will call back all nodes contained in the section in the original trajectory, calculate the Euclidean distance between them and the current breakpoint node, and retrieve the node that is not more than 300 meters away from the previous node and is connectable in the topology graph as a substitute node to insert into the connection path. The 300-meter judgment is based on the average small road connection distance in urban high-density areas. After the path reconstruction is completed, the system generates a new trajectory connection graph according to the final node structure, the node set contains all preserved and supplemented nodes, and the edge set contains all legal connection edges. The structure integrity of the trajectory graph structure is verified and the structure graph is output for use in the net car dispatching path adjustment scene to form the final structured dispatching path result.

[0088] Please refer to Figure 2 and Figure 7 The dispatching path balancing module also includes a dispatching path balancing module that performs load balancing determination on the net car dispatching path adjustment result and the real-time vehicle distribution state, optimizes the dispatching pressure distribution by distributing adjustment instructions, and constructs a dispatching candidate path set.

[0089] The dispatching candidate path set specifically refers to a load evaluation path, a vehicle transfer path, and a regional balancing path.

[0090] The dispatching path balancing module includes:

[0091] The vehicle state extraction submodule calls the target area identifier and trajectory structure of each path in the net car dispatching path adjustment result, collects the real-time positioning data of all idle vehicles in the current period and whether they are in a orderable state, and obtains the vehicle distribution set corresponding to the path.

[0092] The vehicle state extraction submodule calls the target area identifier and trajectory structure of each path in the net car dispatching path adjustment result, first analyzes each trajectory number, spatial node sequence, and dispatching area number corresponding to the path in the path output result, and accesses the vehicle state database with these information as index conditions. The database stores the positioning state and dispatching order switch flag field of all idle vehicles on the platform in real time. In the extraction process, the system first obtains the UTC timestamp of the current system period, which is set to tnow and read all the vehicle current position (x q ,y q ), vehicle unique number v q , whether the vehicle is marked as available for order state δ q =1 (available for order) or δ q =0 (not available for order); in the actual processing, the system filters out all the vehicles with δ q =1 as the current available idle vehicle set, and forms a vehicle point cloud distribution set in coordinate form; then, according to the target area number of each path in the path adjustment result, the spatial matching is performed to construct the area spatial index, the covered area of the trajectory is defined as the buffer polygon area, and the spatial inclusion judgment is performed between the available vehicle point set and the trajectory polygon, and the vehicle coordinates falling into the trajectory buffer area are used as the judgment standard, the system sets the path buffer width to 200 meters, and sets the reference city road average lane width and vehicle static waiting distribution density, if a vehicle position falls into the buffer area, the vehicle is classified into the vehicle distribution set corresponding to the path; for example, the trajectory number TR023 corresponds to the area ID A15, and the corresponding trajectory buffer area covers 23 idle vehicles, the system binds it with the path ID and outputs it as the vehicle distribution set of the path, and finally all the path corresponding vehicle distribution sets are organized in hash structure.

[0093] The dispatch pressure calculation submodule calculates the number of idle vehicles in a specified distance for each path trajectory structure according to the path corresponding vehicle distribution set, and generates a path dispatch pressure index group;

[0094] According to the path corresponding vehicle distribution set, the number of idle vehicles in a specified distance for each path trajectory structure is calculated, first read the vehicle distribution set and the path coordinate structure, process each trajectory in coordinate order, the system divides the path into nodes with a sliding window of 100 meters, generates a number of micro segments, each micro segment region is constructed with the segment as the center line and the buffer area with a width of 200 meters for matching the number of idle vehicles in the spatial range; the number of matching vehicles in each segment buffer area is denoted as n q , the system traverses all the micro segments on the whole trajectory to generate a vehicle number sequence [n1, n2,..., n Q ], where Q is the total number of micro segments; the system further calculates the arithmetic mean of the vehicle number sequence as the average vehicle coverage density of the path, denoted as μ p , and calculates its standard deviation σ p, which measures the fluctuation of vehicle coverage distribution; the segment distance and buffer width here are determined according to the urban road network distribution model and the average vehicle stay point density, and the coverage density of 1.8 vehicles per 100 meters ± 0.9 in the historical trajectory data sample is used as a reference to determine the distribution pattern, and 100 meters is used as the basic statistical window; if the trajectory TR045 is divided into 15 segments, the vehicle coverage number sequence is [2, 3, 2, 1, 2, 0, 1, 1, 4, 3, 3, 2, 2, 1, 0], the coverage mean of the path is 2 vehicles per segment, and the standard deviation is 1.05, and the system records the mean and fluctuation as one of the path scheduling pressure indicators, and finally the system generates μ p , σ p , the maximum vehicle vacancy segment, the minimum coverage density segment, and other pressure-related structural information, forming a path scheduling pressure indicator group.

[0095] The path allocation adjustment submodule identifies abnormal paths with pressure values exceeding the standard deviation range of other paths based on the path scheduling pressure indicator group, and determines the vehicle allocation direction and target path based on the transferability of idle vehicles to generate a scheduling candidate path set.

[0096] Based on the path scheduling pressure indicator group, abnormal paths with pressure values exceeding the standard deviation range of other paths are identified, the system first reads the pressure indicator group of all paths, extracts the coverage mean μ p and standard deviation σ p of each path, calculates the statistical distribution of each indicator in the entire path set, and performs a normalized comparison operation on each path; for example, if the average coverage density of the entire path set is 2.3 and the standard deviation is 0.7, the system sets the high-pressure path limit as follows: any path with a coverage density lower than (2.3-1.5×0.7=1.25) is considered an abnormal path with scheduling pressure; after the judgment is completed, the system further calls the idle vehicle distribution map to identify all idle vehicles located in low-pressure paths and within 500 meters of the high-pressure path area as transferable vehicles, and the system sets the 500-meter transfer radius based on the urban walking scheduling radius and the average vehicle response distance model. In the process of setting the transfer path direction, the system constructs a directional path structure from the low-pressure path to the high-pressure path based on the current orientation of the transferable vehicle and the node position of the path it belongs to, filters all idle vehicles that meet the consistency of direction, distance accessibility, and vehicle state to form a transferable vehicle set, and generates a target path list according to the path ID, and finally outputs the matched scheduling candidate path set.

[0097] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in other forms. Any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes, and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, without departing from the technical solution content of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. An intelligent network car-hailing dispatching system, characterized in that: The system comprises: The aggregation offset identification module collects passenger call request data to obtain the current aggregation position sequence of the passenger, performs spatial and time sequence change analysis on the aggregation position sequence at multiple times, and calculates the aggregation position offset value of the sequence; The dispatching anomaly detection module obtains the order load information of the current dispatching area, and forms a regional state vector after normalizing the aggregation position offset value, calculates the anomaly factor of the regional state vector, and outputs the corresponding dispatching anomaly state instruction according to the anomaly factor; The path dense extraction module filters the abnormal regional trajectory segment in the vehicle trajectory data record through the dispatching anomaly state instruction, and identifies the path dense candidate set through the direction density and node density of the trajectory segment; The optimal path generation module filters the dispatching path through the trajectory segment in the path dense candidate set, and generates a network car dispatching path adjustment result; Further comprising a dispatching path balancing module, which performs load balancing judgment on the network car dispatching path adjustment result and real-time vehicle distribution state, optimizes the dispatching pressure distribution through allocation adjustment instruction, and constructs a dispatching candidate path set; The dispatching path balancing module comprises: The vehicle state extraction submodule calls the target area identifier and trajectory structure of each path in the network car dispatching path adjustment result, collects the real-time positioning data of all idle vehicles in the current period and whether they are in a state of being able to accept orders, and obtains a vehicle distribution set corresponding to the path; The dispatching pressure calculation submodule calculates the number of idle vehicles within a specified distance for each path trajectory structure according to the vehicle distribution set corresponding to the path, and generates a path dispatching pressure index group; The path allocation adjustment submodule identifies abnormal paths with pressure values exceeding the standard deviation range of other paths based on the path dispatching pressure index group, determines the vehicle allocation direction and target path in combination with the transferability condition of idle vehicles, and generates a dispatching candidate path set. 2.The intelligent network car-hailing dispatching system of claim 1, wherein: The aggregation position offset value is specifically a position change amplitude, a coincident aggregation density, and a time sequence fluctuation intensity. The dispatching anomaly state instruction includes a regional burst level, a trigger time node, and a dispatching control label. The path dense candidate set is specifically a node spatial density segment, a direction angle change segment, and a path return aggregation area. The network car dispatching path adjustment result includes a feasible path set, a key node sequence, and a path accessibility relationship. 3.The intelligent network car-hailing dispatching system of claim 1, wherein: The aggregation offset identification module comprises: The request parameter extraction submodule collects the initial positioning coordinates of the passenger call request, the response time delay of the candidate vehicle, and the map click frequency within 1 minute before the request, calculates the Euclidean distance change rate between the request positioning coordinates, the total amount of click frequency, and the sample median of response time delay, and obtains a behavior trajectory basic parameter set; The trajectory variation identification submodule identifies the mutation point section of the Euclidean distance change rate, the total amount of click frequency, and the response time delay in the behavior trajectory basic parameter set, and extracts the mean difference on both sides of the mutation point section to obtain a trajectory change difference index group; The offset value calculation submodule calculates the joint vector module value of the mean difference of each parameter according to the mean difference of the parameters in the trajectory change difference index group, and obtains the aggregation position offset value.

4. The intelligent network car dispatching system of claim 1, wherein: The dispatching anomaly detection module comprises: The dispatch order load extraction submodule obtains three index parameters of a specified area order request frequency, a proportion of a number of dispatchable vehicles, and an average response time delay in a current dispatch area, performs numerical normalization processing on the three indexes, and obtains standardized order load information of the dispatch area; The state vector construction submodule combines the standardized order load information of the dispatch area and the normalized aggregated position offset value into a feature vector structure, and obtains an area state vector; The abnormal level determination submodule inputs the area state vector into an isolation forest algorithm model, calculates an area abnormal score of a path of each sample, performs interval judgment according to an abnormal score and a burst level threshold standard, and outputs a dispatch abnormal state instruction. 5.The intelligent network car-hailing dispatching system of claim 4, wherein: For calculating the area abnormal score s(x) of the path of each sample, a formula is used: where s(x) represents the average isolation degree of the state vector in the multi-tree structure of the Isolation Forest algorithm model, the larger the value, the more abnormal, n is the total number of trees in the forest of the Isolation Forest algorithm model, h i (x) is the path splitting depth corresponding to the state vector x in the i-th tree of the Isolation Forest algorithm model, c(m) is used to standardize the depth under each tree of the Isolation Forest algorithm model, a is a time enhancement coefficient, which determines the influence strength of the time term in the overall abnormal score, b is a time decay rate, which is used to control the decreasing speed of the influence of the sample timestamp on the abnormal score, t is the interval of the state sample from the current scheduling time, e -β·τ is a decay function, which is used to quantify the recency of the sample.

6. The intelligent network car dispatching system of claim 1, wherein: The path density extraction module includes: The trajectory data acquisition submodule acquires vehicle driving trajectory records of a specified time in a marked area in the dispatch abnormal state instruction, extracts a first and last boundary node and all continuous spatial points in each trajectory, calculates a sequence of adjacent distances of the continuous spatial points, a sequence of direction angle changes, and a cumulative number of turn-back nodes, and obtains a group of trajectory structure feature parameters; The trajectory segment screening submodule calls the group of trajectory structure feature parameters, a target area boundary, and an area abnormal score in the dispatch abnormal state instruction, screens segments in which path nodes are continuously distributed in a specified area range in the trajectory data, and obtains a path segment set in the area; The dense structure identification submodule calculates a mean value of direction angle changes and a trajectory node distribution density for each trajectory in the path segment set in the area according to a specified length, judges whether the two parameters simultaneously exceed a dense determination standard, and constructs a path dense candidate set. 7.The intelligent network car-hailing dispatching system of claim 6, wherein: For calculating the node distribution density, a formula is used: where: p(x, y) is a trajectory node distribution density representing the estimated trajectory node distribution at position (x, y), reflecting the degree of aggregation of nodes in the region, k is the total number of nodes in the path segment, (x q ,y q ) is the qth trajectory node coordinate, h is the bandwidth, which controls the influence range of each node, and (x, y) is the position of the current kernel function action point.

8. The intelligent network car dispatching system of claim 1, wherein: The optimal path generation module includes: The trajectory structure analysis submodule calls a node sequence of each trajectory in the path dense candidate set, judges a spatial connectivity relationship between continuous nodes and a direction vector change trend under the premise of limiting path accessibility, and obtains a group of node structure change parameters; The polyline node screening submodule screens a line segment interval in which a continuous direction change value is lower than a direction change threshold value according to the group of node structure change parameters, screens nodes in the target interval by using a Douglas-Pok polyline simplification algorithm, and generates a path simplified node sequence; The path reconstruction verification submodule performs trajectory graph structure reconstruction based on the path simplified node sequence, detects a connection relationship between nodes and path connectivity integrity, and obtains a network car dispatch path adjustment result.

9. The intelligent network car dispatching system of claim 1, wherein: The dispatch candidate path set specifically refers to a load evaluation path, a vehicle transfer path, and an area balancing path.

Images