Dynamic fence planning method based on sign-in track of pile-free shared bicycle

Abstract

The invention discloses a dynamic fence planning method based on a sign-in track of a pile-free shared bicycle, which comprises the following steps of: firstly, determining candidate positions of an electronic fence by adopting a hierarchical clustering method in combination with a spatial point processing method, and then, calculating the inventory condition of each candidate position by using a series of algorithms so as to simulate the actual bicycle activity. And finally, predicting inventory capacity by using the space-time diagram neural network model, and generating a real-time electronic fence deployment strategy. The method can adapt to the spatial and temporal change of the use mode of the shared bicycle according to the sign-in track data of the user, flexibly adjusts the deployment strategy of the electronic fence, can provide a new solution for the challenge brought by the rapid expansion of a pile-free shared bicycle system, and provides a more effective technical scheme for standardizing the parking behavior of an urban area.

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic big data mining, and in particular relates to a dynamic fence planning method based on a check-in trajectory of a pile-free shared bicycle. Background Art

[0002] As shared bikes can provide urban residents with sustainable, low-carbon and affordable transportation options, more and more cities have introduced shared bike systems; however, with the rapid increase in the number of shared bikes, the management of shared bike systems has become increasingly prominent. In order to increase urban cleanliness and ensure that bicycles are properly parked in designated areas, setting up electronic fences is considered an effective means; electronic fences use GPS technology and have the advantages of low carbon and low cost. By preventing riders from parking in non-designated areas, they can help shared bike operators efficiently manage shared bikes.

[0003] Many studies have shown that the distribution of dockless shared bicycles is uneven in space and time; therefore, it is necessary to propose a dynamic electronic fence that can adapt to changes in bicycle usage patterns. The electronic fence aims to regulate the user's parking location while meeting the user's parking needs as much as possible to improve the cleanliness of the city; for dynamic electronic fences, their main feature is that compared with static electronic fences, they can significantly reduce the number of station deployments and flexibly change according to parking patterns in different time and space states.

[0004] At present, most technical solutions only plan electronic fences based on static station locations and parking requirements, without considering the spatiotemporal changes of real-time bicycle inventory. Reference [Zhang, Y., & Mi, Z. (2018). Environmental benefits of bike sharing: A big data-based analysis. Applied Energy, 220] proposes to divide the city into fixed grids to determine the location of electronic fences; Reference [Shi, X., Liang, Z., & Seng, D. (2022). An Electric Fence-Planning Framework for a Dockless Bike-Sharing System Based on a Land Parcel Subdivision and Regional Coverage Maximization. IEEE Intelligent Transportation Systems Magazine, 2–13] proposes a fence planning framework for dockless shared bicycles based on dividing urban regional plots. However, the planning of electronic fences must be based on the prediction of flexible station locations and bicycle inventory at these locations. At the same time, the virtual station and scheduling information of dockless shared bicycles cannot be directly obtained from the check-in trajectory data.

[0005] In summary, the problems and defects of the prior art are mainly as follows:

[0006] ① The existing shared bicycle fence planning method based on grid division is not flexible enough. The boundaries of the grid can easily cause spatial cluster separation, resulting in inaccurate virtual site division.

[0007] ② Derivation of parking demand based solely on origin-to-destination (OD) check-in trip data cannot reflect the real inventory of virtual stations, resulting in inaccurate capacity estimation of each electric fence.

[0008] ③ The existing static fence planning method lacks the ability to dynamically adapt to parking needs and cannot cope with the ever-changing bicycle usage patterns in time and space.

[0009] Therefore, there is an urgent need for a method to calculate and predict bicycle inventory in real time in order to more accurately plan electronic fences for dockless shared bicycles. Summary of the invention

[0010] In view of the above, the present invention provides a dynamic fence planning method based on the check-in trajectory of dockless shared bicycles, which can provide a new solution to the challenges brought about by the rapid expansion of dockless shared bicycle systems and provide a more effective technical solution for regulating parking behavior in urban areas.

[0011] A dynamic fence planning method based on the check-in trajectory of a dockless shared bicycle comprises the following steps:

[0012] (1) Determine candidate sites by clustering based on historical travel data;

[0013] (2) Calculate and determine the maximum inventory of the candidate site in each time window in the past;

[0014] (3) Construct a spatiotemporal graph convolutional neural network. The network takes the maximum inventory of each candidate site in the previous m time windows as input to predict the maximum inventory of each candidate site in the next time window, where m is a natural number greater than 1.

[0015] (4) Using the historical data obtained in step (2) to train the spatiotemporal graph convolutional neural network;

[0016] (5) The trained spatiotemporal graph convolutional neural network is used to predict the maximum inventory of each candidate site in a new time window in the future, and dynamic fence planning is performed based on the initial inventory of each candidate site in this time window.

[0017] Furthermore, the specific implementation method of step (1) is as follows: first, 24 hours a day is divided into multiple time periods. For any time period, the trajectory endpoints of all bicycle historical travel data in the time period are used to perform secondary clustering at the macro and micro levels, so that the entire city range is divided into multiple areas, and the centroid of each area is taken as the cluster center; then, the cluster centers obtained by clustering all time periods are obtained, and the center points that are too close are merged, that is, for two cluster centers A and B whose distance is less than a set threshold, the midpoint of the line connecting A and B is taken and added as a new cluster center, and A and B are deleted at the same time; finally, all cluster centers are matched to the road network, and the foot of the perpendicular between each cluster center and the nearest road line is taken as a candidate site, and the sites that are too close are merged according to the above method.

[0018] Furthermore, the specific implementation method of step (2) is as follows: first, the travel data of all bicycles in a certain time period in the past is obtained, and the time period is divided into multiple time windows. According to the travel data, the bicycles in the locked state in the first time window are counted, and the positions of these bicycles are spatially matched with each candidate station, so as to determine the initial inventory of each candidate station; for any subsequent time window, the number of bicycles that are spatially matched with the candidate station in the time window is counted according to the travel data, and the maximum number of parked bicycles is used as the maximum inventory of the candidate station. At the same time, the final number of parked bicycles is recorded as the initial inventory of the next time window for the next iteration.

[0019] Preferably, the bicycle trip data is subjected to behavior recognition, and by identifying that the previous state of the check-in point is locked and the next state is unlocked, the parking behavior between the two states is extracted from the original trajectory, and the distance between the arrival point and the starting point in the parking behavior is calculated. If the distance is less than the threshold, the parking behavior is determined to be a drift behavior; if the distance is less than the threshold, the parking behavior is determined to be a scheduling behavior, and a virtual trip is created between the arrival point and the starting point of the parking behavior and inserted into the trip data. By correcting the drift data and inserting the virtual trip, the phenomenon of GPS positioning deviation and bicycle operator scheduling in the real world can be truly simulated, making the inventory value calculation more accurate and reasonable.

[0020] Furthermore, the spatiotemporal graph convolutional neural network is composed of a first gated time convolution layer, a graph convolution layer, a second gated time convolution layer, and a fully connected layer connected in sequence, wherein the first gated time convolution layer and the second gated time convolution layer are used to convolve their respective inputs in the time dimension, and the graph convolution layer uses a weighted adjacency matrix to convolve the input in the spatial dimension. The input dimensions of the first three layers are all m×n, and the fully connected layer outputs a 1×n prediction result after comprehensive mapping, where n is the number of candidate sites.

[0021] Furthermore, the specific implementation method of step (4) is as follows: first, a graph structure is established with candidate sites as nodes, and any time window corresponds to a graph structure. The maximum inventory of the candidate site in the time window is the value of the corresponding node, and the distance between the candidate sites is the weight value of the edge between the corresponding nodes; then, historical data is used to calculate the weight of the edge in the time window t1~t m The graph structure is used as the input of the spatiotemporal graph convolutional neural network, with a time window t m+1 The graph structure is used as a label, and then the network parameters are calculated and updated by the gradient descent method according to the loss function between the network output prediction result and the label, thereby completing an iteration; Based on the above, a fixed step size sliding time window (i.e., t2~t m+1 The graph structure is input, t m+2 The graph structure is labeled) for the next iteration, so that the network parameters are continuously updated and optimized.

[0022] Furthermore, the dimension of the weighted adjacency matrix is ​​n×n, wherein the calculation expression of the element value of the i-th row and the j-th column is as follows:

[0023]

[0024] Where: σ is the standard deviation, dist(i,j) is the semi-haversine distance between candidate sites i and j, i and j are the index numbers of the candidate sites, and k is the distance threshold.

[0025] Furthermore, in the step (5), for any candidate site, the predicted maximum inventory is graded to determine the level of the maximum inventory of the candidate site, and it is judged whether the initial inventory of the candidate site in the new time window reaches the maximum inventory level: if so, the number of dynamic fences arranged in the candidate site area is increased; if not, the number of dynamic fences arranged in the candidate site area is reduced.

[0026] Based on the above technical solution, the present invention has the following beneficial technical effects:

[0027] 1. The present invention well combines the characteristics of time and space in large-scale check-in point trajectories to find adaptive fence planning candidate points for the distribution of shared bicycles.

[0028] 2. The present invention takes into account the rebalancing behavior in the real world and accurately simulates the real inventory of the virtual site.

[0029] 3. The present invention uses a dynamically adaptive modeling approach to predict inventory capacity and generate a real-time and dynamic electronic fence deployment plan.

[0030] Therefore, the method of the present invention can adapt to the spatiotemporal changes in the shared bicycle usage pattern according to the user's check-in trajectory data, and flexibly adjust the deployment strategy of the electronic fence. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the flow chart of the dynamic fence planning method of the present invention.

[0032] Figure 2 Schematic diagram of the structure of the spatiotemporal graph convolutional neural network of the present invention. DETAILED DESCRIPTION

[0033] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0034] like Figure 1 As shown, the dynamic fence planning method based on the check-in trajectory of the dockless shared bicycle of the present invention includes the following steps:

[0035] Step S1: Determination of candidate sites. In the selected time period, the endpoints of the shared bicycle check-in data are extracted, the hierarchical clustering method is used to select the cluster center, and the cluster center merging algorithm is used to merge the cluster centers of different periods. The cluster centers are placed on the road network through the road network matching algorithm, and the candidate fences are determined. Specifically:

[0036] S11: Use hierarchical clustering method to mine check-in point centers with different densities.

[0037] Firstly, the data is divided into six time periods according to the distribution of travel volume on weekdays and weekends to accurately identify the time periods with higher frequency of bicycle use. Then, according to the divided time periods, the space is divided into areas of different density at the macro level, and each area is further subdivided at the micro level. In each level of clustering, the K-medoids algorithm is used to take the shared bicycle unlocking point as the centroid to mine the parking point centers with different densities in the shared bicycle dataset.

[0038] S12: Use the cluster center merging method to obtain candidate points.

[0039] First, according to the minimum limit distance L between two points min Then, sort the parking point centers by the number of adjacent points within L; min All points of are merged and their geometric centers are taken; finally, these points are traversed and the point list is updated to ensure that no adjacent points are within each other's L min After this step, candidate points are obtained.

[0040] S13: Match the candidate points with the nearest road network.

[0041] All candidate points are matched with the nearest road network line, and then the perpendicular foot of the point and the line is found, and the candidate points are merged again using the cluster center merging method.

[0042] Step S2: Inventory time series calculation. In order to be close to the real world, the present invention proposes a simulation scheduling algorithm and determines the initial number of parked bicycles in the first time window; then, the inventory calculation algorithm is used to establish a label value for each candidate station in each time window from the recorded data. The specific process is as follows:

[0043] S21: Identify the dispatching behavior and restore the dispatching check-in trajectory. First, by identifying that the previous state of the check-in point is locked and the next state is unlocked, the parking behavior (DO) is extracted from the original trajectory; then, the distance between the arrival point D and the starting point O in the parking behavior is calculated, and the threshold is selected as 200 meters to distinguish between positioning drift and dispatching behavior. When the distance is less than the threshold, it is regarded as drifting behavior, and the longitude and latitude of the arrival point and the starting point are set to the same value; when the distance is greater than the threshold, it is regarded as dispatching behavior, and a virtual trip (DO) with the same start and end point is created and inserted into the check-in trajectory data.

[0044] S22: Calculate the initial inventory quantity of candidate points. First, sort the bicycle records according to bicycle identity ID and time to obtain the first record of each bicycle ID; then, evaluate the lock status and save the unlock record; finally, spatially match the bicycle data with the location of the candidate point; Based on the above steps, we can determine the initial number of bicycles for each candidate site within the first time window.

[0045] S23: Calculate the inventory quantity of candidate sites in each time window. For each time window, two values ​​are generated: one is the maximum number of parked bicycles, which will be output; the other is the final number of parked bicycles, which will be used for the next iteration. Each iteration process includes: first, the bicycle location space is matched, for each record, the distance to the candidate site is calculated, and the nearest candidate site is matched; then the number of bicycles at the candidate site in each time window is grouped, and the bicycle records in each time period group and candidate site group during the observation period are aggregated; finally, the inventory of each candidate site is equal to the initial number of each time window plus the number of changes, which is continuously updated by recording data.

[0046] Step S3: Spatiotemporal graph prediction. Build a graph structure for the current scenario and use the spatiotemporal graph convolutional network (STGCN) to predict the inventory of candidate sites; according to the real-time inventory of the sites, determine the level of the candidate sites and decide whether to activate them; if activated, the capacity of the site will be determined by its level.

[0047] S31: Construct a spatiotemporal graph of candidate sites. Based on the maximum inventory quantity of each time window generated in the above steps, the sites are regarded as nodes in the graph structure, and edges are established based on the distances between the sites to establish a weighted graph structure for the candidate sites. Specifically: define an undirected connected graph G = (V, E, W), W: E → R + ∪0, where V=(v1,v2,…,v n ) is a set of candidate sites, E is a set of edges representing sites; the weighted adjacency matrix is ​​represented by W, w(i,j) is weighted by the distance between site i and site j, where w(i,j) is defined as follows:

[0048]

[0049] Among them: k and σ 2 is the threshold used to control the connectivity between candidate sites, σ is the standard deviation of the distance matrix, dist(i,j) is the haversine distance between candidate sites i and j, R is the radius of the earth, usually set to 6371 km, and x and y are the latitude and longitude of the candidate site, respectively.

[0050] Finally, based on the above weighted adjacency matrix and time series stock value, in the time period T = {t1,…,t m}Establish a series of candidate site graph structures Gs={G t1 ,…,G tm}.

[0051] S32: Predict candidate site inventory using spatiotemporal convolutional neural network. First, the time series inventory data and spatial connectivity matrix of each candidate site are input into Figure 2 The spatiotemporal graph convolutional network shown in the figure predicts the inventory value of each site in the future time window by learning the fusion features of spatiotemporal data; then, the inventory value of each site in the future time window is predicted using the time window from t1 to t h The historical inventory is taken from the image, and 128-dimensional high-level features are extracted by passing each node through a 2-layer ST Conv module. The ST Conv module includes two layers of gated temporal convolution layers and one layer of graph convolution layers, using three layers of channels of 64, 16, and 64 respectively; finally, these features are input into two fully connected layers with 128 and 1 channels respectively, for predicting the future time window t h+△ The bicycle inventory at each station in .

[0052] S33: Determine the capacity level of candidate sites and plan dynamic fences. First, determine a series of inventory values ​​as the thresholds for the capacity level of candidate sites. Then, assign a level to each site based on the real-time inventory results of each candidate site predicted by the spatiotemporal graph convolutional network. If the predicted result is less than the threshold Thr, it means that the site is not activated. Finally, the location and level predictions of all candidate sites are used as the dynamic fence planning results.

[0053] In order to test the performance of the inventory forecasting model invented by this party, we used the evaluation indicators widely used in time series forecasting tasks for verification, including mean absolute error (MAE), root mean square error (RMSE) and weighted mean absolute percentage error (WMAPE), which are used to measure the forecast error. The smaller the error value, the better the performance. This verification example compares the following six benchmark methods on the data sets of two regions: historical average method HA, exponential smoothing method ES, autoregressive integrated moving average method ARIMA, support vector regression method SVR, feedforward neural network FNN and long short-term memory network LSTM, as shown in Table 1:

[0054] Table 1

[0055]

[0056] It can be seen from the results in the above table that since the present invention combines the spatial topological structure information and the inherent information of the time dimension, the prediction effect is more accurate.

[0057] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art to the present invention based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A dynamic fence planning method based on the check-in trajectory of a dockless shared bicycle, comprising the following steps: (1) Determine candidate sites by clustering based on historical travel data; (2) Calculate and determine the maximum inventory of the candidate site in each time window in the past; (3) Construct a spatiotemporal graph convolutional neural network. The network takes the maximum inventory of each candidate site in the previous m time windows as input to predict the maximum inventory of each candidate site in the next time window, where m is a natural number greater than 1. (4) Using the historical data obtained in step (2) to train the spatiotemporal graph convolutional neural network; (5) The trained spatiotemporal graph convolutional neural network is used to predict the maximum inventory of each candidate site in a new time window in the future, and dynamic fence planning is performed based on the initial inventory of each candidate site in this time window.

2. The dynamic fence planning method according to claim 1, characterized in that: The specific implementation method of step (1) is as follows: first, a 24-hour day is divided into multiple time periods. For any time period, the trajectory endpoints of all bicycle historical travel data in the time period are used to perform secondary clustering at the macro and micro levels, thereby dividing the entire city into multiple areas, and taking the centroid of each area as the cluster center; Then, the cluster centers obtained by clustering all time periods are obtained, and the center points that are too close are merged. That is, for two cluster centers A and B whose distance is less than the set threshold, the midpoint of the line connecting A and B is taken and added as the new cluster center, and A and B are deleted at the same time; finally, all cluster centers are matched to the road network, and the foot of the perpendicular between each cluster center and the nearest road line is taken as a candidate site, and the sites that are too close are merged according to the above method.

3. The dynamic fence planning method according to claim 1, characterized in that: The specific implementation method of step (2) is as follows: first, the travel data of all bicycles in a certain period of time in the past is obtained, and the period of time is divided into multiple time windows. According to the travel data, the bicycles in the first time window are counted and locked. The positions of these bicycles are spatially matched with each candidate station, so as to determine the initial inventory of each candidate station; For any subsequent time window, the number of bicycles that match the candidate site space within the time window is counted based on the travel data, and the maximum number of parked bicycles is used as the maximum inventory of the candidate site. At the same time, the final number of parked bicycles is recorded as the initial inventory of the next time window for the next iteration.

4. The dynamic fence planning method according to claim 1, characterized in that: Behavior recognition is performed on the bicycle travel data. By identifying that the previous state of the check-in point is locked and the next state is unlocked, the parking behavior between the two states is extracted from the original trajectory, and the distance between the arrival point and the starting point in the parking behavior is calculated. If the distance is less than the threshold, the parking behavior is determined to be a drifting behavior. If the distance is less than the threshold, the parking behavior is determined to be a scheduling behavior, and a virtual trip is created between the arrival point and the starting point of the parking behavior and inserted into the trip data.

5. The dynamic fence planning method according to claim 1, characterized in that: The spatiotemporal graph convolutional neural network is composed of a first gated time convolutional layer, a graph convolutional layer, a second gated time convolutional layer, and a fully connected layer connected in sequence, wherein the first gated time convolutional layer and the second gated time convolutional layer are used to convolve their respective inputs in the time dimension, and the graph convolutional layer uses a weighted adjacency matrix to convolve the input in the spatial dimension. The input dimensions of the first three layers are all m×n, and the fully connected layer outputs a 1×n prediction result after comprehensive mapping, where n is the number of candidate sites.

6. The dynamic fence planning method according to claim 1, characterized in that: The specific implementation method of step (4) is as follows: first, a graph structure is established with candidate sites as nodes, and any time window corresponds to a graph structure. The maximum inventory of the candidate site in the time window is the value of the corresponding node, and the distance between the candidate sites is the weight value of the edge between the corresponding nodes; then, historical data is used to calculate the weight of the edge in the time window t1~t m The graph structure is used as the input of the spatiotemporal graph convolutional neural network, with a time window t m+1 The graph structure is used as the label, and then the network parameters are calculated and updated by the gradient descent method according to the loss function between the network output prediction result and the label, thereby completing one iteration; based on the above, the next iteration is performed with a fixed step sliding time window, so that the network parameters are continuously updated and optimized.

7. The dynamic fence planning method according to claim 5, characterized in that: The dimension of the weighted adjacency matrix is ​​n×n, where the element value calculation expression of the i-th row and j-th column is as follows: Where: σ is the standard deviation, dist(i,j) is the semi-haversine distance between candidate sites i and j, i and j are the index numbers of the candidate sites, and k is the distance threshold.

8. The dynamic fence planning method according to claim 1, characterized in that: In the step (5), for any candidate site, the predicted maximum inventory is graded to determine the level of the maximum inventory of the candidate site, and it is judged whether the initial inventory of the candidate site in the new time window reaches the maximum inventory level: if so, the number of dynamic fences arranged in the candidate site area is increased; if not, the number of dynamic fences arranged in the candidate site area is reduced.

Images