A method for modeling heterogeneous travel mode of online taxi demand prediction

Abstract

The application provides a method for modeling heterogeneous travel mode of online taxi demand prediction, and belongs to the field of traffic flow prediction. The method establishes a prediction model HTPSTGCN for online taxi demand. Historical order data is converted into grid online taxi demand through data cleaning and reconstruction, and is input into the prediction model. The model aggregates grids into homogeneous regions through a homogeneous region generator, outputs historical online taxi demand of the region, extracts feature vectors representing the heterogeneous inherent influence of different regions through a heterogeneous inherent influence extractor, extracts two groups of feature vectors representing the influence of similar regions and high interaction regions through an external influence extractor, inputs the two groups of feature vectors into a fusion predictor, fuses the two groups of feature vectors and maps them to the demand space, and outputs the demand prediction value. The test proves that the application has good prediction effect on high demand area and low demand area, has robustness, and can meet the demand of online taxi demand prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of traffic flow prediction, and relates to network taxi demand prediction, in particular to a network taxi demand prediction method based on heterogeneous travel modes. BACKGROUND

[0002] As a new type of transportation infrastructure, the network car platform is an important part of the intelligent transportation system. The network car platform has a large number of active users, and billions of orders are generated on it every year. An efficient network car platform can improve the quality of life of residents and the efficiency of road traffic operation, so how to improve the efficiency of the network car platform is a key problem in the construction of an intelligent transportation system. The network car platform has problems such as long-time empty loading of vehicles and long waiting time of passengers, which is caused by the imbalance between travel demand and vehicle supply, and accurate network car demand prediction is the key to solving the problem.

[0003] The demand of online car-hailing is influenced by complex spatio-temporal factors, which can be attributed to intrinsic factors and external factors. Intrinsic factors remain stable for a long period of time, while external factors change dramatically over time. For example, the function and population of a region are intrinsic factors, while the inter-regional interaction traffic flow and the traffic flow of neighboring regions are external factors. These differentiated intrinsic factors and complex external factors make different regions have heterogeneous travel patterns. Generally speaking, there are three types of online car-hailing demand prediction methods: statistical methods, machine learning-based methods, and deep learning-based methods. Earlier research on online car-hailing demand prediction is based on statistical methods, such as autoregressive moving average models, which can only capture the influence of simple intrinsic factors, and are rarely applied to real-world situations because the model can only handle stationary spatio-temporal sequences. Machine learning-based methods can handle non-stationary spatio-temporal sequences, but they only focus on the intrinsic factors of a single region and ignore external factors, which cannot accurately model the heterogeneous travel patterns of different regions. With the increase of computing resources and data, deep learning-based methods dominate the research. Deep learning models capture the influence of intrinsic factors and external factors by building deep networks, among which recurrent neural networks and convolutional neural networks have achieved good results in online car-hailing demand prediction. However, these models ignore the differences in intrinsic factors between regions, use the same parameters for modeling, and the convolutional neural network-based model can only handle data in Euclidean space, while traffic networks are usually represented as graph structures in non-Euclidean space. These models divide the region to be studied into coarse-grained, regular grids, which often cannot guarantee that the regions within the grid have similar travel patterns. Therefore, recent research focuses on graph convolutional neural networks that can handle non-Euclidean space data. Existing graph convolutional network-based methods do not explicitly model the differentiated intrinsic influence of regions and ignore the influence of external factors such as interaction flow.

[0004] To solve the above problems, there is currently a need for a new online car-hailing demand prediction method based on graph convolutional neural networks. SUMMARY

[0005] The problem to be solved by the present application is that in the online car-hailing demand prediction problem, statistical and machine learning-based methods ignore the influence of external factors of regions, and existing deep learning models do not explicitly model differentiated intrinsic factors of regions, so both cannot accurately model heterogeneous travel patterns and cannot give accurate prediction results. Therefore, a new online car-hailing demand prediction method is needed to reduce prediction errors.

[0006] To solve the above problems, the present application provides an online car-hailing demand prediction method for modeling heterogeneous travel patterns, comprising the following steps:

[0007] Step 1. Collect order data from the online car-hailing platform, perform data cleaning, construct a standard order data set, divide the area to be studied into grids according to longitude and latitude, convert the order data set into a demand data set, and calculate the online car-hailing demand and arrival data of each grid in different time periods.

[0008] Step 2. Construct a deep learning model HTPSTGCN for modeling heterogeneous travel patterns and use it for demand prediction.

[0009] The deep learning model HTPSTGCN includes a homogeneous region generator, a heterogeneous inherent influence extractor, an external influence extractor, and a fusion predictor. The input grid historical time period demand is aggregated into homogeneous regions in the homogeneous region generator, and after region generation, the region historical time period demand is calculated, the feature vector representing the inherent influence of regional differences is obtained by the heterogeneous inherent influence extractor, and two feature vectors representing the influence of similar regions and high interaction regions are obtained by the external influence extractor. The two feature vectors output by the external influence extractor are input into the fusion predictor and mapped to the demand space to output the online car-hailing demand prediction value.

[0010] Step 3. Train the established HTPSTGCN model, input the data to be predicted into the trained model, and output the future online car-hailing demand of each region.

[0011] In step 2, the homogeneous region generator includes a grid OD (traffic volume) distribution generation part and a grid aggregation part; the OD distribution generation part calculates the grid time period online car-hailing pickup and drop-off flow to construct an OD distribution vector; the grid aggregation part calculates the grid dissimilarity using the grid OD distribution vector and the distance between grids, and then clusters the grids by grid dissimilarity to aggregate them into homogeneous regions.

[0012] In step 2, the heterogeneous inherent influence extractor uses an independent gated recurrent unit for each homogeneous region to extract a time series feature vector representing the inherent influence of regional differences and input it into the external influence extractor.

[0013] The external influence extractor in step 2 includes a nearby region influence extractor and a high interaction region influence extractor. The nearby region influence extractor constructs a distance geographical graph through region latitude and longitude information, captures the external influence of the nearby region on the current region by using a graph convolutional neural network, and extracts a feature vector representing the external influence of the nearby region on the current region. The high interaction region influence extractor constructs an interaction graph through inter-region interaction order data, captures the external influence of the region with high interaction order quantity on the current region by using a graph convolutional neural network, and extracts a feature vector representing the external influence of the region with high interaction order quantity on the current region. The external influence extractor inputs the two extracted feature vectors into the fusion predictor.

[0014] In step 2, the fusion predictor is composed of a feature fusion part and a feature mapping part. The feature fusion part weights and fuses the two feature vectors input by the external influence extractor through adaptive weights. The feature mapping part maps the fused feature vector to the demand quantity space through a stacked fully connected layer, and outputs a prediction value of the online car-hailing demand quantity. The feature mapping part uses different stacked fully connected layers for mapping for different regions.

[0015] Compared with the prior art, the method has the advantages and positive effects that: first, the method clusters fine-grained grids into regions with similar travel patterns based on distance and OD distribution, so that the neural network can easily capture the rules. Second, the method uses independent neural network structures to model the different local inherent factors of different regions, effectively distinguishing the travel patterns of different regions. Then, the method combines region proximity information and inter-region interaction information, uses a model based on a graph convolutional neural network to capture the external influence of the loaded region, and supplements information for the region online car-hailing demand prediction task. It has been verified that the method is suitable for regions with different traffic levels, can be widely used in region-level online car-hailing demand prediction tasks, has robustness, and the prediction result is accurate. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the deep learning saliency detection method based on global priori and local context of the application;

[0017] Figure 2 is a schematic diagram of region division and reconstruction and a grid OD distribution in the application;

[0018] Figure 3 is the homogeneous region, distance geographical graph and interaction graph generated in the embodiment of the application; (a) homogeneous region, (b) distance geographical graph, (c) interaction graph;

[0019] Figure 4Fig. 1 is a schematic diagram of the prediction results of the present application on the New York taxi dataset; (a) high traffic area 1, (b) high traffic area 2, (c) low traffic area 1, (d) low traffic area 2. DETAILED DESCRIPTION

[0020] The technical solutions of the present application will be further described in detail below in combination with the drawings and examples.

[0021] The present application provides a method for predicting the demand of online car-hailing, which converts historical online car-hailing order data into grid online car-hailing demand data through data cleaning and reconstruction, and predicts future online car-hailing demand based on historical online car-hailing demand. First, the area to be studied is divided into fine-grained grids, then homogenized areas are constructed based on geographical proximity and OD distribution similarity, and demand information is aggregated for different areas. For different areas with heterogeneous travel patterns, a separate deep learning model is used to model the differential inherent influences on regional demand, obtaining a time-dimension feature vector. Then, a distance geography graph and an interaction graph are constructed using the distance between regions and the interaction flow, and the distance geography graph and the interaction graph are used to model complex external influences, further obtaining a space-dimension feature vector. By fusing the space-dimension feature vector, the prediction result is obtained. The present application is suitable for online car-hailing demand prediction, has good robustness, and the detection result is accurate. An implementation process of the method of the present application is shown in Fig. 1, and the specific implementation steps of the present application will be described in detail below in combination with the drawings. Figure 1

[0022] Step 1. The order data collected by the online car-hailing platform is cleaned to construct a standard order data set, the order data set is converted into a demand data set, and the demand data of each grid at different time periods is obtained as the input of the model.

[0023] The original order data includes the following eight kinds of information: order number, departure longitude, departure latitude, arrival longitude, arrival latitude, departure time, arrival time, and travel time. In the data cleaning process, the order data with travel time less than 0 or travel time greater than 15 hours is removed. The area to be studied is divided into I×J fine-grained grids according to the longitude and latitude. According to the departure longitude, departure latitude, and departure time information in the order, the online car-hailing demand in the corresponding grid within the same time interval is counted, and the demand of grid g(i, j) in the kth time period is denoted as If the demand of a certain grid is always 0, the grid will not be considered in the subsequent model. According to the arrival longitude, arrival latitude, and arrival time information in the order, the online car-hailing arrival quantity in the corresponding grid within the same time interval is counted, which is used as the input of the deep learning model HTPSTGCN for modeling heterogeneous travel patterns.

[0024] ​Step 2. Construct a deep learning model HTPSTGCN that models heterogeneous travel patterns and use it for demand prediction.

[0025] The deep learning model HTPSTGCN includes a homogeneous region generator, a heterogeneous intrinsic influence extractor, an external influence extractor, and a fusion predictor. The input raster historical time-segmented demand is aggregated into homogeneous regions by the homogeneous region generator, and then aggregated into regional historical time-segmented demand. The heterogeneous intrinsic influence extractor then obtains a sequence of feature vectors representing the inherent influence of regional differences. The external influence extractor then obtains two sets of feature vector sequences representing the influence of adjacent regions and highly interactive regions, respectively. These two feature vector sequences output by the external influence extractor are input into the fusion predictor, mapped onto the demand space, and output as a predicted value for ride-hailing demand.

[0026] Step 2.1, design and implement a homogeneous region generator.

[0027] Raster partitioning and homogeneous region construction, such as Figure 2 As shown, a homogeneous region is a region within which the constituent parts have similar travel patterns. The innovation of this invention lies in the fact that other methods, by dividing regions into coarse-grained grids or by road networks, cannot guarantee similar travel patterns within the regions, and therefore cannot model the travel patterns within the regions. In contrast, this invention aggregates grids with similar travel patterns to obtain homogeneous regions, thereby enabling downstream models to better model consistent travel patterns within the regions.

[0028] This invention proposes that grids within the same area should possess the following two characteristics:

[0029] Geographic proximity: If the grids in the aggregated region are scattered, then the prediction for that region is meaningless, and ride-hailing platforms cannot dispatch vehicles based on the prediction results for that region. Therefore, the grids in the region should have geographical proximity.

[0030] Travel pattern similarity: Grids within the same area should have similar travel patterns, because the patterns of grids with similar travel patterns will become the travel patterns of the area after aggregation. However, if the travel patterns of grids within an area are very different, the aggregated area may not have travel patterns that can be modeled.

[0031] Based on these two characteristics, the geographic distance is first calculated using raster latitude and longitude. The geographic distance between raster g(i,j) and g(p,q) is defined as follows: Next, an OD distribution is defined for the grid, which represents the proportion of vehicle boarding and alighting traffic at different times for that grid. In this embodiment of the invention, a week is divided into five time periods, as shown in Table 1.

[0032] Table 1. Time Period Division Method

[0033]

[0034]

[0035] Traffic flow was counted separately for different time periods. i (1≤i≤5) and the alighting flow d i (1≤i≤5), where i takes values ​​from 1 to 5, corresponding to the 5 time periods in Table 1. For each grid cell, weekdays and weekends are counted separately, and the passenger flow for each time period of each day is counted to obtain the OD distribution M of a grid cell as follows:

[0036]

[0037] The OD dissimilarity between rasters is calculated based on the OD distribution. The OD dissimilarity between rasters g(i,j) and g(p,q) is:

[0038]

[0039] Where M i，j and M p，q Let g(i,j) and g(p,q) represent the OD distribution vectors of grids g(i,j) and g(p,q), respectively. Then, the global dissimilarity is calculated by weighted summation of OD dissimilarity and geographic distance, expressed as:

[0040]

[0041] ρ1 is an adjustable hyperparameter used to adjust the weight of the two dissimilarity levels.

[0042] Calculate the global dissimilarity among all rasters and construct a global dissimilarity matrix. Based on this matrix, use K-means clustering to group the rasters into N clusters, which are called N regions. It's important to note that rasters with historically low traffic should be removed before clustering. For example, a minimum traffic threshold can be set, and rasters that do not meet this threshold should be deleted to enhance clustering stability. An example of homogeneous regions generated by clustering is shown below. Figure 3 As shown in (a). Finally, the grid demand of the region is generated by summing the grid demand of the region, which is expressed as region P in the following formula. m Demand at time k

[0043]

[0044] in, Representing region P m The demand for ride-hailing services in the middle grid g(i,j) at time k.

[0045] Step 2.2, design and implement the heterogeneous inherent influence extractor.

[0046] This invention uses independent neural network structures to model the heterogeneous intrinsic effects of each region. The heterogeneous intrinsic effect extractor is a stack of multiple gated linear units, which model the differential intrinsic effects of different regions through multiple independent gated linear units. The heterogeneous intrinsic effect extractor in this embodiment uses two layers of gated recurrent units. Gated recurrent units are a variant of recurrent neural networks, which solve the gradient vanishing and gradient exploding problems during training of recurrent neural networks through reset and update gates. Their recurrent structure is suitable for modeling time series features. For the historical demand of each region... The output vector of the hidden layer at the current time can be obtained based on the following formula. as follows:

[0047]

[0048]

[0049]

[0050]

[0051] in, W and U are trainable model parameters, ⊙ represents element-wise matrix multiplication, and σ(·) represents the sigmoid function. To update the door, To reset the door. In an embodiment of the invention, The dimension of W is 64*12. The dimension of U is 64*64. The hidden layer vector output at the last time step is used as the temporal feature vector output by the network. The feature vectors extracted from each region are concatenated to obtain a temporal feature matrix with a dimension of 54*64, denoted as H. Here, 54 is the total number of clustered regions in step 2.1, and 64 is the feature vector dimension, which is consistent with the output dimension of the hidden layer of the gated recurrent unit. The resulting temporal feature matrix is ​​used as the input of the external influence extractor.

[0052] Step 2.3, design and implement the external influence extractor.

[0053] The external influence extractor is used to capture the complex external influences affecting the current region. It consists of two parts: a proximity region influence extractor and a high-interaction region influence extractor. Based on a graph convolutional neural network (GCN), GCNs are better suited for modeling data in non-Euclidean space. The key lies in constructing a suitable graph structure, which is generally expressed in the form of an adjacency matrix. This invention uses two different topological structures to encode proximity region influences and high-interaction region influences.

[0054] Distance-based geographic map: The distance-based geographic map is constructed based on the distances between regions. Assuming that regions with close geographical distances are more likely to influence each other, the adjacency matrix of the distance-based geographic map in this invention is calculated as follows:

[0055]

[0056] in, Represents the adjacency matrix A geo The value at position [i, j]; σ is the filter width of the Gaussian kernel function, a hyperparameter, which is set to 2 in this embodiment; dis ij For region P i and P j The center-line distance, threshold geo A distance threshold is set to consider areas above the threshold as having no mutual influence. In this embodiment of the invention, the distance threshold is set to 2 kilometers. An example of a distance-based geographic map constructed by this invention is shown below. Figure 3 As shown in (b).

[0057] Interaction Graph: The interaction graph is constructed based on the interaction traffic between regions. Interaction traffic refers to interaction order data. This invention assumes that two regions with higher interaction traffic are more likely to have mutual influence. The OD traffic matrix between regions is defined as follows:

[0058]

[0059] Among them, od ij This indicates that historical data from region P i To P j The sum of traffic. The inter-regional traffic matrix is ​​defined as:

[0060]

[0061] Here, the superscript T represents the matrix transpose operation. An example of an interaction graph constructed in this invention is shown below. Figure 3 As shown in (c).

[0062] Finally, the adjacency matrix of the interaction graph is defined as:

[0063]

[0064] Where D is The degree matrix.

[0065] Next, based on the two graph structures mentioned above, parallel graph convolutional neural networks were constructed, corresponding to the proximity region influence extractor and the high interaction region influence extractor, respectively. The output of a single-layer graph neural network is shown in the following equation:

[0066]

[0067]

[0068] Among them, H l H l+1 represent the hidden vectors of layers l and l+1, respectively, with a dimension of 54*64 in this embodiment. When l=0, it is the same as the output vector of the heterogeneous intrinsic influence extractor. || represents the matrix concatenation operation, referred to here as a skip connection. This design can prevent over-smoothing during graph convolution and avoid gradient vanishing and gradient exploding during training. W int and W geo These are trainable parameters; in this embodiment of the invention, their dimension is 128*64, and the final output feature vector dimension is 54*64. The subscript 'int' represents the interaction graph, and 'geo' represents the distance-based geographic graph.

[0069] Step 2.4, design the fusion predictor.

[0070] The fusion predictor is divided into a feature fusion part and a feature mapping part. In the feature fusion part, it receives the output feature vector from the external influence extractor and performs feature fusion using adaptive weights to obtain the fused feature H. fuse as follows:

[0071]

[0072] Here, 0≤ρ2≤1 are trainable weights used to adjust the importance of the two external influences.

[0073] In the feature mapping part, similar to the heterogeneous intrinsic influence extractor, this invention uses an independent fully connected layer for each region to map the extracted fused features onto the demand space, region P. i The calculation process for future demand is as follows:

[0074]

[0075] in and These are trainable parameters, and in this embodiment of the invention, their dimensions are 1*100 and 100*64, respectively. For the final predicted region P i The demand for ride-hailing services.

[0076] Step 3: Train the HTPSTGCN model established in Step 2. Input the data to be predicted into the trained model and output the future demand for ride-hailing services in each region so that the system platform can schedule vehicles.

[0077] This invention uses the New York taxi order dataset to train the HTPSTGCN model. The New York taxi order dataset is a large-scale public dataset containing taxi order data from Manhattan and surrounding areas. The dataset used in this invention extracts order data from April 1, 2016 to June 30, 2016, and the study area is a rectangular region from 40.67°N to 40.80°N and 73.92°W to 74.02°W.

[0078] (1) For the New York taxi dataset, the rectangular area to be studied is divided into fine-grained 130*100 grids. The dataset contains approximately 30 million taxi order data, each order data includes the following eight pieces of information: order number, departure longitude, departure latitude, arrival longitude and arrival latitude, departure time, arrival time, and trip time. The dataset is cleaned and transformed into a grid taxi demand dataset according to step 1, with time periods of 20 minutes. After cleaning, the total number of grids is 6397, and the dataset dimension is 6397*6552, representing the taxi demand of each grid in each time period. The regional taxi demand dataset is obtained through the homogeneous region generator in step 2.1. The regional generation result is as follows. Figure 4 As shown in (a), a total of 54 regions were generated, with a dataset dimension of 54*6552, representing the taxi demand in each region at each time period. The dataset was divided into a training set (80%), a validation set (10%), and a test set (10%) along the time dimension. This invention treats the demand prediction task as a spatiotemporal sequence prediction task, using historical 4-hour regional demand data to predict regional demand in the next 20 minutes.

[0079] (2) The HTPSTGCN model of the present invention is trained. The data is Z-score standardized, and all parameters in the HTPSTGCN model are randomly initialized.

[0080] Training was performed on a Linux operating system using an Intel(R) Core(TM) i9-10980XE CPU and a GeForce RTX3090 GPU. The batch processing parameter was set to 72, the initial learning rate was set to 0.001, and the loss function was the L2 loss function, as shown below:

[0081]

[0082] in X and X represent the predicted value and the actual value, respectively.

[0083] Using the Adam optimization algorithm and an exponentially decaying dynamic learning rate strategy, the model was trained for 200 epochs on the complete dataset. The model was validated using L2 loss at each training epoch, and the optimal model was saved based on the L2 loss. An Early Stopping strategy was used during training, terminating training early if the L2 loss did not decrease for 50 consecutive epochs.

[0084] (3) The test results of the HTPSTGCN model on the test set are as follows: Figure 4 As shown, 4(a) and 4(b) are areas of high demand. Figure 4 (c) and 4(d) are low-demand areas. The results show that the model HTPSTGCN of this invention has high predictive performance for both high-demand and low-demand areas and is robust.

[0085] The method of this invention was compared with existing technologies, and regional taxi demand was predicted on the same dataset. The comparison results are shown in Table 2. The prediction results were evaluated using mean absolute error (MAE) and root mean square error (RMSE). The lower the error, the better the prediction effect. The calculation method is as follows:

[0086]

[0087]

[0088] This invention compares six existing technical solutions.

[0089] The first type is the classic time series forecasting model ARIMA, which captures the correlation of time series and removes the influence of noise through autoregression and moving average strategies. However, it does not consider spatial correlation, so its performance is poor.

[0090] The second and third types are two major variants of recurrent neural networks, GRU and LSTM, which are widely used in time series prediction problems. They capture time dependencies through recurrent structures, but ignore the inherent effects of heterogeneity between regions and spatial correlations, thus resulting in poor prediction performance.

[0091] The fourth type is TGCN, which captures both spatial correlation and temporal dependence by combining GRU and GCN, but ignores the inherent impact of heterogeneity between regions. Its prediction performance is stronger than the first three models, but it still has shortcomings.

[0092] The fifth type is GraphWaveNet, which further improves prediction performance through a dynamic graph structure, but does not consider the external influences caused by highly interactive regions in the current area.

[0093] The sixth type is STGCN, which uses one-dimensional convolution to model temporal features and also ignores the external influences caused by highly interactive regions in the current region.

[0094] Table 2 Comparison of Regional Demand Forecast Results

[0095]

[0096]

[0097] The experimental results show that the prediction performance of this model is better than that of existing models.

[0098] (4) The present invention conducted ablation experiments on several innovative points to prove its effectiveness. The experimental results are shown in Table 3.

[0099] This invention compares three post-ablation models.

[0100] The first method removes the heterogeneous intrinsic effects extractor. This part in the original model is replaced with a separate two-layer gated recurrent unit, preventing it from extracting heterogeneous intrinsic effects.

[0101] The second approach removes the influence extractor for adjacent regions. This part is directly removed from the original model, and only the output of the influence extractor for high-interaction regions is retained as the output of the external influence extractor.

[0102] The third approach removes the high-interaction region influence extractor. This part is directly removed from the original model, and only the output of the proximity region influence extractor is retained as the output of the external influence extractor.

[0103] Table 3 Comparison of ablation test results

[0104] Comparative approach MAE RMSE Heterogeneous-inherent-removal extractor 8.02 14.42 High-interaction-area-removal extractor 8.33 14.85 Nearby-area-removal extractor 8.10 14.13 Inventive HTP STGCN 7.88 14.08

[0105] The experimental results show that each module in the model of this invention is effective, and the high-interaction region influence extractor contributes the most to the model effect.

[0106] Except for the technical features described in the specification, all other technologies are known to those skilled in the art. Descriptions of well-known components and technologies are omitted in this invention to avoid redundancy and unnecessary limitation. The embodiments described above do not represent all embodiments consistent with this application. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the protection scope of this invention.

Claims

1. A method for modeling heterogeneous trip patterns for demand forecasting of ride-hailing, characterized in that, The method comprises the following steps: Step 1, collecting order data of a network car-hailing platform, constructing a standard order data set, dividing a to-be-studied area into grids according to longitude and latitude, and according to the order data set, counting network car-hailing demand and arrival amount of each grid in each time period; Step 2, constructing a deep learning model HTPSTGCN of heterogeneous travel modes for predicting network car-hailing demand; The model HTPSTGCN comprises a homogeneous region generator, a heterogeneous inherent influence extractor, an external influence extractor and a fusion predictor; The homogeneous region generator comprises a grid OD distribution generation part and a grid aggregation part; OD represents traffic volume; the grid historical network car-hailing demand in each time period is input into the homogeneous region generator, the OD distribution generation part counts the network car-hailing boarding and alighting flow of each grid in each time period, and constructs a grid OD distribution vector; the grid aggregation part calculates the dissimilarity of the grid by using the grid OD distribution vector and the distance between grids, and then clusters and aggregates the grids into homogeneous regions according to the dissimilarity; the network car-hailing demand in each time period in the homogeneous region is input into the heterogeneous inherent influence extractor; The heterogeneous inherent influence extractor uses an independent gated recurrent unit for each homogeneous region to extract a time sequence feature vector representing the inherent influence of the difference between regions, and inputs the time sequence feature vector into the external influence extractor; The external influence extractor comprises a similar region influence extractor and a high interaction region influence extractor; the similar region influence extractor constructs a distance geographical map by using the longitude and latitude of the region, and uses a graph convolutional neural network to extract a feature vector representing the external influence of the region similar to the current region; the high interaction region influence extractor constructs an interaction graph by using the interaction flow between regions, and uses a graph convolutional neural network to extract a feature vector representing the external influence of the region with high interaction flow; the external influence extractor inputs the two extracted feature vectors into the fusion predictor; The fusion predictor comprises a feature fusion part and a feature mapping part; The feature fusion part weights and fuses the two input feature vectors by using adaptive weights; the feature mapping part maps the fusion feature of each homogeneous region to the network car-hailing demand space by using a stacked fully connected layer, and outputs a network car-hailing demand prediction value; Step 3, training the established model HTPSTGCN, and predicting the future network car-hailing demand of each region by using the trained model.

2. The method of claim 1, wherein, The step 1 comprises: Constructing a standard order data set, wherein each order data in the standard order data set comprises an order number, a departure longitude, a departure latitude, an arrival longitude, an arrival latitude, a departure time, an arrival time and a travel time; cleaning the order data with a travel time less than 0 or a travel time greater than 15 hours; According to the departure longitude, the departure latitude and the departure time in the order, counting the network car-hailing demand of each grid in each time period; if the demand of a certain grid is always 0, the grid is no longer considered; According to the arrival longitude, the arrival latitude and the arrival time in the order, counting the network car-hailing arrival amount of each grid in each time period.

3. The method of claim 1, wherein, In step 2, the homogeneous region generator calculates the dissimilarity between grids by using the grid OD distribution vector and the inter-grid distance, and then clusters the grids by using the K-means clustering algorithm; Wherein, the grid OD distribution is the statistics of the boarding and alighting flow ratio of the grid in different time periods; one week is divided into two cases of weekdays and weekends, and the time periods are divided in the weekdays and weekends respectively; The distance between the two grids g(i, j) and g(p, q) is calculated as The OD dissimilarity is The dissimilarity D of the two grids is then obtained by weighted summation i,j,p,q .

4. The method of claim 3, wherein, In step 2, before clustering the grids, the grids that do not satisfy the minimum flow threshold are deleted.

5. The method of claim 1, wherein, In step 2, the heterogeneous inherent influence extractor includes two layers of gated recurrent units corresponding to the number of homogeneous regions, and extracts the time sequence feature vectors of each region respectively.

6. The method of claim 1, wherein, In step 2, the proximity region influence extractor constructs a distance geographical graph based on the distance between regions, the adjacency matrix A of the distance geographical graph geo This is calculated as follows: wherein, denotes the adjacency matrix A geo at the [i,j] position; σ is the filter width of the Gaussian kernel function; dis ij is the center straight line distance between the regions P i and P j ; threshold geo is a preset distance threshold. The high-interaction region influence extractor constructs an interaction graph based on the interaction flow between regions, and then performs an analysis on two regions P. i and P j Statistical analysis from historical data of region P i To P j Total flow of od ij , will od ij As an inter-regional OD flow matrix W od The value at position [i,j] is used to further obtain the interaction flow matrix between regions. The superscript T represents matrix transpose. This is used to calculate the adjacency matrix of the interaction graph. D is The degree matrix; A parallel graph convolutional neural network is constructed, which corresponds to the similar region influence extractor and the high interaction region influence extractor respectively, and the output of the single-layer graph neural network is as follows: wherein, respectively represent the hidden vectors of the l-th and l+1-th layers of the neural network of the region-wise influence extractor, respectively represent the hidden vectors of the l-th and l+1-th layers of the neural network of the high-interaction region-wise influence extractor, W int and W geo are trainable parameters, and || represents a matrix concatenation operation, which is a skip connection in this case.

7. The method of claim 1, wherein, In step 2, in the feature fusion part of the fusion predictor, adaptive weights are used for feature fusion, which is represented as follows: wherein p2 is a trainable weight, 0 < p2 < 1; H fuse is the fused feature, geo is the feature vector extracted by the nearby region influence extractor, H int is the feature vector extracted by the high interaction region influence extractor; The fused features are input into two full connection layers corresponding to the homogeneous regions, and the predicted value of the online car-hailing demand is mapped and output. in, and For trainable parameters, For region P i The fusion characteristics For the predicted region P i The demand for ride-hailing services.

Images