Bus travel time prediction method based on space-time diagram attention network

Abstract

The invention discloses a bus travel time prediction method based on a space-time diagram attention network, and the method comprises the steps: carrying out the preprocessing of bus arrival time data, including bus number definition, duplicate record deletion, abnormal record deletion and missing record restoration; calculating the interval travel time of each train number based on the bus arrival time data to obtain the average travel time of each interval in different time periods; constructing a directed network by taking midpoints of the station intervals as nodes of the network and connecting lines between the midpoints of the station intervals as edges of the network, and fusing a distance matrix and a communication matrix of the nodes to represent spatial features between the nodes so as to weight the network of the station intervals; the historical average travel time of the station interval and the travel time of the latest train number serve as input, a transfomrer encoder, a graph attention network and a full connection layer are adopted to build a space-time graph attention network, time features and space features are extracted and fused, and multi-step prediction of the travel time of the station interval is achieved. According to the invention, more accurate bus travel time prediction can be realized.

Description

Technical Field

[0001] The invention relates to a bus travel time prediction method based on a spatiotemporal graph attention network, and belongs to the field of public transportation big data mining. Background Art

[0002] An efficient and convenient public transportation system is an important part of a sustainable city. Giving priority to the development of public transportation can ease urban traffic congestion and reduce carbon emissions, helping to achieve the "dual carbon" strategic goal. The operation of public transportation is subject to interference from a variety of internal and external factors, resulting in unstable vehicle operation time, reducing the efficiency of public transportation and increasing passengers' distrust of public transportation.

[0003] With the continuous improvement of the level of public transportation informatization and intelligence, buses are equipped with real-time positioning terminals, which provide important data support for real-time understanding of the operation status of buses. Carrying out bus travel time prediction based on real-time location information of buses will help improve the reliability of bus operations and provide passengers with better travel services, thereby attracting more passengers to choose public transportation.

[0004] Existing relevant prediction methods mainly consider the correlation of interval travel time historical series in the time dimension to carry out prediction. For example, "Bus arrival time prediction method and system based on GPS data and time-space correlation" (Application No.: 202410307360.X) discloses a bus arrival time prediction method that uses bus GPS data and integrates LSTM model, Seq2Seq structure and attention mechanism. "Bus travel time prediction method based on dual attention mechanism and bidirectional double-layer LSTM" (Application No.: 202210007787.9) discloses a bus travel time prediction method, which takes into account the influence of travel time in similar time periods of the previous day and similar time periods of the same characteristic day of the previous week.

[0005] In a traffic network, road conditions often show spatial correlation. Some prediction methods construct a network with bus stops as nodes, and predict the bus arrival time by considering the operating status of adjacent nodes. For example, "A Bus Arrival Time Prediction Method for Missing Data" (Application Number: 202111022557.1) discloses a bus arrival time prediction model based on a multi-head spatiotemporal graph attention network. However, this method has a cumbersome process when constructing the network. It is necessary to determine the weights of intersections and stations based on the number of GPS points and clustering algorithms to represent the geographical structure of each line; it is necessary to use external data sources such as POI data to construct an adjacency matrix, and the POI data source needs to be updated regularly; it uses the similarity between different time periods, but ignores the more fine-grained similarity between adjacent trains.

[0006] In summary, the existing technology does not fully explore the implicit knowledge in the real-time bus positioning data, and lacks effective utilization of the spatiotemporal correlation between adjacent station intervals and adjacent train running times. Summary of the invention

[0007] In view of the problems and improvement needs of the existing technology, the present invention proposes a bus travel time prediction method based on a spatiotemporal graph attention network, which can fully capture the spatiotemporal dependency of travel time between adjacent station areas and between adjacent trains in the bus network based on the historical data of bus arrival time, and realize more accurate bus travel time prediction.

[0008] The technical solution of the present invention is as follows: Step 1: Preprocessing of bus arrival time data; Step 2: Calculate the interval travel time based on the bus arrival time data; Step 3: Construction of directed weighted network between sites; Step 4: Build a bus travel time prediction model based on the spatiotemporal graph attention network.

[0009] Step 1: Preprocessing of bus arrival time data, including:

[0010] Step 1.1: The fields of bus arrival time data include line name, line direction, bus number, stop number, stop name, arrival distance, arrival timestamp, etc. Use the arrival distance field to filter and only keep the arrival time records of buses in the arrival state. Sort the bus arrival time data in ascending order according to the four fields of line name, line direction, bus number, and stop number. The process of each bus traveling from stop number 1 to the maximum number in the line direction is defined as a trip.

[0011] Step 1.2: If the number of arrival time records of a bus at a bus stop i in a trip is greater than or equal to 2, then only the latest arrival record of the bus at stop i is retained.

[0012] Step 1.3: Traverse the arrival time records of each train after sorting, and check whether the arrival time field is also sorted in ascending order. If the arrival time of a record is earlier than the arrival time of the previous record, the record is deleted as an abnormal record;

[0013] Step 1.4: For stations with missing arrival time records within a train, linear interpolation is used to fill the gap using the arrival times of the previous and next stations.

[0014] Step 2 calculates the interval travel time based on the bus arrival time data, including:

[0015] Step 2.1: Define the road section between any two adjacent bus stops as a stop interval. The starting point and end point of stop interval i are denoted as stop i and stop i+1 respectively. and The arrival time of train number k at the i-th station and the i+1-th station, respectively, and the travel time of train number j through the i-th station interval is equal to the difference in the arrival time of the bus at two adjacent bus stops, that is, Thus, the interval travel time series of each train is obtained.

[0016] Step 2.2: Take the average of the interval travel time of all lines and different buses passing through any station interval i in a certain period of time as the average travel time series of the station interval in this period of time. If there is no bus passing through a station interval in this period of time, the average travel time of the previous period of time is used to fill it.

[0017] Step 3: Constructing a directed weighted network between sites, including:

[0018] Step 3.1: Take the midpoints of the site intervals as the nodes of the network and the lines between the midpoints of the site intervals as the edges of the network to construct a directed network. The specific steps are as follows:

[0019] Since bus routes are usually divided into two directions, up and down, this network is designed as a directed network. The midpoint of the line connecting the stations at both ends of the station interval represents the station interval and is used as a node of the network. The longitude and latitude coordinates of the midpoint of the interval can be obtained by taking the arithmetic average of the longitude and latitude coordinates of the bus stations at both ends of the interval. i |i=1,2,…,N} represents a set of nodes, where v i represents the ith bus route interval in the network. The line connecting the midpoints of the station interval is used as the edge, and the set E = {e ij |i,j=1,2,…,N} represents the set of edges, where e ij represents the edge between the ith site interval and the jth site interval. Set W = {w ij |i,j=1,2,…,N} is the adjacency matrix of the network, where w ij for e ij The weight of .

[0020] Step 3.2: Calculate the Haversine distance between the midpoints of any two station intervals, and construct a distance matrix after taking the inverse. Construct a connectivity matrix based on the connectivity between any two station intervals, normalize the distance matrix and the connectivity matrix respectively, and then add them to get W. Use the normalized W to weight the edges of the network, specifically including:

[0021] S1. Construct a distance matrix using node v i and v j The latitude and longitude coordinates (x i ,y i ) and (x j ,y j ), and the Haversine formula is used to calculate the straight-line distance between two points. Since the closer the spatial distance between two stations, the more relevant the traffic operation status of the two is, the reciprocal of the distance is used to construct the distance matrix The larger the value, the greater the correlation between the two station intervals; conversely, it means that the correlation between the two intervals is smaller. p Perform normalization.

[0022] S2. Constructing connectivity matrix If two station intervals are connected in space, that is, the end station of interval i is the starting station of interval j, then the weight value of the corresponding edge is Set to 1; otherwise, set to 0.

[0023] S3. Add the distance matrix and the connectivity matrix, that is, W = W p +W a , and normalized to obtain the matrix W used to represent the spatial connection relationship between nodes.

[0024] Step 4 builds a bus travel time prediction model based on the spatiotemporal graph attention network, which includes:

[0025] Step 4.1: Concatenate the average travel time series in the first t time periods of each station interval with the travel time of the last n trains passing through each station interval in the first t time periods to form a spatiotemporal series As model input, where i is the number of nodes, t is the time step, and n is the number of neighboring cars selected.

[0026] Step 4.2: In the temporal feature extraction module, the Transformer encoder structure is used as the core component of temporal feature extraction. The Transformer encoder is composed of multiple stacked encoding layers, each of which contains a multi-head attention mechanism, a feedforward neural network, a normalization layer, and a residual connection. The multi-head self-attention mechanism is used to capture the mutual influence between different time periods in the sequence.

[0027] Step 4.3: In the spatial feature extraction module, the features output by the temporal feature extraction module are concatenated with the adjacency matrix. The Graph Attention Networks (GAT) are used to process the spatial information of the site interval directed weighted network. The multi-head attention mechanism is used to dynamically adjust the model's attention to each spatial position in the input sequence, and the features of each node are updated by weighted summing the features of neighboring nodes.

[0028] Step 4.4: In the spatiotemporal fusion module, after passing through the Transformer encoder and the GAT network, a joint representation of temporal and spatial features is obtained. These joint features are sent to a fully connected layer. The fully connected layer learns based on the extracted spatiotemporal features and generates a predicted value for the travel time of a certain train in the station interval i during the t+m period.

[0029] The mean square error (MSE) is used as the loss function, and the Adam optimizer and back-propagation algorithm are used to continuously adjust and optimize the hyperparameters of the model to minimize the MSE value between the predicted value and the true value, and to train the spatiotemporal graph attention network.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] Firstly, a lightweight spatiotemporal graph attention network that integrates the transformer encoder and the GAT network is proposed to extract the spatiotemporal correlation of travel time between adjacent station intervals. At the same time, the correlation of travel time between adjacent trains in each station interval is considered to achieve accurate prediction of bus travel time.

[0032] Secondly, a directed network is constructed with the midpoints of the station intervals as the nodes of the network and the lines between the midpoints of the station intervals as the edges of the network. When the adjacency matrix is ​​used to weight the network, the distance and connectivity between nodes are integrated, which can better reflect the spatial relationship between intervals.

[0033] Finally, only the bus arrival time data and the longitude and latitude coordinates of the bus stops are used as the data source of the model, without relying on other data sources, and the implicit knowledge in the bus arrival time data can be fully utilized. The entire prediction method has the characteristics of simple process and convenient operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 This is a schematic diagram of the construction of a directed weighted network between sites proposed by the present invention;

[0036] Figure 3 This is a framework diagram of the spatiotemporal graph attention network proposed in this invention. Specific implementation methods

[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0038] Reference Figure 1 The present invention provides a method for predicting bus travel time based on a spatiotemporal graph attention network, comprising the following steps: Step 1: Preprocessing of bus arrival time data; Step 2: Calculate the interval travel time based on the bus arrival time data; Step 3: Construction of directed weighted network between sites; Step 4: Build a bus travel time prediction model based on the spatiotemporal graph attention network.

[0039] The specific implementation is as follows:

[0040] Step 1: Preprocessing of bus arrival time data.

[0041] Step 1.1: The fields of bus arrival time data include line name, line direction, bus number, stop number, stop name, arrival distance, arrival timestamp and other fields. At present, all buses in Beijing are equipped with real-time positioning devices, and the bus arrival time data records the distance of each bus from the next stop at the current moment. When the arrival distance is -1, it means that the bus is at the station, and this moment is the bus arrival time. This embodiment selects the arrival time data of buses passing through the vicinity of Beijing University of Technology, covering 132 bus stop intervals, with a time range from 8:00 on August 30, 2020 to 8:00 on October 1, 2020, with a total of approximately 1 million records. An example of the original data obtained is shown in Table 1:

[0042] Table 1

[0043] Only records with a distance to the station value of -1 are retained. The bus arrival time data is sorted in ascending order according to the four fields: line name, line direction, vehicle number, and arrival time. The process of each bus traveling from station number 1 to the largest station number in the line direction is defined as a trip.

[0044] Step 1.2: Within a trip, if the number of arrival time records of a bus at a bus stop is greater than or equal to 2, only the last arrival record is retained.

[0045] Step 1.3, traverse the arrival time records of each train after sorting, and check whether the arrival time field is also sorted in ascending order. If the arrival time of a record is earlier than the arrival time of the previous record, the record is deleted as an abnormal record;

[0046] Step 1.4: For stations with missing arrival times within a train, fill in the gaps using the arrival times of the previous and next stations and the linear interpolation method.

[0047] Step 2, calculating the interval travel time based on the bus arrival time data, specifically including: Step 2.1: Define the road section between any two adjacent bus stops as a stop interval. The starting point and the end point of stop interval i are denoted as stop i and stop i+1 respectively. The interval travel time series of each bus trip can be obtained by calculating the difference in the arrival time of each bus passing through two adjacent bus stops. The calculation formula is shown in (1), thereby obtaining the interval travel time series of each bus. The interval name is represented by the name of the starting point and the end point, and the timestamp is the time when the vehicle arrives at the starting point. The calculation results are shown in Table 2.

[0048] In the formula, is the bus running time of the ith station interval with bus number k, and are the arrival times of the i-th station and the i+1-th station of train number k respectively.

[0049] Table 2

[0050] Step 2.2: Take the average of the interval travel time of all lines and different trains passing through any station interval within a certain period of time as the average travel time series of the station interval within the period of time. In this example, the time interval is 5 minutes. The calculation formula (2) is as follows:

[0051] Where K is the number of trains passing through station interval i in a certain period of time.

[0052] If there is no bus passing through a certain station interval during this period, the average travel time of the previous period is used to fill the gap. The final average travel time series is shown in Table 3.

[0053] Table 3

[0054] Step 3: Construct a directed weighted network of station intervals. Considering that bus routes are usually divided into two directions, up and down, when constructing a directed network, the bus routes are considered as directed graphs. The following are the specific steps:

[0055] Step 3.1: First, the interval between two adjacent bus stops is taken as a node in the network. To simplify the representation, the midpoint of each stop interval is selected as the representative of the interval and as the node of the network. The set V = {v i |i=1,2,…,N}, indicating that v i represents the i-th bus route interval in the network.

[0056] The connection between the midpoints of the station interval is used as the edge in the network, and the set is E = {e i |i=1,2,…,N} means, where e ij Represents the edge between the i-th site interval and the j-th site interval.

[0057] Finally, define an adjacency matrix W = {w ij |i,j=1,2,…,N}, where w ij for e ij According to the above steps, the network construction of this example is as follows Figure 2 shown.

[0058] Step 3.2: In order to more accurately describe the spatial connection relationship in the network, this example calculates the Haversine distance between the midpoints of any two site intervals, builds an adjacency matrix based on connectivity, and then combines the two for network weighting. The following are the specific steps:

[0059] S1. Construct a distance matrix:

[0060] Using node v i and v j The latitude and longitude coordinates v i (x i ,y i ), v j (x j ,y j ), use the Haversine formula to calculate the straight-line distance d between two points ij The calculation formula (3) is as follows:

[0061] Where r is the radius of the Earth.

[0062] Using distance d ij The reciprocal of Constructing the adjacency distance matrix Weight value The larger the value is, the greater the correlation between the two sites is; conversely, the smaller the correlation between the two sites is.

[0063] S2. Construct a connectivity matrix.

[0064] Based on the definition of connectivity, if two station intervals are connected, that is, the end station of interval i is the starting station of interval j, then the weight value of the corresponding edge is is set to 1; otherwise, it is set to 0. The corresponding formula (4) is:

[0065]

[0066] S3. Sum and normalize.

[0067] The distance matrix and the connectivity matrix are added together to obtain the adjacency matrix W. The adjacency matrix W combines the two factors of distance and connectivity, and comprehensively reflects the spatial connection relationship between nodes. The matrix W is normalized so that the values ​​of all elements are between 0 and 1.

[0068] Step 4: Build a bus travel time prediction model based on the spatiotemporal graph attention network. The framework of the spatiotemporal graph attention network is as follows: Figure 3 shown.

[0069] Step 4.1: Data preparation and input

[0070] In this example, the average travel time series of the first 12 periods of the 132 station intervals are spliced ​​with the travel time of the last train that passed through the first 12 periods of each interval to form the interval travel time series. as model input.

[0071] Step 4.2: Temporal feature extraction module.

[0072] The temporal feature extraction module uses the Transformer encoder as the core. The Transformer encoder structure consists of multiple stacked encoding layers. One encoding layer includes a multi-head attention mechanism, a feedforward neural network, a normalization layer, and a residual connection. The encoder input X is obtained by adding position encoding to the historical interval travel time series X input , self-attention can learn the dependencies between different time steps of interval travel time series. Self-attention coefficient in time dimension The calculation formula (5) is:

[0073] In the formula, Q = X input W q , K = X input W k 、V=X input Wv They are the input feature matrix X input The query matrix, key matrix and value matrix of d k is the dimension of the key; softmax is the normalization function.

[0074] The output of each head is obtained by weighted summation:

[0075]

[0076] Multi-head attention allows the model to learn different attention patterns in different subspaces by computing multiple self-attention heads in parallel. For each attention head, the input feature matrix X is calculated using different weight matrices. input Query key Sum For each attention head, calculate the attention output The output of each head is concatenated together and the feature representation of node i is obtained through a linear transformation:

[0077]

[0078] In order to maintain the stability of the gradient, residual connection and layer normalization are performed. The calculation formula (8) is as follows:

[0079]

[0080] Each self-attention layer is followed by a feed-forward neural network:

[0081]

[0082] Perform residual connection and layer normalization again:

[0083]

[0084] After stacking multiple layers of encoders, the feature vector representation of each network node can be obtained

[0085] Step 4.3: Spatial feature extraction module.

[0086] The spatial feature extraction module is based on the graph attention network, which includes a multi-head attention mechanism, a node feature update mechanism, a residual connection, and layer normalization. Concatenate with the adjacency matrix W as the input of GAT:

[0087]

[0088] GAT uses a multi-head attention mechanism to calculate the relevance between nodes. Each head learns the focus points of each spatial position in different input sequences. For the fth attention head, the attention coefficient between nodes is calculated.

[0089] Where a is a learnable weight vector; W f is a linear transformation matrix.

[0090] Normalize the node attention coefficient:

[0091]

[0092] N(i) is the set of neighbor nodes of node i.

[0093] GAT updates each node feature by weighted summing the features of neighboring nodes. The calculation formula (14) is:

[0094]

[0095] Concatenate the outputs of all heads to form a higher-dimensional feature representation:

[0096]

[0097] After residual connection and layer normalization:

[0098]

[0099] Step 4.4: Spatiotemporal fusion module and prediction output

[0100] A fully connected layer is used for feature fusion and multi-step prediction. The input of the fully connected layer is all the node features H output by GAT GAT , the output is the predicted value of each node in the next m time steps, that is, the predicted value of the travel time of a certain train at a specific station interval in m time steps

[0101] in

[0102] This example uses the mean square error (MSE) as the loss function to calculate the square average of the difference between the predicted value and the true value. The Adam optimizer and back propagation algorithm are used to continuously adjust and optimize the model's hyperparameters, including the number of encoding layers and the number of attention heads, to minimize the MSE value and train the spatiotemporal graph attention network. The calculation formula (18) of MSE is as follows:

[0103] in is the true value.

[0104] The spatiotemporal graph attention network proposed in this application is compared with the long short-term memory neural network (LSTM) and the gated recurrent unit (GRU) in an experiment. At the same time, the influence of the interval travel time of adjacent trains on the prediction accuracy of the spatiotemporal graph attention network is analyzed. The selected comparison indicators include the mean absolute error (MAE) and the root mean square error (RMSE). The results are shown in Table 4.

[0105] It can be found from Table 4 that compared with the LSTM and GRU models that only consider time features, the spatiotemporal graph attention network proposed in this application can significantly reduce the prediction error by integrating the spatial features of the station interval. By introducing the interval travel time of adjacent trains as one of the input features, the prediction error is also significantly reduced, proving the effectiveness of introducing the interval travel time of adjacent trains.

[0106] Table 4

Claims

1. A bus travel time prediction method based on spatiotemporal graph attention network, characterized in that: The following steps are involved: Step 1: Preprocessing of bus arrival time data; Step 2: Calculate the interval travel time based on the preprocessed bus arrival time data; Step 3: Construct a directed weighted network based on site intervals; Step 4: Build a bus travel time prediction model based on the spatiotemporal graph attention network.

2. The bus travel time prediction method based on spatiotemporal graph attention convolution network according to claim 1 is characterized in that: The bus arrival time data preprocessing step described in step 1 specifically includes: Step 1.1: The fields of bus arrival time data include line name, line direction, bus number, stop number, stop name, arrival distance, and arrival timestamp fields; use the arrival distance field to filter and only keep the arrival time records of buses in the arrival state; sort the bus arrival time data in ascending order according to the four fields of line name, line direction, bus number, and stop number; define the process of each bus traveling from stop number 1 to the maximum number in the line direction as a trip; Step 1.2: If the number of arrival time records of a bus at a bus stop i in a trip is greater than or equal to 2, then only the latest arrival time record of the bus at stop i is retained; Step 1.3: Traverse the arrival time records of each train after sorting, and check whether the arrival time field is also sorted in ascending order. If the arrival time of a record is earlier than the arrival time of the previous record, the record is deleted as an abnormal record; Step 1.4: For stations with missing arrival time records within a train, fill in the gaps using the arrival times of the previous and next stations and linear interpolation.

3. The bus travel time prediction method based on spatiotemporal graph attention network according to claim 1 is characterized in that: The calculation of the interval travel time based on the bus arrival time data described in step 2 specifically includes: Step 2.1: Define the road section between any two adjacent bus stops as a stop interval; the starting point and the end point of stop interval i are denoted as stop i and stop i+1 respectively; and The arrival time of train number k at the i-th station and the i+1-th station, respectively, and the travel time of train number k passing through the i-th station interval is equal to the difference in the arrival time of the bus at two adjacent bus stops, that is, Thus, the interval travel time series of each train is obtained; Step 2.2: Take the average of the interval travel time of all lines and different buses passing through any station interval within a certain period of time as the average travel time series of the station interval within the period of time; if no bus passes through a station interval within the period of time, the average travel time of the previous adjacent period of time is used to fill the gap.

4. The bus travel time prediction method based on spatiotemporal graph attention network according to claim 1 is characterized in that: The construction of the site interval directed weighted network described in step 3 specifically includes: Step 3.1: Construct a directed network with the midpoints of the station intervals as nodes of the network and the lines between the midpoints of the station intervals as edges of the network; Step 3.2: Calculate the Haversine distance between the midpoints of any two site intervals to construct a distance matrix. Based on the connectivity between any two site intervals, construct a connectivity matrix. Normalize the distance matrix and connectivity matrix separately, add them together, and normalize them again to obtain the adjacency matrix. Use the adjacency matrix to weight the network.

5. The bus travel time prediction method based on spatiotemporal graph attention network according to claim 1 is characterized in that: The steps of constructing a bus travel time prediction model based on the spatiotemporal graph attention network described in step 4 specifically include: Step 4.1: Concatenate the average travel time series in the first t time periods of each station interval with the travel time of the last n trips passing through each station interval in the first t time periods to form a spatiotemporal series as the input of the bus travel time prediction model; Step 4.2: In the temporal feature extraction module, the Transformer encoder is selected as the temporal feature extraction module. The Transformer encoder structure consists of multiple stacked encoding layers, each of which contains a multi-head attention mechanism, a feedforward neural network, a normalization layer, and a residual connection; the multi-head self-attention mechanism is used to capture the dependencies between different time periods in the sequence; Step 4.3: In the spatial feature extraction module, the features output by the temporal feature extraction module are concatenated with the adjacency matrix, and the graph attention network GAT is used to process the spatial information of the site interval directed weighted network. The multi-head attention is used to dynamically adjust the model's attention to different spatial positions in the input sequence, and the features of each node are updated by weighted summing the features of neighboring nodes. Step 4.4: Sequentially connect the temporal feature extraction module and the spatial feature extraction module, and finally pass through a fully connected layer to obtain the predicted value of the travel time of a certain train in the station interval i during the t+m period; use the mean error MSE as the loss function, use the Adam optimizer and back-propagation algorithm to continuously adjust and optimize the model's hyperparameters, and train the spatiotemporal graph attention network.

Images