Short-time rail transit flow prediction method and device based on decoupling process

Abstract

The invention discloses a short-time rail transit flow prediction method and device based on a decoupling process. The method belongs to the technical field of public transport data analysis and prediction, and comprises the following operation steps: collecting subway AFC card punching data to obtain a passenger OD data set; defining a region to establish a rail transit reality topological structure, and designing a multi-modal convolution feature map through entity subway road measurement data and an OD data set; the method comprises the following steps of: improving an initial Cell structure of an LSTM (Long Short Term Memory), fusing a Swin Transform module into the Cell, and adjusting an original gating structure; and constructing a prediction network by using the IST-LSTM as a new prediction unit and the SBULSTM in parallel. According to the method, AFC clock-in data is collected, a new multi-modal convolution feature map is constructed by using a dynamic time warping method, and a new multi-modal convolution feature map is established by simulating a signal decoupling process; meanwhile, an LSTM unit is improved, and the attention function of Swin Transform is integrated, so that all time features in a time period can be captured, global decoupling information provided by the multi-modal convolution feature map can be enhanced, and the prediction accuracy of the rail traffic flow is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of public transportation data analysis and prediction, and relates to a short-term rail transit flow prediction method and device based on a decoupling process. Background Art

[0002] As an important means of daily commuting for modern urban residents, subway travel has the characteristics of high speed, high efficiency, large passenger volume, and good punctuality compared with driving. However, the surge in passenger flow and the complexity of routes have brought potential safety hazards to the subway system. Therefore, how to build a more accurate passenger flow prediction model to accurately predict the future short-term passenger flow of the subway has important practical significance and application prospects for optimizing traffic resource allocation and reducing the accident rate.

[0003] In recent years, data-driven research methods have gradually become a hot topic in traffic prediction. Data-driven traffic flow prediction methods are mainly based on statistics and machine learning. Traditional statistics-based prediction methods mainly include support vector machines, autoregressive moving average methods, Kalman filtering methods, etc., and their ability to capture non-linear features between spatio-temporal sequences is not ideal. Due to its deep data expression ability, deep learning has been widely used in traffic prediction. In the early research process, the recurrent neural network (RNN) was directly trained with traffic data. Cui et al. explored the performance of stacked long short-term memory networks (LSTM) in flow prediction and achieved excellent results using a pure RNN structure.

[0004] However, the RNN-based model only uses the temporal features of travel behavior and ignores the potential spatial dependence in the transportation network. To capture the spatial dependence in this data, a number of methods have emerged that introduce Convolutional Neural Networks (CNNs) to build prediction models. Shi [8] et al. incorporated convolution into the gating mechanism to construct a Conv-LSTM model for predicting recent precipitation in a region. However, CNN-based models only consider the absolute distance relationship between stations in two-dimensional Euclidean space, and the non-Euclidean structural characteristics in the transportation network are not fully learned. Thus, the proposal of graph convolution generalizes traditional convolution to non-Euclidean structured data, and graph neural networks with different structures have greatly promoted related research on traffic flow prediction. Although the traffic flow prediction algorithms based on graph neural networks have been quite diverse, most methods still do not fully utilize complex spatio-temporal correlations. Shao Zezhi et al. believe that for time series, there are two different types of signals, namely diffusion signals and intrinsic signals. Based on this, more accurate modeling of different parts of traffic data is beneficial to improving the accuracy of prediction. Ma Chi et al. proposed a decoupled traffic flow prediction model based on an optimized complete graph attention network. The decoupling layer of this model uses quadratic decomposition to decompose traffic flow into trend-time representations, and separately processes the stable characteristics of trends and the fluctuating characteristics of events.

[0005] Based on the above problems, there is an urgent need for a prediction method that can stably capture coupled information and maintain high accuracy. Summary of the Invention

[0006] In view of the above problems, the object of the present invention is to propose a short-term rail transit flow prediction method and device based on a decoupling process. By decoupling the original flow signal, the ability of the traffic flow prediction network to learn complex spatio-temporal correlations is strengthened, and the accuracy and generalization of the short-term flow of rail transit passengers under complex conditions are improved.

[0007] The technical solution of the present invention is as follows: A short-term rail transit flow prediction method based on a decoupling process according to the present invention has the following operating steps:

[0008] (1): AFC card swiping data collection and processing; The data collection uses the rail transit route as the research object, with a half-hour statistical interval, and records the passenger card swiping data into a database file; The data processing is to screen and eliminate redundant values, outliers, and missing values in the collected data, and finally obtain passenger OD data;

[0009] (2): Construct a multi-modal convolutional feature map; the multi-modal convolutional feature map is based on the physical topology of the real rail transit within the expressway surrounding area, uses the OD data before and after dynamic time warping as content values, decouples them into passenger flow signals in traffic lines and at traffic stations, and establishes a multi-modal convolutional feature map.

[0010] (3): Design an IST-LSTM (Implement Swin Transformer Long Short-Term Memory) prediction unit; the IST-LSTM prediction unit adopts the basic Cell structure of LSTM, integrates the Swin Transformer module into the Cell and adjusts the original gating structure to establish an IST-LSTM prediction unit as a new Cell.

[0011] (4): Construct a parallel prediction network; the parallel prediction network adopts a two-branch structure. The first branch stacks IST-LSTM units to capture the decoupled traffic flow features, and the second branch stacks SBULSTM (Stacked bidirectional and unidirectional Long Short-Term Memory) and LSTM (Long Short-Term Memory) units to learn the original traffic flow information before decoupling. Information interaction is carried out by the time difference method at the end of the two branches, and the prediction result is calculated through Dropout and the fully connected layer to obtain the evaluation index.

[0012] Furthermore, in step (1), the AFC card swiping data collection uses the rail transit route as the research object, with a half-hour statistical interval, and saves the passenger card swiping records to the database file.

[0013] The data processing is to screen and eliminate redundant values, outliers, and missing values in the collected data, and finally obtain the passenger OD data.

[0014] The specific process is as follows: First, since two AFC card swiping behaviors are required to enter and exit the subway station, one passenger OD travel behavior can be defined by this rule.

[0015] Second, one OD travel can be located through the passenger's one-card face number, the station number where the transaction occurs, and the transaction time. And records with missing corresponding in and out station transaction information or too long in and out station transaction time need to be deleted.

[0016] Finally, all OD data is statistically analyzed at time intervals of 15 minutes, 30 minutes, and 60 minutes to form the passenger OD data.

[0017] Furthermore, in step (2), the multi-modal convolutional feature map includes a physical topology feature map Gp , Passenger flow similarity graph G S and correlation ratio graph G c :

[0018] Among them, the physical topology feature graph G p is directly constructed according to the subway physical topology structure of the selected area; first, by constructing a physical connection matrix P ∈ R N×N , if the corresponding stations i and j are connected in the real world, then P(i, j) = 1, and if not, P(i, j) = 0. After determining the physical connection matrix P, perform row normalization operation on P to obtain the edge weight matrix W p , and the element W p (i, j) in the weight matrix is calculated as follows:

[0019]

[0020] In the formula, P(i, k) represents the connection relationship between stations i and j;

[0021] Passenger flow similarity graph G S is defined by the similarity of passenger flows; first, use Dynamic Time Warping (DTW) to evaluate the deviation between two time series, calculate the similarity score S(i, j) between stations i and j, and obtain the similarity score matrix S ∈ R N×N , and the calculation formula of DTW is as follows:

[0022] S(i, j) = exp(-DTW(X i , X j ))

[0023] In the formula, X i and X j represent the complete passenger flow time series of stations i and j, and exp(·) is the calculation formula of the exponential function;

[0024] Then, perform normalization on S to calculate the edge weight matrix W s , and its calculation formula is as follows:

[0025]

[0026] In the formula, when E s contains the edge connecting nodes i and k, L(E s , i, k) = 1, otherwise L(E s , i, k) = 0;

[0027] Correlation ratio graph G c is constructed from the distribution of passengers' departure and destination locations. First, by establishing a correlation ratio matrix C ∈ R N×N, and its calculation formula is as follows:

[0028]

[0029] In the formula, D(i, j) represents the total number of passengers from station i to j, then is the total number of passengers departing from station i.

[0030] Furthermore, in step (3), the designed IST-LSTM prediction unit adopts the LSTM basic Cell structure, integrates the Swin Transformer module into the Cell and adjusts the original gating structure, and establishes the IST-LSTM prediction unit as the new Cell. Its calculation process is as follows:

[0031] f t = σ(W f × [h t-1 , x t + c f )

[0032] i t = STB(W i × [h t-1 , x t + c i )

[0033] C t = f t × C t-1 + i t

[0034] O t = σ(W o × [h t-1 , x t + c o )

[0035] h t = O t × tanh(C t )

[0036] In the formula, f t is the forgetting gate, i t is the improved input gate, STB(·) is the Swin Transformer Block, C t is the memory cell, O t is the output gate, W f 、W i 、W o are weight matrices, σ and tanh are activation functions, x t is the input at time t, c f 、c i 、co is the bias;

[0037] Among them, the STB(·) is the window-based multi-head self-attention (W-MSA) and the shifted window-based multi-head self-attention (SW-MSA), and its calculation process is as follows:

[0038]

[0039]

[0040]

[0041]

[0042] In the formula, LN(·) represents layer normalization, MLP(·) is a multi-layer perceptron layer, and W-MSA(·) and SW-MSA(·) are multi-head self-attention and shifted window multi-head self-attention operations.

[0043] Furthermore, in step (4), the dual-branch parallel prediction network is as follows: the first branch is composed of stacked IST-LSTM units, and the second branch is composed of a stack of SBULSTM and LSTM, where the SBULSTM layer is a bidirectional LSTM layer and an LSTM layer;

[0044] The specific time difference method is as follows: combining the dynamic programming method in the Bootstrapping manner and the Monte Carlo method based on the sampling idea, and its calculation and update process formula is as follows:

[0045] V(S t ) ← V(S t ) + α(R t+1 + γV(S t+1 ) - V(S t ))

[0046] In the formula, R t+1 is the immediate reward of this state, V(S t+1 ) is the state value of the next step, γ is the attenuation coefficient, and R t+1 + γV(S t+1 ) is the time difference target;

[0047] Using the mean absolute error (MSE) as the model loss function, and the mean square error (MAE) and mean absolute percentage error (MAPE) as the final prediction result evaluation indicators, and their mathematical expressions are as follows:

[0048]

[0049]

[0050]

[0051] Furthermore, a short-term rail transit flow device based on a decoupling process includes a memory and a processor;

[0052] wherein, the memory is used to store a computer program that can run on the processor;

[0053] The processor is used to execute the steps of the short-term rail transit flow prediction method based on the decoupling process as described above when running the computer program.

[0054] The beneficial effects of the present invention are as follows: 1. The present invention decouples the diffusion of the original subway passenger flow from the inherent signal, constructs a multi-modal convolution feature map, and on the basis of being able to obtain the spatial feature map structure of the subway topological network, decouples the passenger flow signal into the passenger flow signals of the traffic lines and traffic stations, enabling the model to effectively learn the dynamic dependence of the space; 2. Aiming at the complex spatio-temporal features in traffic flow prediction, by integrating the Swin Transformer module and improving the gating mechanism in the original LSTM, the present invention proposes a new recurrent unit - IST-LSTM, which can effectively filter and transmit information, strengthen the model's learning of long sequence dependencies, and improve the representation ability of the input sequence. Through the improved gating unit, the model can achieve more accurate and flexible attention allocation; 3. The present invention proposes an improved dual-branch prediction neural network. The first branch uses the IST-LSTM unit, and improves the gating mechanism of the LSTM by integrating the Swin Transformer module to enhance the ability to capture time features and spatial features, thereby improving the processing effect of complex spatio-temporal data. The second branch uses the SBULSTM and LSTM units, and performs in-depth processing through time difference learning to improve the understanding of the correlation between sequences and the feature extraction ability, thereby improving the prediction accuracy and robustness of the overall network. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is the flowchart of the method of the embodiment of the present invention;

[0056] Figure 2 is the physical topology feature map of the embodiment of the present invention;

[0057] Figure 3 is the passenger flow similarity map of the embodiment of the present invention;

[0058] Figure 4 is the correlation ratio map of the embodiment of the present invention;

[0059] Figure 5 is the schematic diagram of the IST-LSTM model of the embodiment of the present invention;

[0060] Figure 6 It is a schematic diagram of the LSTM model according to an embodiment of the present invention;

[0061] Figure 7 It is a schematic diagram of SBULSTM according to an embodiment of the present invention;

[0062] Figure 8 It is a schematic diagram of Swin Transformer according to an embodiment of the present invention;

[0063] Figure 9 It is an experimental result diagram of the time difference method in step 4 of an embodiment of the present invention. Detailed implementation manners

[0064] The following further elaborates on the specific technical solutions of the present invention in combination with specific examples.

[0065] As shown in the figure, a short-term rail transit flow prediction method based on a decoupling process according to the present invention decouples the original subway passenger flow signal, constructs a multi-modal convolution feature map, and improves the gating mechanism of LSTM by integrating the Swin Transformer module, and proposes a new recurrent unit IST-LSTM. The parallel stacked SBULSTM model and the LSTM model are used to process the time dependence relationship, which can strengthen the extraction of the real space dependence and the capture of the long-term context sequence characteristics, and improve the comprehensive spatio-temporal prediction ability.

[0066] Such as Figure 1 shown, specifically including the following steps:

[0067] Step (1): Collection and processing of AFC card swiping data; taking the rail transit route within the area surrounded by the second ring expressway in a certain city as the research object, with a half-hour statistical interval, and saving the passenger card swiping records to a database file;

[0068] The data processing is as follows: First, two AFC card swiping behaviors are required to enter and exit the subway station, so a passenger OD travel behavior can be defined by this rule; second, a single OD travel can be located through the passenger's one-card face number, the station number where the transaction occurs, and the transaction time; and, records with missing corresponding inbound and outbound transaction information or too long inbound and outbound transaction times need to be deleted; finally, all OD data is statistically analyzed with time intervals of 15 minutes, 30 minutes, and 60 minutes respectively to form passenger OD data;

[0069] Step (2): Construct a multi-modal convolution feature map, based on the real rail transit physical topology within the area surrounded by the second ring expressway in a certain city, using the OD data before and after dynamic time warping as the content value, decoupling it into passenger flow signals in traffic lines and traffic stations, and establishing a multi-modal convolution feature map:

[0070] Physical topology feature graph G p Constructed directly according to the subway physical topology of the selected area; first, by constructing a physical connection matrix P ∈ R N×N , if the corresponding stations i and j are connected in the real world, then P(i,j) = 1, and if not, P(i,j) = 0. After determining the physical connection matrix P, perform row normalization on P to obtain the edge weight matrix W p , and the element W p (i,j) is calculated as follows:

[0071]

[0072] In the formula, P(i,k) represents the connection relationship between stations i and j;

[0073] Passenger flow similarity graph G S Defined by the similarity of passenger flows; first, use Dynamic Time Warping (DTW) to evaluate the deviation between two time series, calculate the similarity score S(i,j) between stations i and j, and obtain the similarity scoring matrix S ∈ R N×N , and the calculation formula of DTW is as follows:

[0074] S(i,j) = exp(-DTW(X i ,X j ))

[0075] In the formula, X i and X j represent the complete passenger flow time series of stations i and j, and exp(·) is the exponential function calculation formula;

[0076] Then, normalize S to calculate the edge weight matrix W s , and its calculation formula is as follows:

[0077]

[0078] In the formula, when E s contains the edge connecting nodes i and k, L(E s ,i,k) = 1, otherwise L(E s ,i,k) = 0;

[0079] Correlation ratio graph G c Constructed from the distribution of passengers' departure and destination locations. First, by establishing a correlation ratio matrix C ∈ R N×N , and its calculation formula is as follows:

[0080]

[0081] where \(D(i,j)\) represents the total number of passengers from station \(i\) to \(j\), then is the total number of passengers departing from station \(i\);

[0082] Step (3): Design the IST-LSTM prediction unit. Adopt the basic LSTM Cell structure, integrate the Swin Transformer module into the Cell and adjust the original gating structure to establish the IST-LSTM prediction unit as the new Cell. Its calculation process is as follows:

[0083] f t =\(\sigma(W\) f \(\times [h\) t-1 ,x\) t +c\) f )

[0084] i t =STB(W\) i \(\times [h\) t-1 ,x\) t +c\) i )

[0085] C t =f t \(\times C\) t-1 +i t

[0086] O t =\(\sigma(W\) o \(\times [h\) t-1 ,x\) t +c\) o )

[0087] h t =O t \(\times \tanh(C\) t )

[0088] where \(f\) t is the forget gate, \(i\) t is the improved input gate, \(STB(\cdot)\) is the Swin Transformer Block, \(C\) t is the memory cell, \(O\) t is the output gate, \(W\) f ,W\) i ,W\) o are weight matrices, \(\sigma\) and \(\tanh\) are activation functions, \(x\) t is the input at time \(t\), \(c\) f ,c\) i ,c\) o are biases;

[0089] The STB(·) is the window-based multi-head self-attention (W-MSA) and the shifted window-based multi-head self-attention (SW-MSA), and its calculation process is as follows:

[0090]

[0091]

[0092]

[0093]

[0094] In the formula, LN(·) represents layer normalization, MLP(·) is a multi-layer perceptron layer, and W-MSA(·) and SW-MSA(·) are multi-head self-attention and shifted window multi-head self-attention operations;

[0095] Step (4): Construct a parallel prediction network with a two-branch structure. The first branch is composed of stacked IST-LSTM units, and the second branch is composed of a stack of SBULSTM and LSTM. The SBULSTM layer is a bidirectional LSTM layer and an LSTM layer;

[0096] The time difference method is as follows: Combine the dynamic programming method of the Bootstrapping method and the Monte Carlo method based on the sampling idea, and its calculation update process formula is as follows:

[0097] V(S t )←V(S t )+α(R t+1 +γV(S t+1 )-V(S t ))

[0098] In the formula, R t+1 is the immediate reward of this state, V(S t+1 ) is the state value of the next step, γ is the attenuation coefficient, and R t+1 +γV(S t+1 ) is the time difference target;

[0099] Use the mean absolute error (MSE) as the model loss function, and the mean square error (MAE) and mean absolute percentage error (MAPE) as the evaluation indicators for the final prediction result. The mathematical expressions are as follows:

[0100]

[0101]

[0102]

[0103] An apparatus according to the present invention includes a memory and a processor, wherein:

[0104] The memory is configured to store a computer program that can run on the processor;

[0105] The processor is configured to execute the steps of the short-term rail transit flow prediction method based on the decoupling process as described above when running the computer program.

Claims

1. A short-term rail traffic flow prediction method based on a decoupling process, characterized in that: The operation steps are as follows: (1): AFC punch-in data collection and processing; (2): Construct a multimodal convolutional feature map; (3): Design IST-LSTM prediction unit; (4): Build a parallel prediction network.

2. According to claim 1, a method for predicting short-term rail traffic flow based on a decoupling process is characterized in that: In step (1), the AFC punch-in data collection and processing are specifically: The data collection uses rail transit routes as the research object, takes half an hour as the statistical interval, and saves the passenger check-in records into the database file; The data processing is to screen and remove redundant values, abnormal values ​​and missing values ​​in the collected data, and finally obtain the passenger OD data.

3. The short-term rail traffic flow prediction method based on the decoupling process according to claim 2 is characterized in that: The specific process of the data processing is: First, two AFC punch-in behaviors are required to enter and exit the subway station, and this rule defines one passenger OD travel behavior; Secondly, locate an OD trip through the passenger's card number, the station number where the transaction occurred, and the time when the transaction occurred; and delete the records with missing entry and exit transaction information or too long entry and exit transaction time; Finally, all OD data are counted with time intervals of 15 minutes, 30 minutes, and 60 minutes to form passenger OD data.

4. The short-term rail traffic flow prediction method based on the decoupling process according to claim 1 is characterized in that: In step (2), the construction of the multimodal convolutional feature map is specifically: based on the actual rail transit physical topology, using the OD data before and after dynamic time warping as the content value, decoupling it into passenger flow signals in the traffic line and at the traffic station, and establishing the multimodal convolutional feature map.

5. The short-term rail traffic flow prediction method based on the decoupling process according to claim 4 is characterized in that: The multimodal convolutional feature map includes a physical topological feature map G p , passenger flow similarity graph G S Correlation ratio graph G c : Among them, the physical topological feature graph G p Directly build according to the physical topology of the subway in the selected area; first, by constructing a physical connection matrix P∈R N×N , if the corresponding sites i and j are connected in the real world, then P(i,j) = 1, if they are not connected, then P(i,j) = 0. After determining the physical connection matrix P, normalize the rows of P to obtain the edge weight matrix W p , the element W at each position of the weight matrix p The calculation formula for (i,j) is as follows: In the formula, P(i,k) represents the connectivity relationship between sites i and j; Passenger flow similarity graph G S Defined by the similarity of passenger flows; first, dynamic time warping is used to evaluate the deviation between the two time series, and the similarity score S(i,j) between stations i and j is calculated to obtain the similarity score matrix S∈R N×N , the calculation formula of DTW is as follows: S(i,j)=exp(-DTW(X i ,X j )) Where, X i With X j represents the complete passenger flow time series of station i and j, exp(·) is the exponential function calculation formula; Then, S is normalized to calculate the edge weight matrix W s , which is calculated as follows: In the formula, when E s When it contains an edge connecting nodes i and k, L(E s ,i,k)=1, otherwise L(E s ,i,k)=0; Correlation ratio graph G c It is constructed by the distribution of passengers' origin and destination. First, a correlation ratio matrix C∈R is established. N ×N , which is calculated as follows: Where D(i,j) represents the total number of passengers from station i to station j. is the total number of passengers departing from station i.

6. The short-term rail traffic flow prediction method based on the decoupling process according to claim 1 is characterized in that: In step (3), the designed IST-LSTM prediction unit adopts the LSTM basic Cell structure, integrates the SwinTransformer module into the Cell and adjusts the original gating structure to establish the IST-LSTM prediction unit as a new Cell. The calculation process is as follows: f t =σ(W f ×[h t-1 ,x t ]+c f ) i t =STB(W i ×[h t-1 ,x t ]+c i ) C t =f t ×C t-1 +i t The t =σ(W o ×[h t-1 ,x t ]+c o ) h t =O t ×tanh(C t ) In the formula, f t For the forget gate, i t is the improved input gate, STB(·) is the Swin Transformer Block, C t For memory cells, O t is the output gate, W f , W i , W o is the weight matrix, σ and tanh are activation functions, x t is the input at time t, c f 、c i 、c o For bias.

7. The short-term rail traffic flow prediction method based on the decoupling process according to claim 6 is characterized in that: in, The STB(·) is the window-based W-MSA and the shift window-based SW-MSA, and its calculation process is: Where LN(·) represents layer normalization, MLP(·) is a multi-layer perception layer, W-MSA(·) and SW-MSA(·) are multi-head self-attention and shift window multi-head self-attention operations.

8. The short-term rail traffic flow prediction method based on the decoupling process according to claim 1 is characterized in that: In step (4), the parallel prediction network is constructed using a dual-branch structure, the first branch stacks IST-LSTM units to capture the traffic characteristics after decoupling, and the second branch stacks SBULSTM and LSTM units to learn the original traffic information before decoupling; At the end of the two branches, the time difference method is used to exchange information, and the prediction results are calculated after the Dropout and fully connected layers to obtain the evaluation index. The SBULSTM layer is a bidirectional LSTM layer and an LSTM layer.

9. The short-term rail traffic flow prediction method based on the decoupling process according to claim 8 is characterized in that: The time difference method is specifically: combining the dynamic programming method of Bootstrapping with the Monte Carlo method based on sampling idea, and its calculation update process is as follows: V(S t )←V(S t )+α(R t+1 +γV(S t+1 )-V(S t )) In the formula, R t+1 is the immediate reward of this state, V(S t+1 ) is the state value of the next step, γ is the attenuation coefficient, R t+1 +γV(S t+1 ) is the time difference target; Using MSE as the model loss function, MAE and MAPE as the final prediction result evaluation indicators, the mathematical expressions are as follows:

10. A short-term rail traffic flow device based on a decoupling process, characterized in that: including memory and processor; Wherein, the memory is used to store a computer program that can be run on the processor; The processor is used to execute the steps of the short-term rail traffic flow prediction method based on the decoupling process as described above when running the computer program.

Images