Rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning

Abstract

The invention discloses a rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning, and the method comprises the steps: obtaining original passenger flow data collected by a subway AFC system, carrying out the data cleaning, and extracting subway passenger flow distribution features; the method comprises the following steps: acquiring subway passenger flow environment characteristics including climate and date characteristics, station attribute characteristics and inter-station adjacency characteristics, and preprocessing the characteristics; a multi-dimensional feature fusion model GLFCN (GCN-BiLSTM-FCN) is integrated, and a space-time passenger flow prediction module is constructed. The GCN module is used for processing subway passenger flow distribution characteristics and inter-station adjacency characteristics; the BiLSTM module is responsible for processing climate and date characteristics; the FCN module is used for processing site attribute features; and performing weighted fusion on the features processed by each module through a Gating mechanism, and outputting a prediction result. According to the method, the influence of the subway passenger flow environment characteristics on passenger flow prediction is deeply mined through the integrated model GLFCN, a subway operator is helped to optimize scheduling and operation strategies, the management efficiency and the service level are improved, the travel experience of passengers is improved, and the congestion problem is reduced.

Description

Technical Field

[0001] The invention relates to a rail transit network passenger flow prediction method based on multi-dimensional feature integrated learning, and belongs to the technical field of artificial intelligence. Background Art

[0002] In recent years, with the acceleration of urbanization and the rapid economic development, urban traffic pressure has continued to increase, especially in large and medium-sized cities. Traffic congestion has become an important issue affecting the quality of life of residents and the efficiency of economic operation. Urban rail transit, as a safe, convenient, punctual and environmentally friendly public transportation method, has become an important way to solve urban traffic problems because of its high carrying capacity and efficiency.

[0003] According to statistical data, the passenger flow of urban rail transit shows a rapid growth trend, especially during peak hours in the morning and evening, when subways and other rail transit vehicles carry a large number of commuters in the city; this growth not only increases the load on the urban rail transit network, but also puts higher demands on its operation scheduling, service quality and passenger experience; at the same time, the passenger flow of urban rail transit has obvious temporal and spatial fluctuation characteristics, and is affected by many factors, such as weather, holidays, social activities and emergencies; therefore, how to accurately predict the passenger flow of rail transit has become a key issue in optimizing the allocation of urban rail transit resources and improving operational efficiency and service levels.

[0004] Passenger flow forecasting can provide a basis for macro-planning, passenger flow control and guidance of urban rail transit networks, and is also an important reference for urban traffic management. Currently, existing rail transit passenger flow forecasting methods mostly use single-mode data input, such as historical passenger flow data or simple time series analysis. However, this traditional method is often difficult to fully capture the multi-dimensional characteristics of passenger flow when dealing with complex and dynamic urban environments. With the continuous advancement of data collection and processing technology, especially the development of big data, artificial intelligence and Internet of Things technology, there is an urgent need for multi-dimensional features to be used in passenger flow forecasting models, such as historical passenger flow data, weather data, traffic event data, urban activity data, etc. By comprehensively utilizing these spatiotemporal data, the changing trend of rail transit passenger flow can be more accurately portrayed, thereby improving the accuracy of the forecast. Summary of the invention

[0005] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning, which can fully integrate multiple data sources and use artificial intelligence technology to make high-precision predictions of the passenger flow of the rail transit system. This not only helps subway operators optimize scheduling and operation strategies, but also improves passengers' travel experience, reduces congestion, and thus realizes efficient and intelligent management of urban rail transit.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning, the method comprising the following steps:

[0007] Step 1: Obtain the original passenger flow data collected by the AFC system for all subway stations, including station names, entry and exit times, and passenger flow; perform data cleaning on the original passenger flow data, including outlier processing and selecting appropriate time granularity to obtain subway passenger flow distribution characteristics; obtain subway passenger flow environmental characteristics, including climate and date characteristics, station attribute characteristics, and inter-station adjacency characteristics; and perform normalization on the above characteristics;

[0008] Step 2: Integrate the multi-dimensional feature fusion model GLFCN to build a spatiotemporal passenger flow prediction module. The GCN module emphasizes the relationship and structure between nodes. When learning node features, it considers the node's neighbor node information, which is used to process passenger flow distribution characteristics and inter-station adjacency characteristics; the BiLSTM module can capture bidirectional time dependencies, process long-term and short-term dynamic information, and model nonlinear relationships, which is used to process climate and period characteristics; the FCN module can receive inputs of any size and retain spatial structures. Through convolution operations, it can effectively extract local dependencies of features, which is used to process station attribute features; Gating is a mechanism for dynamically adjusting information flow. It generates weights based on the input features to control which information passes and which information is suppressed. The features processed by each module are weighted and fused through the Gating mechanism to output the prediction results;

[0009] Step 3: Use the subway passenger flow distribution characteristics and environmental characteristics as the input of the spatiotemporal passenger flow prediction module to fit the GLFCN model; input the test set into the model to infer the subway passenger flow, compare the performance with the benchmark model and evaluate the performance of the GLFCN model.

[0010] As an improvement of the present invention, in addition to the passenger flow distribution characteristics, the step 1 of the present invention also mines the external environment characteristics of the passenger flow, including climate and date characteristics, station attribute characteristics and inter-station adjacency characteristics; the subway passenger flow is obtained by the subway AFC system. When passengers enter and exit the station through the gate, the AFC system will record the passenger's entry time, entry station, passenger type, exit time, exit station and travel cost; the original data recorded by the system may contain erroneous or incomplete information, and the original data needs to be deduplicated, outliers and missing values ​​are processed; the cleaned data is summarized according to time, that is, every hour, to form passenger flow distribution data; it is divided into granularity with every 15 minutes as a time node to obtain the passenger flow distribution characteristics W f ;

[0011] The climate and date characteristics include temperature data, rainfall data, holiday and working day data. The above data are linearly spliced ​​and divided into granularities with every 15 minutes as a time node to obtain the climate and date characteristics W d , whose length is consistent with the characteristic length of passenger flow distribution;

[0012] The external environment characteristics of subway passenger flow in the present invention include: POI data within 400 to 700 meters of each node, land mixed entropy data, traffic buffer data, station type data and traffic accessibility data; the above data are linearly spliced ​​without granularity division processing to obtain the station characteristics W s ;

[0013] The inter-station adjacency feature is the actual distance between adjacent stations of each subway station. According to the total number of nodes N, a feature matrix W with a shape of N×N is constructed g , determine whether the subway stations are adjacent and assign values ​​to them;

[0014] Normalize the subway passenger flow distribution characteristics and environmental characteristics.

[0015] As an improvement of the present invention, the specific method of integrating the multidimensional feature fusion model GLFCN in step 2 of the present invention to construct a spatiotemporal passenger flow prediction module includes:

[0016] Construct the GCN module and analyze the inter-station adjacency feature W g Normalization is performed to make the information propagation in the convolution operation more stable, which is beneficial to the effectiveness of graph convolution and obtains the processed inter-station adjacency features.

[0017]

[0018] Where W g is the original adjacency matrix, I N is the identity matrix of size N×N, where N is the number of nodes, used to add self-loops to each node, and D is the adjacency matrix W after adding self-loops g +I N The degree matrix of the degree matrix, the diagonal elements of the degree matrix D ii is the degree of each node, It is the inverse of the square root of the degree matrix and is used for normalization.

[0019] Input passenger flow distribution characteristics W f The dimension is B×2N×T (batch size B, number of nodes N, number of time steps T). The first N node features are subjected to two linear transformations and graph convolutions and activated to obtain the inbound passenger flow features.

[0020]

[0021] Reuse the GCN module, perform two linear transformations and graph convolutions on the features of the last N nodes to obtain the outbound passenger flow features.

[0022]

[0023] Where W f [:,:N,:]、W f [:,N:,:] represents the selected flow distribution feature W f The first and last N node features are the inbound and outbound passenger flow features. W1 and W2 are the weight matrices of the two linear transformations. is the normalized adjacency matrix and ReLU is the activation function.

[0024] The characteristics of the incoming passenger flow and outbound passenger flow characteristics Linear superposition is performed on the second dimension, the node dimension, to output the passenger flow distribution characteristics after processing by the GCN module The dimensions are B×2N×1:

[0025]

[0026] Construct a BiLSTM module and input the climate and date features W d The dimension is B×T×F (batch size B, time steps T, features F). The BiLSTM module consists of two LSTM networks. The first one is a bidirectional LSTM network. The output h1 of the bidirectional LSTM concatenates the hidden states in the two directions together. The output shape is B×T×2·H hid That is, each time step will have hidden state information in two directions:

[0027]

[0028] In the formula, x t is the climate and date characteristics input at time t, h t-1 , c t-1 are the hidden state and memory state of the previous time step, and h1 is the concatenation of the forward and backward hidden states;

[0029] The second one is a unidirectional LSTM network, which processes time series data only in the forward direction. h1 is input, and after being processed by the unidirectional LSTM, h2 is output, with a shape of B×T×1. Mean pooling is performed on the second dimension of h2, i.e., the time dimension, to obtain h mean , the shape is B×1; for h mean Perform dimension expansion and broadcast to all nodes, outputting the time series features after BiLSTM processing The shape is B×2N×1:

[0030]

[0031] In the formula, unsqueeze(dim=1) is a dimension expansion operation, repeat(1, 2N, 1) is a broadcast operation, and N is the number of nodes.

[0032] Construct the FCN module, input the site attribute feature W s The dimension is B×N×F (batch size B, number of nodes N, number of features F), for site feature W s Perform two convolutions. The first convolution is a 1D convolution with a kernel size of 3, an input channel of F, and an output channel of hid_c. The convolution features are batch normalized and activated with the ELU function, with a shape of B×hid c ×N:

[0033] h1=ELU(BatchNorm(W1*W s +b1)) (11)

[0034] Where W1 is a convolution kernel of size hid_c×F×3, * represents the convolution operation, b1 is the bias vector, BatchNorm standardizes the data of each channel and learns the scaling and offset parameters, and ELU is the activation function.

[0035] The feature h1 after convolution is subjected to Dropout processing, and a part of neurons are randomly discarded to prevent overfitting. The second convolution is performed. The second convolution is 1D convolution, the convolution kernel size is 3, the input channel is hid_c, and the output channel is 1. The feature after convolution is activated with the ReLU function, and the output shape is B×1×N:

[0036] h2=ReLU(W2*h1+b2) (12)

[0037] Where W2 is the convolution kernel of size 1×F×3, and b2 is the bias vector.

[0038] Exchange the second and third dimensions of the convolutional feature h2, broadcast the second dimension to all nodes for feature matching, and output the environmental features processed by the FCN module The shape is B×2N×1:

[0039]

[0040] Where permute(0, 2, 1) is a dimension swap operation, which means swapping the second and third dimensions.

[0041] Gating is a mechanism for dynamically adjusting information flow. It generates weights based on input features to control which information is passed and which is suppressed. It uses the Gating mechanism to weight multiple features to achieve feature fusion. Initialize three learnable weights α f , β d , γ s , the features output by GCN, BiLSTM, and FCN modules Perform weighted fusion to obtain the predicted passenger flow:

[0042]

[0043] As an improvement of the present invention, in step 3 of the present invention, the test set is input into the model to infer the subway passenger flow, and the performance is compared with the benchmark model to evaluate the performance of the model, including:

[0044] The present invention conducts experiments based on a real traffic flow dataset of a city subway. The experiment uses the passenger flow data of the entire network in 2018, of which the data of the first ten months are used as a training set, the data of the last two months are used as a validation set, and the passenger flow data of the entire network in December 2019 is used as a test set. The dataset takes 15 minutes as a time step. Through comparative experiments on real datasets, the effectiveness and accuracy of the rail transit network passenger flow prediction method based on multidimensional feature ensemble learning proposed in the present invention are verified.

[0045] Beneficial effects:

[0046] 1. The present invention can mine the distribution characteristics of passenger flow, and can also mine the external environmental characteristics of passenger flow, including the influence of climate and date characteristics, station attribute characteristics and inter-station adjacency characteristics.

[0047] 2. The present invention proposes a multi-dimensional feature integration model GLFCN, which is used to construct a spatiotemporal passenger flow prediction module. By combining a variety of deep learning technologies, in addition to the passenger flow distribution characteristics, the model also mines the external environment characteristics of subway passenger flow, which will help subway operators optimize scheduling and operation strategies, improve passengers' travel experience, reduce congestion, and thus realize efficient and intelligent management of urban rail transit.

[0048] 3. The multidimensional feature integration model GLFCN proposed in the present invention has been verified to be effective and accurate by comparing it with the benchmark model on a real data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The present invention is a flow chart of the method.

[0050] Figure 2It is the internal structure diagram of the model of the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0052] like Figure 1 As shown, the present invention provides a rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning, which specifically includes the following:

[0053] The subway passenger flow is obtained by the subway AFC system. When passengers enter and exit the station through the gate, the AFC system will record the passenger's entry time, entry station, passenger type, exit time, exit station and travel cost; the original data recorded by the system may contain erroneous or incomplete information, and it is necessary to remove duplicates, outliers and missing values ​​from the original data; the cleaned data is summarized according to time, that is, every hour, to form passenger flow distribution data; it is divided into granularity with every 15 minutes as a time node to obtain the passenger flow distribution feature W f ;

[0054] The climate and date characteristics include temperature data, rainfall data, holiday and working day data. The above data are linearly spliced ​​and divided into granularities with every 15 minutes as a time node to obtain the climate and date characteristics W d , whose length is consistent with the characteristic length of passenger flow distribution;

[0055] The site attribute features include: point of interest (POI) data, which can provide rich information about the business, transportation and pedestrian density of a certain area. According to the land use type, POIs are classified into six categories: residential areas, commercial areas, shopping centers, colleges and universities, medical facilities and scenic spots. POI data within 400 to 700 meters of each subway station are counted;

[0056] Land mixing entropy data. Land mixing entropy refers to the degree of mixing of land use types within the direct influence range of the station. The larger the value, the more land use types there are within the station's attraction range and the more evenly distributed they are. It is an indicator for evaluating the diversity and balance of land use in a specific area. The land mixing entropy data within 400 to 700 meters of each subway station are counted; the calculation formula is:

[0057] E=-∑ i p i ln(p i ) (15)

[0058] Where E is the land use mix index of each station, p i is the proportion of land use category i within the attraction range of the station, and i is the total number of land use categories;

[0059] Station type data: The passenger flow of stations of the same type has similar trends and characteristics. K-shape clustering is used to classify subway stations based on the historical passenger flow of in and out stations. The elbow rule is used to divide subway stations into four categories: commercial service, work-residence combination, colleges and universities, and scenic spots.

[0060] Traffic buffer zone data: bus stop buffer zones around subway stations have a significant impact on passenger flow, especially in urban transportation systems. As two important public transportation modes, subways and buses have a far-reaching impact on passenger flow through their linkage effect. The number of bus stop buffer zones within 400 to 700 meters of each subway station is counted.

[0061] Traffic accessibility data. The analysis of traffic accessibility reflects the attractiveness of urban rail transit routes and sites. The higher the traffic accessibility, the more attractive the site is. Therefore, it is crucial to the short-term passenger flow prediction of urban rail transit. By combining the types of subway stations, the data of passengers' accessibility in walking, cycling, and driving are counted; the calculation formula is:

[0062]

[0063] Where W represents the weight of the known point, d is the distance between the known point and the unknown point, and p is a specified power value used to control the degree of weight attenuation;

[0064] The above data are linearly spliced ​​without granularity division to obtain the site attribute feature W s ;

[0065] The subway passenger flow distribution characteristics and environmental characteristics are normalized using the minimum and maximum value method; the calculation formula is:

[0066]

[0067] In the formula, X represents the sample data, Min is the minimum value in the sample, Max is the maximum value in the sample, and X * Represents the sample data after normalization.

[0068] The inter-station adjacency feature is the actual distance between adjacent stations of each subway station. According to the total number of nodes N, a feature matrix W with a shape of N×N is constructed g , determine whether the nodes are adjacent and assign values ​​to them;

[0069] The calculation formula is:

[0070]

[0071] In the formula, j and v represent two adjacent or non-adjacent nodes respectively, and Distance (j，v)is the actual spatial distance between two adjacent nodes. Formula (18) is used for two adjacent nodes, and formula (19) is used for two non-adjacent nodes.

[0072] Integrate the multi-dimensional feature fusion model GLFCN (GCN-LSTM-FCN) to build a spatiotemporal passenger flow prediction module such as Figure 2 As shown;

[0073] Construct the GCN module and analyze the inter-station adjacency feature W g Normalization is performed to make the information propagation in the convolution operation more stable, which is beneficial to the effectiveness of graph convolution and obtains the processed inter-station adjacency features.

[0074]

[0075] Where W g is the original adjacency matrix, I N is the identity matrix of size N×N, where N is the number of nodes, used to add self-loops to each node, and D is the adjacency matrix W after adding self-loops g +I N The degree matrix of the degree matrix, the diagonal elements of the degree matrix D ii is the degree of each node, It is the inverse of the square root of the degree matrix and is used for normalization.

[0076] Input passenger flow distribution characteristics W f The dimension is B×2N×T (batch size B, number of nodes N, number of time steps T). The first N node features are subjected to two linear transformations and graph convolutions and activated to obtain the inbound passenger flow features.

[0077]

[0078] Reuse the GCN module, perform two linear transformations and graph convolutions on the features of the last N nodes to obtain the outbound passenger flow features.

[0079]

[0080] Where W f [:,:N,:]、W f [:,N:,:] represents the selected flow distribution feature W f The first and last N node features are the inbound and outbound passenger flow features. W1 and W2 are the weight matrices of the two linear transformations. is the normalized adjacency matrix, and the activation function is ReLU.

[0081] The characteristics of the incoming passenger flow and outbound passenger flow characteristics Perform linear superposition on the second dimension and output the passenger flow distribution features after processing by the GCN module The dimensions are B×2N×1:

[0082]

[0083] Construct a BiLSTM module and input the climate and date features W d The dimension is B×T×F (batch size B, time steps T, features F). The BiLSTM module consists of two LSTM networks. The first one is a bidirectional LSTM network. The output h1 of the bidirectional LSTM concatenates the hidden states in the two directions together. The output shape is B×T×2·H hid That is, each time step will have hidden state information in two directions:

[0084]

[0085] In the formula, x t is the climate and date characteristics input at time t, h t-1 , c t-1 are the hidden state and memory state of the previous time step, respectively, and h1 is the concatenation of the forward and backward hidden states;

[0086] The second one is a unidirectional LSTM network, which processes time series data only in the forward direction. h1 is input, and after being processed by the unidirectional LSTM, h2 is output, with a shape of B×T×1. Mean pooling is performed on the second dimension of h2, i.e., the time dimension, to obtain h mean , the shape is B×1; for h mean Perform dimension expansion and broadcast to all nodes, outputting the time series features after BiLSTM processing The shape is B×2N×1:

[0087]

[0088] In the formula, unsqueeze(dim=1) is a dimension expansion operation, repeat(1, 2N, 1) is a broadcast operation, and N is the number of nodes.

[0089] Construct the FCN module, input the site attribute feature W s The dimension is B×N×F (batch size B, number of nodes N, number of features F), for site feature W s Perform two convolutions. The first convolution is a 1D convolution with a kernel size of 3, an input channel of F, and an output channel of hid_c. The convolution features are batch normalized and activated with the ELU function, with a shape of B×hid c ×N:

[0090] h1=ELU(BatchNorm(W1*W s +b1)) (30)

[0091] Where W1 is a convolution kernel of size hid_c×F×3, * represents the convolution operation, b1 is the bias vector, BatchNorm standardizes the data of each channel and learns the scaling and offset parameters, and ELU is the activation function.

[0092] The feature h1 after convolution is subjected to Dropout processing, and a part of neurons are randomly discarded to prevent overfitting. The second convolution is performed. The second convolution is 1D convolution, the convolution kernel size is 3, the input channel is hid_c, and the output channel is 1. The feature after convolution is activated with the ReLU function, and the output shape is B×1×N:

[0093] h2=ReLU(W2*h1+b2) (31)

[0094] Where W2 is the convolution kernel of size 1×F×3, and b2 is the bias vector.

[0095] Exchange the second and third dimensions of the convolutional feature h2, broadcast the second dimension to all nodes, and output the environmental features processed by the FCN module The shape is B×2N×1:

[0096]

[0097] Where permute(0, 2, 1) is a dimension swap operation, which means swapping the second and third dimensions.

[0098] Gating is a mechanism for dynamically adjusting information flow. It generates weights based on input features to control which information is passed and which is suppressed. It uses the Gating mechanism to weight multiple features to achieve feature fusion. Initialize three learnable weights α f , β d , γ s , the features output by GCN, BiLSTM, and FCN modules Perform weighted fusion to obtain the predicted passenger flow:

[0099]

[0100] The test set is input into the model to infer the subway passenger flow, and the performance is compared with the benchmark model to evaluate the performance of the model. The present invention experiments with a real traffic flow dataset of a city subway, with the first ten months of the station passenger flow in 2018 as the training set, the last two months as the validation set, and the station passenger flow in December 2019 as the test set. The dataset has a time step of 15 minutes; the effectiveness and accuracy of the rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning proposed by the present invention are verified through comparative experiments on real datasets. The experimental results are shown in Table 1; the present invention is compared with other benchmark methods: including CNN, RNN, LSTM, GRU, Attention-LSTM, CNN-LSTM and GCN; three metrics are used for evaluation, including mean absolute error MAE, root mean square error RMSE and mean absolute percentage error MAPE. The values ​​of the above three metrics are all as small as possible, and the corresponding calculation formulas are as follows:

[0101]

[0102] In the formula, y i is the observed value of subway passenger flow, is the predicted value of subway passenger flow, n represents the length of the time series to be predicted, and i represents the i-th predicted time step.

[0103] It can be found from Table 1 that the prediction performance of the GLFCN model is better than other benchmark models in all three indicators on the two datasets. The GLFCN model can obtain a more semantically rich spatial feature model by extracting spatial dependencies from multiple angles, which enables the model to have better prediction performance.

[0104]

[0105] Table 1: Performance comparison of GLFCN and baseline models

[0106] The present invention proposes a rail transit network passenger flow prediction method based on multi-dimensional feature integrated learning. Through spatiotemporal data analysis, external factors such as temperature, rainfall, POI, land mixing entropy, traffic accessibility, and station distance are introduced to quantitatively analyze the time, space, and station characteristics of the transportation railway network; an end-to-end multi-dimensional feature integrated model GLFCN is constructed, which will help subway operators optimize scheduling and operation strategies, improve passengers' travel experience, reduce congestion, and thus realize efficient and intelligent management of urban rail transit.

[0107] The above descriptions are merely embodiments of the present invention and are not intended to limit the present invention in any other form. Any modifications or equivalent changes made based on the technical essence of the present invention still fall within the scope of protection required by the present invention.

Claims

1. A rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning, characterized in that: The method comprises the following steps: Step 1: Obtain the original passenger flow data collected by the AFC system for all subway stations, including station names, entry and exit times, and passenger flow; perform data cleaning on the original passenger flow data, including outlier processing and selecting appropriate time granularity to obtain subway passenger flow distribution characteristics; obtain subway passenger flow environment characteristics, including climate and date characteristics, station attribute characteristics, and inter-station adjacency characteristics; Normalize the above features; Step 2: Integrate the multi-dimensional feature fusion model GLFCN to build a spatiotemporal passenger flow prediction module, in which the GCN module emphasizes the relationship and structure between nodes and considers the node's neighbor node information when learning node features, which is used to process passenger flow distribution characteristics and inter-station adjacency characteristics; The LSTM module can capture bidirectional time dependencies, process long-term and short-term dynamic information, and model nonlinear relationships, and is used to process climate and period characteristics. The FCN module accepts inputs of any size and retains spatial structures. It effectively extracts local dependencies of features through convolution operations, and is used to process site attribute features. Gating is a mechanism for dynamically adjusting information flow. It generates weights based on input features to control which information is passed and which is suppressed. The features processed by each module are weighted and fused through the Gating mechanism to output the prediction results. Step 3: Use the subway passenger flow distribution characteristics and environmental characteristics as the input of the spatiotemporal passenger flow prediction module to fit the GLFCN model; input the test set into the model to infer the subway passenger flow, compare the performance with the benchmark model and evaluate the performance of the GLFCN model.

2. According to claim 1, a rail transit network passenger flow prediction method based on multi-dimensional feature ensemble learning is characterized in that: The step 1 comprises: The subway passenger flow is obtained by the subway AFC system. When passengers enter and exit the station through the gate, the AFC system will record the passenger's entry time, entry station, passenger type, exit time, exit station and travel fee; the original data recorded by the system may contain erroneous or incomplete information, and the original data needs to be deduplicated, outliers and missing values ​​processed; the cleaned data is summarized according to time, that is, every hour, to form passenger flow distribution data; it is divided into granularity with every 15 minutes as a time node to obtain passenger flow distribution characteristics; The climate and date characteristics include temperature data, rainfall data, holiday and weekday data, which are linearly spliced ​​and divided into granularities with 15 minutes as a time node, and the length of which is consistent with the passenger flow distribution feature length; The site attribute features include: point of interest (POI) data, which provides rich information about the business, transportation and pedestrian density of a certain area. According to the land use type, POIs are classified into six categories: residential areas, commercial areas, shopping centers, colleges and universities, medical facilities and scenic spots. POI data within 400 to 700 meters of each subway station are counted; Land mixing entropy data. Land mixing entropy refers to the degree of mixing of land use types within the direct influence range of the station. The larger the value, the more land use types there are within the station's attraction range and the more evenly distributed they are. It is an indicator for evaluating the diversity and balance of land use in a specific area. The land mixing entropy data within 400 to 700 meters of each subway station are counted; the calculation formula is: E=-∑ i pp i ln(p i ) (1) Where E is the land use mix index of each station, p i is the proportion of land use category i within the attraction range of the station, and i is the total number of land use categories; Station type data: The passenger flow of stations of the same type has similar trends and characteristics. K-shape clustering is used to classify subway stations based on the historical passenger flow of in and out stations. The elbow rule is used to divide subway stations into four categories: commercial service, work-residence combination, colleges and universities, and scenic spots. Traffic buffer zone data: bus stop buffer zones around subway stations have a significant impact on passenger flow, especially in urban transportation systems. As two important public transportation modes, subways and buses have a far-reaching impact on passenger flow through their linkage effect. The number of bus stop buffer zones within 400 to 700 meters of each subway station is counted. Traffic accessibility data. The analysis of traffic accessibility reflects the attractiveness of urban rail transit routes and sites. The higher the traffic accessibility, the more attractive the site is. Therefore, it is crucial to the short-term passenger flow prediction of urban rail transit. By combining the types of subway stations, the data of passengers' accessibility in walking, cycling, and driving are counted; the calculation formula is: Where W represents the weight of the known point, d is the distance between the known point and the unknown point, and p is a specified power value used to control the degree of weight attenuation; Linearly concatenate the above data without granularity division to obtain site attribute characteristics; The subway passenger flow distribution characteristics and environmental characteristics are normalized using the minimum and maximum value method; the calculation formula is: In the formula, X represents the sample data, Min is the minimum value in the sample, Max is the maximum value in the sample, and X* represents the sample data after normalization; The inter-station adjacency feature is the actual distance between adjacent stations of each subway station. According to the total number of nodes N, a feature matrix W with a shape of N×N is constructed g , determine whether the nodes are adjacent and assign values ​​to them; The calculation formula is: In the formula, j and v represent two adjacent or non-adjacent nodes respectively, and Distance (j，v) is the actual spatial distance between two adjacent nodes. Formula (4) is used for two adjacent nodes, and formula (5) is used for two non-adjacent nodes.

3. The method for predicting passenger flow of rail transit network based on multi-dimensional feature ensemble learning according to claim 1, characterized in that: The specific method of integrating the multi-dimensional feature fusion model in step 2 to construct a spatiotemporal passenger flow prediction module includes: Construct a GCN module. The graph convolutional network emphasizes the relationship and structure between nodes. When learning node features, the node's neighbor node information is considered, which is used to process passenger flow distribution characteristics and inter-station adjacency characteristics; construct a BiLSTM module. The bidirectional long short-term memory recurrent network can capture bidirectional time dependencies, process long-term and short-term dynamic information, and model nonlinear relationships. It is used to process climate and period characteristics; construct an FCN module. The fully convolutional network receives input of any size and retains the spatial structure. It effectively extracts the local dependencies of features through convolution operations, which is used to process site attribute characteristics; Gating is a mechanism for dynamically adjusting information flow. It generates weights based on the input features to control which information passes and which information is suppressed. The Gating mechanism assigns weights to the three types of features after processing. Finally, the three types of features are weighted fused to obtain the final prediction result.

4. The method for predicting passenger flow of rail transit network based on multi-dimensional feature ensemble learning according to claim 1, characterized in that: Step 3 inputs the test set into the model to infer the subway passenger flow, compares the performance with the benchmark model and evaluates the performance of the model, including: The present invention conducts experiments based on a real traffic flow dataset of a city subway. The experiment uses the passenger flow data of the entire network in 2018, of which the data of the first ten months are used as a training set, the data of the last two months are used as a validation set, and the passenger flow data of the entire network in December 2019 is used as a test set. The dataset takes 15 minutes as a time step. Through comparative experiments on real datasets, the effectiveness and accuracy of the rail transit network passenger flow prediction method based on multidimensional feature ensemble learning proposed in the present invention are verified.

Images