WO2021174755A1 - 一种基于深度学习的轨道交通客流需求预测方法和装置 - Google Patents

一种基于深度学习的轨道交通客流需求预测方法和装置 Download PDF

Info

Publication number
WO2021174755A1
WO2021174755A1 PCT/CN2020/104902 CN2020104902W WO2021174755A1 WO 2021174755 A1 WO2021174755 A1 WO 2021174755A1 CN 2020104902 W CN2020104902 W CN 2020104902W WO 2021174755 A1 WO2021174755 A1 WO 2021174755A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
convolution
passenger flow
dimensional
periodic
Prior art date
Application number
PCT/CN2020/104902
Other languages
English (en)
French (fr)
Inventor
韦伟
刘岭
刘军
张波
王舟帆
Original Assignee
北京全路通信信号研究设计院集团有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京全路通信信号研究设计院集团有限公司 filed Critical 北京全路通信信号研究设计院集团有限公司
Priority to EP20923491.3A priority Critical patent/EP4099237A4/en
Publication of WO2021174755A1 publication Critical patent/WO2021174755A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/40—Business processes related to the transportation industry
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning

Definitions

  • the present disclosure belongs to the field of rail transit passenger flow prediction, and in particular relates to a method and device for predicting rail transit passenger flow demand based on deep learning.
  • Rail transit is an important mode of transportation in my country, the backbone of transportation in my country, and the lifeline project of the city. It has the characteristics of large capacity, all-weather, safety, low energy consumption, and less pollution. It is very important for realizing the sustainable development of the city. Significance. With the rapid development of information technology, intelligence and informatization have become an important development direction of rail transit, and passenger flow demand forecasting is an important link in the realization of a modern rail transit system network. Accurate short-term passenger flow demand forecasting helps optimize the pre-allocation of transportation resources, reduce rail transit operating costs, and improve passenger travel convenience.
  • Passenger flow demand has dynamic time and space characteristics.
  • the key to accurately predicting passenger flow demand lies in how to accurately perceive the time and space dependence of passenger flow demand.
  • the passenger flow demand of origin-destination (ORIGIN-DESTINATION, OD) has a certain periodicity and trend.
  • the OD passenger flow demand at a specific moment will not only be dependent on the passenger flow demand of the nearby historical moment, but also related to the passenger flow demand of the historical moment in different cycles. For example, the OD passenger flow demand at the same time of the previous day, the same time of the previous week, and the same time of the previous month will have a certain impact on the OD passenger flow demand at that moment.
  • the demand for OD passenger flow has a coexistence of multiple periodic patterns in the time dimension, that is, multi-period correlation.
  • the spatial dimension the passenger flow demand between stations or lines in different spatial locations will affect each other, and the passenger flow demand from the same departure place to adjacent destinations or from adjacent departure places to the same destination will also affect each other.
  • OD passenger flow demand has more common and significant characteristics in the spatial dimension, which are the origin dependence and destination dependence. If there are many companies in place A, there is a clear correlation between the passenger flow from the adjacent departure place to place A in the morning, which reflects the departure place dependence (or correlation) of passenger flow demand. In the afternoon, the passenger flow from point A to the nearby destination is also correlated, which reflects the destination dependence (or correlation) of passenger flow demand.
  • the passenger flow demand of rail transit refers to the passenger flow demand from a specific departure place to a specific destination, and the passenger flow demand has complex dynamic non-linear temporal and spatial characteristics.
  • most of the existing rail transit OD passenger flow forecasting methods predict OD passenger flow demand based on time series characteristics, ignoring the widespread spatial correlation in OD passenger flow data. In a few methods to predict OD passenger flow demand based on temporal and spatial characteristics, only the spatial correlation between different OD passenger flow demands is considered, and the difference between the origin and destination dependence of OD passenger flow demand is ignored. In terms of time feature extraction, the time characteristics of OD passenger flow demand are not deeply explored, and the correlation of OD passenger flow demand under different cycle modes is not refined, and the multi-period correlation of rail transit OD passenger flow demand cannot be perceived.
  • the present disclosure proposes a deep learning-based rail transit passenger flow demand prediction method, which includes:
  • the extracting the spatial feature data in the periodic OD two-dimensional graph sequence data includes:
  • connection processing is performed on the destination-associated spatial characteristic information, the departure place-associated spatial characteristic information, and the mixed-associated spatial characteristic information to obtain spatial characteristic data.
  • performing the convolution residual processing on the data includes:
  • the one-dimensional row-stacked OD data uses a first convolution pooling function to perform convolution pooling processing
  • the first convolution pooling function is expressed as:
  • pool is the pooling layer
  • f is the activation function
  • w o (k) is the weight that the k-th network unit needs to learn
  • b o (k) is the bias that the k-th network unit needs to learn
  • the one-dimensional column stacked OD data uses a second convolution pooling function to perform convolution pooling processing
  • the second convolution pooling function is expressed as:
  • pool is the pooling layer
  • f is the activation function
  • w d (k) is the weight that the k-th network unit needs to learn
  • b d (k) is the bias that the k-th network unit needs to learn
  • the periodic OD two-dimensional graph sequence data is subjected to convolution pooling processing using a third convolution pooling function
  • the third convolution pooling function is expressed as:
  • X conv2 (k) pool(f(w conv2 (k)*X conv2 (k-1)+b conv2 (k))) ⁇ (3)
  • pool is the pooling layer
  • f is the activation function
  • w conv2 (k) is the weight that the k-th network unit needs to learn
  • b conv2 (k) is the bias that the k-th network unit needs to learn.
  • the step of performing feature extraction on the temporal feature data includes:
  • the prediction method further includes:
  • the prediction accuracy of the prediction method meets the requirements.
  • the present disclosure also provides a rail transit passenger flow demand prediction device based on deep learning, and the prediction device includes:
  • a data input module for collecting OD data, converting the OD data into periodic OD two-dimensional graph sequence numbers, and outputting the periodic OD two-dimensional graph sequence data;
  • a spatial feature information extraction module which uses a spatially complex correlation convolution residual network model to extract spatial feature data in the periodic OD two-dimensional graph sequence data;
  • the time information extraction module in the multi-period mode is used to extract the time feature data in the spatial feature data
  • the data output module performs feature extraction on the time feature data to obtain the OD passenger flow predicted value at the predicted time.
  • the spatially complex correlation convolution residual network model includes several groups of spatial correlation convolution residual networks, and the spatial correlation convolution residual network is used to extract spatial correlation feature data of periodic OD two-dimensional graph sequence data.
  • the spatially associated convolution residual network includes a data transformation module, a convolution residual module, and a spatial feature information connection module;
  • the data conversion module includes a row stack conversion module and a column stack conversion module
  • the row stack transformation module receives periodic OD two-dimensional graph sequence data transmitted from the data input layer, performs row stack transformation on the periodic OD two-dimensional graph sequence data, and outputs one-dimensional row stack OD data;
  • the column stack transformation module receives periodic OD two-dimensional graph sequence data transmitted from the data input layer, performs column stack transformation on the periodic OD two-dimensional graph sequence data, and outputs one-dimensional column stack OD data;
  • the convolution residual module includes a first convolution residual unit, a second convolution residual unit, and a third convolution residual unit;
  • the first convolution residual unit receives the one-dimensional row stack OD data transmitted by the row stack transformation module, performs convolution residual processing on the one-dimensional row stack OD data, and outputs destination associated spatial feature information ;
  • the second convolution residual unit receives the one-dimensional column stacking OD data transmitted from the column stacking transformation module, performs convolution residual processing on the one-dimensional column stacking OD data, and outputs the origin associated spatial feature information ;
  • the third convolution residual unit receives periodic OD two-dimensional graph sequence data transmitted from the data input layer, performs convolution residual processing on the periodic OD two-dimensional graph sequence data, and outputs mixed correlation spatial feature information ;
  • the spatial feature information connection module receives the spatial feature information transmitted by the first convolution residual unit, the second convolution residual unit and the third convolution residual unit, and is used to compare the The destination is associated with the spatial characteristic information, the departure place is associated with the spatial characteristic information, and the mixed-associated spatial characteristic information is connected for connection processing.
  • the first convolution residual unit includes a first convolution neural unit and a residual connection
  • the second convolution residual unit includes a second convolution neural unit and a residual connection
  • the third convolution residual unit includes a third convolution neural unit and a residual connection
  • the first convolutional neural unit performs convolution pooling processing on one-dimensional row-stacked OD data, and is a one-dimensional convolution unit;
  • the second convolutional neural unit performs convolution pooling processing on the one-dimensional column stacked OD data, and is a one-dimensional convolution unit;
  • the third convolutional neural unit performs convolution pooling processing on the periodic OD two-dimensional graph sequence data, and is a two-dimensional convolution unit.
  • the convolution pooling processing method of the first convolutional neural unit is:
  • X o (k) pool(f(w o (k)*X o (k-1)+b o (k)))
  • pool is the pooling layer
  • f is the activation function
  • w o (k) is the weight that the k-th network unit needs to learn
  • b o (k) is the bias that the k-th network unit needs to learn
  • the second convolutional neural unit convolution pooling processing method is:
  • X d (k) pool(f(w d (k)*X d (k-1)+b d (k))) ⁇
  • pool is the pooling layer
  • f is the activation function
  • w d (k) is the weight that the k-th network unit needs to learn
  • b d (k) is the bias that the k-th network unit needs to learn
  • the third convolutional neural unit convolution pooling processing method is:
  • X conv2 (k) pool(f(w conv2 (k)*X conv2 (k-1)+b conv2 (k))) ⁇
  • pool is the pooling layer
  • f is the activation function
  • w conv2 (k) is the weight that the k-th network unit needs to learn
  • b conv2 (k) is the bias that the k-th network unit needs to learn.
  • the time information extraction module in the multi-period mode receives the space feature data transmitted by the space feature information extraction module, and outputs the time feature data
  • the spatial feature information extraction layer is an extraction temporal feature model.
  • the extraction time feature model is a long and short-term memory neural network model.
  • the data output module receives the time characteristic data transmitted by the time information extraction module in the multi-period mode, and outputs the OD passenger flow predicted value at the predicted time;
  • the data output module includes a connection module and a dense module.
  • the connection module receives the time characteristic data transmitted by the time information extraction module in the multi-period mode, merges the time characteristic data, and outputs the fused time characteristic;
  • the intensive module receives the transmission from the connection module to fuse time features, performs feature extraction on the fused time features, and outputs the OD passenger flow prediction value at the predicted time.
  • the prediction device further includes an evaluation device
  • the evaluation device is used to evaluate the prediction accuracy of the rail transit passenger flow demand prediction method and device based on deep learning by using the average absolute error and the mean square error.
  • the present disclosure analyzes the multi-period relevance of the OD data, extracts characteristic data, and obtains the OD passenger flow prediction value at the prediction time, and the prediction accuracy is high.
  • Fig. 1 shows a flowchart of a method for predicting passenger flow demand for rail transit based on deep learning according to an embodiment of the present disclosure
  • Fig. 2 shows a model diagram of a device for predicting passenger flow demand for rail transit based on deep learning according to an embodiment of the present disclosure
  • Fig. 3 shows a schematic diagram of a spatial correlation convolution residual network structure according to an embodiment of the present disclosure
  • Fig. 4 shows a schematic structural diagram of a convolution residual unit according to an embodiment of the present disclosure
  • Fig. 5 shows a schematic structural diagram of a one-dimensional convolution unit according to an embodiment of the present disclosure
  • Fig. 6 shows a schematic structural diagram of a two-dimensional convolution unit according to an embodiment of the present disclosure.
  • the present disclosure is not limited to predicting rail traffic passenger flow, but is also applicable to other traffic passenger flows limited by destinations such as long-distance buses, ships, shipping, and so on.
  • prediction of rail traffic passenger flow is taken as an example.
  • the present disclosure provides a rail transit passenger flow demand prediction method based on deep learning.
  • the prediction method can adopt but is not limited to the following processes. Illustratively, as shown in FIG. 1, the prediction method includes:
  • Step 1 Collect OD data, and convert the OD data into periodic OD two-dimensional graph sequence data.
  • the OD data is collected, and the OD data is converted into periodic OD two-dimensional graph sequence data.
  • the OD data collection cycle is preset, and OD data is collected regularly within one collection cycle. Arrange and transform the collected OD data into two-dimensional OD images. According to the corresponding acquisition period, all OD two-dimensional images are divided into several groups of periodic OD two-dimensional image sequence data.
  • the OD data is data indicating the origin-destination passenger flow demand data.
  • the sampling period of periodic OD sequence 1 is 12 weeks
  • the sampling period of periodic OD sequence 2 is 7 days
  • the sampling period of periodic OD sequence 3 is 1 day
  • the sampling period of periodic OD sequence 4 is the same day. All data from the beginning to the moment before the forecast. In a cycle, OD data is collected every 5 minutes.
  • the OD two-dimensional map may adopt but not limited to the following structures:
  • the origin number and destination number may be different, and the following methods can be used but not limited to:
  • the origin is the following regions: Beijing, Shanghai, Tianjin, Chongqing..., serial number: 0, 1, 2, 3 ....
  • the destinations are the following areas: Beijing, Shanghai, Tianjin..., serial number: 0, 1, 2....
  • Beijing-Shanghai passenger flow demand is 50, Beijing-Tianjin passenger flow demand is 200, Shanghai-Beijing passenger flow demand is 60, Shanghai-Tianjin passenger flow demand is 7, Tianjin-Beijing passenger flow demand is 300, Tianjin- The demand for passenger flow in Shanghai is 8, the demand for Chongqing-Beijing passenger flow is 15, the demand for Chongqing-Shanghai passenger flow is 2, and the demand for Chongqing-Tianjin passenger flow is 18.
  • the corresponding OD graph is as follows:
  • Step 2 Use a spatially complex correlation convolution residual network model to extract spatial feature data from the periodic OD two-dimensional graph sequence data.
  • row stacking transformation is performed on the periodic OD two-dimensional graph sequence data to obtain one-dimensional row stacking OD data.
  • the data after row stacking transformation is:
  • i is the origin number
  • j is the destination number
  • n and m are positive integers
  • i is an integer between 0-n
  • j is an integer between 0-m.
  • column stacking transformation is performed on the periodic OD two-dimensional graph sequence data to obtain one-dimensional column stacking OD data.
  • the input data after column stacking transformation is:
  • i is the origin number
  • j is the destination number
  • n and m are positive integers
  • i is an integer between 0-n
  • j is an integer between 0-m.
  • convolution residual processing is performed on the one-dimensional row-stacked OD data to obtain destination-associated spatial feature information.
  • the convolution residual processing includes: convolution pooling processing for one-dimensional row-stacked OD data.
  • one-dimensional row stacking OD data requires multiple convolution pooling processing.
  • the first convolutional pooling function uses the first convolutional pooling function to perform a convolutional pooling process on the one-dimensional row-stacked OD data to obtain the destination-associated spatial feature information 1; again use the first convolutional pooling function to determine the output destination
  • the associated spatial feature information 1 is processed by convolution pooling to obtain the destination associated spatial feature information 2... and so on, the nth time uses the first convolutional pooling function to associate the output destination with spatial feature information n-1 Perform convolution pooling processing to obtain destination-associated spatial feature information n. In this way, multiple convolutional pooling processing for one-dimensional row-stacked OD data is completed.
  • the first convolution pooling function is expressed as:
  • X o (k) pool(f(w o (k)*X o (k-1)+b o (k)))
  • pool is the pooling layer
  • f is the activation function
  • w o (k) is the weight that the k-th network unit needs to learn
  • b o (k) is the bias that the k-th network unit needs to learn.
  • the pool pooling layer can use but is not limited to the following methods: Average pool pooling or Max pool pooling; the activation function f can be used but is not limited to the following activation functions: Relu activation function or Tanh activation function.
  • the convolution residual processing is performed on the one-dimensional column stacked OD data to obtain the associated spatial feature information of the departure place.
  • the convolution residual processing includes: convolution pooling processing for one-dimensional column stacking OD data.
  • one-dimensional column stacking OD data requires multiple convolution pooling processing.
  • the second convolutional pooling function uses the second convolutional pooling function to perform a convolutional pooling process on the one-dimensional column stacked OD data to obtain the origin associated spatial feature information 1; again use the second convolutional pooling function to determine the output of the origin Associated spatial feature information 1 is processed by convolution pooling, and the origin associated spatial feature information 2 is obtained.
  • the second convolution pooling function is used for the nth time to associate the output origin with spatial feature information n-1 Perform convolution pooling processing to obtain the associated spatial feature information n of the departure place. In this way, multiple convolution pooling processing for one-dimensional column stacked OD data is completed.
  • the second convolution pooling function is expressed as:
  • X d (k) pool(f(w d (k)*X d (k-1)+b d (k))) ⁇
  • pool is the pooling layer
  • f is the activation function
  • w d (k) and b d (k) are the weights and biases that the k-th network unit needs to learn, respectively.
  • the pool pooling layer can use but is not limited to the following methods: Average pool pooling or Max pool pooling; the activation function f can be used but is not limited to the following activation functions: Relu activation function or Tanh activation function.
  • convolution residual processing is performed on the periodic OD two-dimensional graph sequence data to obtain hybrid correlation spatial feature information.
  • the convolution residual processing includes: convolution pooling processing on the periodic OD two-dimensional graph sequence data.
  • the periodic OD two-dimensional graph sequence data requires multiple convolution pooling processing.
  • the third convolution pooling function to perform a convolution pooling process on the periodic OD two-dimensional graph sequence data to obtain mixed correlation spatial feature information 1; again use the third convolution pooling function to perform the output of the mixed correlation Spatial feature information 1 is processed by convolution pooling to obtain hybrid correlation spatial feature information 2...and so on, the third convolution pooling function is used for the nth time to convolve the output hybrid correlation spatial feature information n-1 Pooling process to obtain the mixed correlation spatial feature information n. In this way, multiple convolution pooling processing of the periodic OD two-dimensional graph sequence data is completed.
  • the sequence data of the input period OD two-dimensional graph is:
  • i is the origin number
  • j is the destination number
  • n and m are positive integers
  • i is an integer between 0-n
  • j is an integer between 0-m.
  • the third convolution pooling function is expressed as:
  • X conv2 (k) pool(f(w conv2 (k)*X conv2 (k-1)+b conv2 (k))) ⁇
  • pool is the pooling layer
  • f is the activation function
  • w conv2 (k) and b conv2 (k) are the weights and biases that the k-th network unit needs to learn, respectively.
  • the pool pooling layer can use but is not limited to the following methods: Average pool pooling or Max pool pooling; the activation function f can be used but is not limited to the following activation functions: Relu activation function or Tanh activation function.
  • connection processing is performed on the destination-associated spatial characteristic information, the departure place-associated spatial characteristic information, and the mixed-associated spatial characteristic information to obtain spatial characteristic data.
  • the output spatial feature data is expressed as:
  • X ti represents the spatial feature data at time ti output by the spatially associated convolution residual network.
  • an extraction time feature model is used to extract the time feature data in the spatial feature data.
  • the spatial feature is the spatial feature of consecutive time points under multiple cycles.
  • the model for extracting temporal features can use, but is not limited to, the following models: Long and Short-Term Memory Neural Network Model (LSTM).
  • LSTM Long and Short-Term Memory Neural Network Model
  • the long- and short-term memory neural network can learn the long-term time dependence of the sequence, input the spatial feature data X into the long- and short-term memory neural network, and output the time feature data.
  • the step of performing feature extraction on the temporal feature data includes:
  • the time feature data is fused to obtain time feature fusion data; the time feature fusion data is subjected to feature extraction to obtain the OD passenger flow prediction value at the desired prediction time.
  • the prediction method further includes: using the average absolute error and the mean square error to evaluate the rail transit passenger flow demand prediction value based on deep learning.
  • the mean absolute error (MAE) is:
  • MSE mean square error
  • the first threshold and the second threshold are preset values, which can be modified according to actual conditions.
  • a rail transit passenger flow demand prediction device based on deep learning includes:
  • a data input module for collecting OD data, converting the OD data into periodic OD two-dimensional graph sequence numbers, and outputting the periodic OD two-dimensional graph sequence data;
  • a spatial feature information extraction module which uses a spatially complex correlation convolution residual network model to extract spatial feature data in the periodic OD two-dimensional graph sequence data;
  • the time information extraction module in the multi-period mode is used to extract the time feature data in the spatial feature data
  • the data output module performs feature extraction on the time feature data to obtain the OD passenger flow predicted value at the predicted time.
  • the prediction device may adopt but is not limited to the following model: a rail transit passenger flow demand prediction device model based on deep learning, as shown in FIG. 2.
  • the model includes four layers: a data input layer, a spatial feature information extraction layer, a time information extraction layer in a multi-period mode, and a data output layer.
  • the data input layer collects OD data and outputs four-period OD two-dimensional graph sequence data.
  • the spatial feature information extraction layer uses a spatially complex correlation convolution residual network model to receive periodic OD two-dimensional graph sequence data transmitted from the data input layer, and output spatial feature data.
  • the spatially complex correlation convolution residual network model includes several groups of spatial correlation convolution residual networks (CNNRES_net).
  • the time information extraction layer uses a long and short-term memory network model, receives the space feature data transmitted by the space feature information extraction layer, and outputs the time feature data.
  • the data output layer includes a connection layer (Concatenate) and a dense layer (Dense), and receives the time characteristic data transmitted by the time information extraction layer in the multi-period mode, and outputs the predicted value of the OD passenger flow at the predicted time.
  • the spatially complex correlated convolutional residual network model includes several groups of spatially correlated convolutional residual networks.
  • the spatial correlation convolution residual network includes a data transformation module, a convolution residual module and a spatial feature information connection module.
  • the spatial correlation convolution residual network adopts but is not limited to the following structure, as shown in FIG. 3.
  • the spatially associated convolution residual network includes a row stack transformation module, a column stack transformation module, three convolution residual units and a spatial feature information connection module.
  • the three convolution residual units are respectively a first convolution residual unit, a second convolution residual unit and a third convolution residual unit.
  • the row stacking transformation unit is arranged at the input end of the first convolution residual unit, and the first convolution residual unit is used for extracting the spatial feature information associated with the destination.
  • the column stacking transformation unit is arranged at the input end of the second convolution residual unit, and the second convolution residual unit is used for extracting the spatial feature information associated with the departure place.
  • the periodic OD two-dimensional graph sequence data is directly input to the third convolution residual unit, and the third convolution residual unit is used to extract the mixed correlation spatial feature information.
  • the output ends of the first convolution residual unit, the second convolution residual unit, and the third convolution residual unit are connected to the input end of the spatial feature information connection module, and the spatial feature information connection module is used to compare the first convolution residual The unit, the second convolution residual unit and the third convolution residual unit output data fusion, and then output spatial feature data.
  • the first convolution residual unit includes a first convolution neural unit and a residual connection
  • the second convolution residual unit includes a second convolution neural unit and a residual connection
  • the third convolution residual unit includes a third convolution neural unit and a residual connection.
  • the first, second, and third convolutional residual units can adopt but are not limited to the following structures, as shown in Figure 4, where a multi-layer convolutional neural network unit is added to the superposition structure of a multi-layer convolutional neural network.
  • the residual connection of the output layer (or other layers close to the output layer).
  • Data is input to the first layer of convolutional neural network unit from the input, and after convolution pooling, it is passed to the second layer of convolutional neural network unit; after convolutional pooling, it is passed to the third layer of convolutional neural network.
  • the network unit ...
  • the OD data When the OD data is processed through the multi-layer convolutional neural network unit, the information will be lost; the residual connection is used to prevent the disappearance of the gradient information.
  • the first convolutional neural unit performs convolution pooling processing on one-dimensional row-stacked OD data, and is a one-dimensional convolution unit (Conv1Unit);
  • the second convolutional neural unit performs convolution pooling processing on the one-dimensional column stacked OD data, and is a one-dimensional convolution unit;
  • the third convolutional neural unit performs convolution pooling processing on the periodic OD two-dimensional graph sequence data, and is a two-dimensional convolution unit (Conv2Unit).
  • FIG. 5 Exemplarily, the structure of a one-dimensional convolution unit is shown in FIG. 5, including a one-dimensional convolution layer (Conv1) and a one-dimensional maximum pooling layer (maxpooling1D).
  • Conv1 one-dimensional convolution layer
  • maxpooling1D one-dimensional maximum pooling layer
  • the two-dimensional convolution unit structure is shown in FIG. 6, and includes a two-dimensional convolution layer (Conv2) and a two-dimensional maximum pooling layer (maxpooling2D).
  • Conv2 two-dimensional convolution layer
  • maxpooling2D two-dimensional maximum pooling layer
  • the first convolutional neural unit convolutional pooling processing method is:
  • X o (k) pool(f(w o (k)*X o (k-1)+b o (k)))
  • pool is the pooling layer
  • f is the activation function
  • w o (k) and b o (k) are the weights and biases that the k-th network unit needs to learn, respectively.
  • the pool pooling layer can use but is not limited to the following methods: Average pool pooling or Max pool pooling; the activation function f can be used but is not limited to the following activation functions: Relu activation function or Tanh activation function.
  • the second convolutional neural unit convolution pooling processing method is:
  • X d (k) pool(f(w d (k)*X d (k-1)+b d (k))) ⁇
  • pool is the pooling layer
  • f is the activation function
  • w d (k) and b d (k) are the weights and biases that the k-th network unit needs to learn, respectively.
  • the pool pooling layer can use but is not limited to the following methods: Average pool pooling or Max pool pooling; the activation function f can be used but is not limited to the following activation functions: Relu activation function or Tanh activation function.
  • the third convolutional neural unit convolution pooling processing method is:
  • X conv2 (k) pool(f(w conv2 (k)*X conv2 (k-1)+b conv2 (k))) ⁇
  • pool is the pooling layer
  • f is the activation function
  • w conv2 (k) and b conv2 (k) are the weights and biases that the k-th network unit needs to learn, respectively.
  • the pool pooling layer can use but is not limited to the following methods: Average pool pooling or Max pool pooling; the activation function f can be used but is not limited to the following activation functions: Relu activation function or Tanh activation function.
  • the time information extraction layer in the multi-period mode receives the space feature data transmitted by the space feature information extraction layer, and outputs the time feature data.
  • the spatial feature information extraction layer is an extraction temporal feature model.
  • the extraction time feature model adopts but is not limited to the following model: a long and short-term memory neural network model.
  • the time feature information extraction in the multi-period mode is to construct a time feature information extraction unit for different periods of input information, and each extraction unit is composed of a series of long and short-term memory neural network layers.
  • the data output layer receives the time characteristic data transmitted by the time information extraction layer in the multi-period mode, and outputs the predicted value of the OD passenger flow at the predicted time.
  • the data output layer includes a connection layer and a dense layer.
  • connection layer is set at the input end of the dense layer for fusing the temporal feature data.
  • the dense layer is used to perform feature extraction on the fusion time feature to obtain the OD passenger flow prediction value at the prediction time.
  • the prediction device further includes an evaluation device.
  • the evaluation device is used for evaluating the rail transit passenger flow demand prediction method and device based on deep learning by using the average absolute error and the mean square error.
  • the evaluation device adopts but is not limited to the following devices: the average absolute error device and the mean square error device; the following devices can also be used but not limited to: a device that includes both the average absolute error evaluation function and the mean square error evaluation function .

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • General Physics & Mathematics (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • General Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Development Economics (AREA)
  • Primary Health Care (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Educational Administration (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Game Theory and Decision Science (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Traffic Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于深度学习的轨道交通客流需求预测方法和装置，预测方法包括：采集OD数据，数据转化为周期OD二维图序列数据；将周期OD二维图序列数据输入至空间复杂关联卷积残差网络模型，输出空间特征数据；将空间特征数据输入至提取时间特征信息模型，输出时间特征数据；使用时间特征数据进行特征提取，得到预测时刻OD客流值；根据需要对预测方法进行评估。该方法通过对OD数据的多重周期关联性进行分析，提取特征数据，得到预测时刻的OD客流预测值，预测精度高。

Description

一种基于深度学习的轨道交通客流需求预测方法和装置
本申请要求于2020年3月2日递交的中国专利申请第202010133886.2号的优先权，在此全文引用上述中国专利申请公开的内容以作为本申请的一部分。
技术领域
本公开属于轨道交通客流量预测领域，特别涉及一种基于深度学习的轨道交通客流需求预测方法和装置。
背景技术
轨道交通是我国重要的运输方式，是我国交通运输的“骨干”，是城市的生命线工程，具有运量大、全天候、安全、低能耗、少污染等特点，对于实现城市的可持续发展具有非常重要的意义。随着信息技术的高速发展，智能化、信息化成为了轨道交通的重要发展方向，而客流需求预测是实现现代化轨道交通系统网络的重要环节。精准的短时客流需求预测有助于优化交通资源的预分配、降低轨道交通运营成本、提高旅客的出行便捷性。
客流需求具有动态时空特性。精准预测客流需求的关键在于如何准确感知客流需求的时间依赖性和空间依赖性。在时间维度上，出发地——目的地(ORIGIN-DESTINATION，OD)的客流需求存在一定的周期性、趋势性。某一特定时刻的OD客流需求不仅会与其临近的历史时刻客流需求存在依赖性，还与其不同周期的历史时刻客流需求有关。例如前一天的同期时刻、前一周的同期时刻、前一个月的同期时刻等的OD客流需求都会对该时刻的OD客流需求产生一定的影响。因此，OD客流需求在时间维度上存在多种周期模式并存的依赖关系，即多重周期关联性。在空间维度上，不同空间位置的车站或线路之间 的客流需求会相互影响，同一个出发地到相邻目的地或相邻出发地到同一目的地之间的客流需求也会互相影响。此外，OD客流需求在空间维度上还存在比较常见和显著的特性是出发地依赖性和目的地依赖性。假若A地有很多公司，则早晨从相邻出发地到达A地的客流量存在明显相关性，这体现了客流需求的出发地依赖(或关联)。下午从A地到达临近的目的地的客流量也存在相互关联性，这体现了客流需求的目的地依赖(或关联)。
目前，大多采取的建模方法是基于站点历史客流数据来预测短时进出站客流量。这种方法不能准确体现轨道交通客运需求量的大小。轨道交通客流需求是指特定的出发地到特定的目的地的客流需求，客流需求具有复杂的动态非线性时空特性。目前，既有的轨道交通OD客流量预测方法大多是基于时序特征预测OD客流需求，忽略了OD客流数据中广泛存在的空间关联性。在少数基于时空特征预测OD客流需求的方法中，仅仅考虑了不同OD客流需求之间的空间相关性，而忽略了OD客流需求的出发地依赖和目的地依赖的差异性。在时间特征提取方面，未深度挖掘OD客流需求的时间特性，未细化不同周期模式下OD客流需求的关联性，不能感知轨道交通OD客流需求的多重周期关联性。
发明内容
针对上述问题，本公开提出一种基于深度学习的轨道交通客流需求预测方法，所述方法包括：
采集OD数据，将所述OD数据转化为周期OD二维图序列数据；
使用空间复杂关联卷积残差网络模型，提取所述周期OD二维图序列数据中的空间特征数据；
使用提取时间特征模型，提取所述空间特征数据中的时间特征数 据；
对所述时间特征数据进行特征提取处理，得到预测时刻的OD客流预测值。
进一步地，所述提取所述周期OD二维图序列数据中的空间特征数据包括：
对所述周期OD二维图序列数据进行行堆叠变换，得到一维行堆叠OD数据；
对所述周期OD二维图序列数据进行列堆叠变换，得到一维列堆叠OD数据；
对所述一维行堆叠OD数据进行卷积残差处理，得到目的地关联空间特征信息；
对所述一维列堆叠OD数据进行卷积残差处理，得到出发地关联空间特征信息；
对所述周期OD二维图序列数据进行卷积残差处理，得到混合关联空间特征信息；
对所述目的地关联空间特征信息、出发地关联空间特征信息、混合关联空间特征信息进行连接处理，得到空间特征数据。
进一步地，对数据进行所述卷积残差处理包括：
对所述一维行堆叠OD数据卷积池化处理；
对所述一维列堆叠OD数据卷积池化处理；
对所述周期OD二维图序列数据卷积池化处理。
进一步地，所述一维行堆叠OD数据使用第一卷积池化函数进行卷积池化处理；
所述第一卷积池化函数表述为：
|X o(k)＝pool(f(w o(k)*X o(k-1)+b o(k)))·(1)·
式(1)中，pool是池化层，f是激活函数，w o(k)是第k层网络 单元需要学习的权重，b o(k)是第k层网络单元需要学习的偏置；
所述一维列堆叠OD数据使用第二卷积池化函数进行卷积池化处理；
所述第二卷积池化函数表述为：
X d(k)＝pool(f(w d(k)*X d(k-1)+b d(k)))···(2)
式(2)中，pool是池化层，f是激活函数，w d(k)是第k层网络单元需要学习的权重，b d(k)是第k层网络单元需要学习的偏置；
所述周期OD二维图序列数据使用第三卷积池化函数进行卷积池化处理；
所述第三卷积池化函数表述为：
X conv2(k)＝pool(f(w conv2(k)*X conv2(k-1)+b conv2(k)))··(3)
式(3)中，pool是池化层，f是激活函数，w conv2(k)是第k层网络单元需要学习的权重，b conv2(k)是第k层网络单元需要学习的偏置。
进一步地，对所述时间特征数据进行特征提取步骤为：
对所述时间特征数据进行融合处理，得到时间特征融合数据；
对所述时间特征融合数据进行特征提取，得到OD客流预测值。
进一步地，所述预测方法还包括：
利用平均绝对误差和均方误差对所述OD客流预测值进行评估；
获取真实值与预测值的平均绝对误差和均方误差，在所述平均绝对误差小于等于第一阈值时且所述均方误差小于等于第二阈值时，所述预测方法的预测精度满足要求。
本公开还提供一种基于深度学习的轨道交通客流需求预测装置，所述预测装置包括：
数据输入模块，用于采集OD数据，将所述OD数据转化为周期OD二维图序列数，并输出所述周期OD二维图序列数据；
空间特征信息提取模块，所述空间特征信息提取模块使用空间复杂关联卷积残差网络模型，用于提取所述周期OD二维图序列数据中 的空间特征数据；
多重周期模式下时间信息提取模块，用于提取所述空间特征数据中的时间特征数据；
数据输出模块，对所述时间特征数据进行特征提取，得到预测时刻的OD客流预测值。
进一步地，所述空间复杂关联卷积残差网络模型包括若干组空间关联卷积残差网络，所述空间关联卷积残差网络用于提取周期OD二维图序列数据的空间关联特征数据。
进一步地，所述空间关联卷积残差网络包括数据变换模块、卷积残差模块和空间特征信息连接模块；
所述数据变换模块包括行堆叠变换模块和列堆叠变换模块；
所述行堆叠变换模块，接收所述数据输入层传输来的周期OD二维图序列数据，对所述周期OD二维图序列数据进行行堆叠变换，输出一维行堆叠OD数据；
所述列堆叠变换模块，接收所述数据输入层传输来的周期OD二维图序列数据，对所述周期OD二维图序列数据进行列堆叠变换，输出一维列堆叠OD数据；
所述卷积残差模块包括第一卷积残差单元、第二卷积残差单元和第三卷积残差单元；
所述第一卷积残差单元，接收所述行堆叠变换模块传输来的一维行堆叠OD数据，对所述一维行堆叠OD数据进行卷积残差处理，输出目的地关联空间特征信息；
所述第二卷积残差单元，接收所述列堆叠变换模块传输来的一维列堆叠OD数据，对所述一维列堆叠OD数据进行卷积残差处理，输出出发地关联空间特征信息；
所述第三卷积残差单元，接收所述数据输入层传输来的周期OD二维图序列数据，对所述周期OD二维图序列数据进行卷积残差 处理，输出混合关联空间特征信息；
所述空间特征信息连接模块，接收所述第一卷积残差单元、所述第二卷积残差单元和所述第三卷积残差单元传输来的空间特征信息，用于对所述目的地关联空间特征信息、出发地关联空间特征信息、混合关联空间特征信息进行连接处理。
进一步地，所述第一卷积残差单元包括第一卷积神经单元和残差连接；
所述第二卷积残差单元包括第二卷积神经单元和残差连接；
所述第三卷积残差单元包括第三卷积神经单元和残差连接；
所述第一卷积神经单元对一维行堆叠OD数据进行卷积池化处理，为一维卷积单元；
所述第二卷积神经单元对一维列堆叠OD数据进行卷积池化处理，为一维卷积单元；
所述第三卷积神经单元对周期OD二维图序列数据进行卷积池化处理，为二维卷积单元。
进一步地，所述第一卷积神经单元卷积池化处理方式为：
X o(k)＝pool(f(w o(k)*X o(k-1)+b o(k)))
其中，pool是池化层，f是激活函数，w o(k)是第k层网络单元需要学习的权重，b o(k)是第k层网络单元需要学习的偏置；
所述第二卷积神经单元卷积池化处理方式为：
X d(k)＝pool(f(w d(k)*X d(k-1)+b d(k)))·
其中，pool是池化层，f是激活函数，w d(k)是第k层网络单元需要学习的权重，b d(k)是第k层网络单元需要学习的偏置；
所述第三卷积神经单元卷积池化处理方式为：
X conv2(k)＝pool(f(w conv2(k)*X conv2(k-1)+b conv2(k)))·
其中，pool是池化层，f是激活函数，w conv2(k)是第k层网络单元需要学习的权重，b conv2(k)是第k层网络单元需要学习的偏置。
进一步地，所述多重周期模式下时间信息提取模块，接收所述空间特征信息提取模块传输来空间特征数据，输出时间特征数据；
所述空间特征信息提取层为提取时间特征模型。
进一步地，所述提取时间特征模型为长短期记忆神经网络模型。
进一步地，所述数据输出模块，接收所述多重周期模式下时间信息提取模块传输来时间特征数据，输出预测时刻的OD客流预测值；
进一步地，所述数据输出模块包括连接模块和密集模块。
所述连接模块，接收所述多重周期模式下时间信息提取模块传输来时间特征数据，将所述时间特征数据融合，输出融合时间特征；
所述密集模块，接收所述连接模块传输来融合时间特征，对所述融合时间特征进行特征提取，输出预测时刻的OD客流预测值。
进一步地，所述预测装置还包括评估装置；
所述评估装置用于利用平均绝对误差和均方误差对基于深度学习的轨道交通客流需求预测方法和装置预测精度进行评估。
本公开通过对OD数据的多重周期关联性进行分析，提取特征数据，得到预测时刻的OD客流预测值，预测精度高。
本公开的其它特征和优点将在随后的说明书中阐述，并且，部分地从说明书中变得显而易见，或者通过实施本公开而了解。本公开的目的和其他优点可通过在说明书、权利要求书以及附图中所指出的结构来实现和获得。
附图说明
为了更清楚地说明本公开实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本公开的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1示出了根据本公开实施例的基于深度学习的轨道交通客流需求预测方法流程图；
图2示出了根据本公开实施例的基于深度学习的轨道交通客流需求预测装置模型图；
图3示出了根据本公开实施例的空间关联卷积残差网络结构示意图；
图4示出了根据本公开实施例的卷积残差单元结构示意图；
图5示出了根据本公开实施例的一维卷积单元结构示意图；
图6示出了根据本公开实施例的二维卷积单元结构示意图。
具体实施方式
为使本公开实施例的目的、技术方案和优点更加清楚，下面将结合本公开实施例中的附图，对本公开实施例中的技术方案进行清楚、完整地说明，显然，所描述的实施例是本公开一部分实施例，而不是全部的实施例。基于本公开中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本公开保护的范围。
本公开不限于预测轨道交通客流，还适用于长途汽车、轮船、航运等出发地目的地限定的其他交通客流，此处以预测铁路交通客流为例。本公开提供了一种基于深度学习的轨道交通客流需求预测方法，所述预测方法可以采用但不限于下列流程，示例性地，如图1所示，所述预测方法包括：
步骤一、采集OD数据，将所述OD数据转化为周期OD二维图序列数据。
具体的，采集OD数据，将OD数据转化为周期OD二维图序列数据。
预设OD数据的采集周期，在一个采集周期内，定时采集OD数据。将采集到的各OD数据整理转化为OD二维图。根据对应的采集周期，将所有OD二维图划分为若干组周期OD二维图序列数据。
所述OD数据是指出发地--目的地客流需求数据。
示例性的，预设四种采集周期：周期OD序列1的采样周期为 12周，周期OD序列2的采样周期为7天，周期OD序列3的采样周期为1天，周期OD序列4是当日从开始至预测前一时刻的所有数据。在一个周期内，每间隔5min，采集一次OD数据。
示例性的，所述OD二维图可以采用但不限于下列结构：
Figure PCTCN2020104902-appb-000001
其中，
Figure PCTCN2020104902-appb-000002
是t时刻车站i出发到车站j的客流需求量，i是出发地编号，j是目的地编号；n、m为正整数，i为0-n之间整数，j为0-m之间的整数。
示例性的，所述出发地编号与目的地编号可以不同，可以采用但不限于下列方式：如出发地为以下地区：北京、上海、天津、重庆…，依次编号：0、1、2、3…。
目的地为以下地区：北京、上海、天津…，依次编号：0、1、2…。
t时刻，北京-上海客流需求量为50，北京-天津客流需求量为200，上海-北京客流需求量为60，上海-天津客流需求量为7，天津-北京客流需求量为300，天津-上海客流需求量为8，重庆-北京客流需求量为15，重庆-上海客流需求量为2，重庆-天津客流需求量为18，则对应的OD二维图如下：
Figure PCTCN2020104902-appb-000003
步骤二、使用空间复杂关联卷积残差网络模型，提取所述周期OD二维图序列数据中的空间特征数据。
具体的，对所述周期OD二维图序列数据进行行堆叠变换，得到一维行堆叠OD数据。
示例性的，经过行堆叠变换后的数据为：
Figure PCTCN2020104902-appb-000004
其中，i是出发地编号，j是目的地编号；n、m为正整数，i为0-n之间整数，j为0-m之间的整数。
具体的，对所述周期OD二维图序列数据进行列堆叠变换，得到一维列堆叠OD数据。
示例性的，经过列堆叠变换后的输入数据为：
Figure PCTCN2020104902-appb-000005
其中，i是出发地编号，j是目的地编号；n、m为正整数，i为0-n之间整数，j为0-m之间的整数。
具体的，对所述一维行堆叠OD数据进行卷积残差处理，得到目的地关联空间特征信息。
具体的，所述卷积残差处理包括：对一维行堆叠OD数据卷积池化处理。
示例性的，一维行堆叠OD数据需要进行多次卷积池化处理。
利用第一卷积池化函数，对一维行堆叠OD数据进行一次卷积池化处理，得到目的地关联空间特征信息1；再次利用第一卷积池化函数，对输出的所述目的地关联空间特征信息1进行卷积池化处理，得到目的地关联空间特征信息2……以此类推，第n次利用第一卷积池化函数，对输出的目的地关联空间特征信息n-1进行卷积池化处理，得到目的地关联空间特征信息n。这样就完成对一维行堆叠OD数据的多次卷积池化处理。
具体的，所述第一卷积池化函数表述为：
X o(k)＝pool(f(w o(k)*X o(k-1)+b o(k)))
式(1)中，pool是池化层，f是激活函数，w o(k)是第k层网络单元需要学习的权重，b o(k)是第k层网络单元需要学习的偏置。
示例性的，pool池化层可使用且不限于下列方式：Average pool 池化或Max pool池化；激活函数f可使用且不限于下列激活函数：Relu激活函数或Tanh激活函数。
具体的，对所述一维列堆叠OD数据进行卷积残差处理，得到出发地关联空间特征信息。
具体的，所述卷积残差处理包括：对一维列堆叠OD数据卷积池化处理。
示例性的，一维列堆叠OD数据需要进行多次卷积池化处理。
利用第二卷积池化函数，对一维列堆叠OD数据进行一次卷积池化处理，得到出发地关联空间特征信息1；再次利用第二卷积池化函数，对输出的所述出发地关联空间特征信息1进行卷积池化处理，得到出发地关联空间特征信息2……以此类推，第n次利用第二卷积池化函数，对输出的出发地关联空间特征信息n-1进行卷积池化处理，得到出发地关联空间特征信息n。这样就完成对一维列堆叠OD数据的多次卷积池化处理。
具体的，所述第二卷积池化函数表述为：
X d(k)＝pool(f(w d(k)*X d(k-1)+b d(k)))·
其中，pool是池化层，f是激活函数，w d(k)和b d(k)分别是第k层网络单元需要学习的权重和偏置。
示例性的，pool池化层可使用且不限于下列方式：Average pool池化或Max pool池化；激活函数f可使用且不限于下列激活函数：Relu激活函数或Tanh激活函数。
具体的，对所述周期OD二维图序列数据进行卷积残差处理，得到混合关联空间特征信息。
具体的，所述卷积残差处理包括：对所述周期OD二维图序列数据卷积池化处理。
示例性的，周期OD二维图序列数据需要进行多次卷积池化处理。
利用第三卷积池化函数，对周期OD二维图序列数据进行一次卷积池化处理，得到混合关联空间特征信息1；再次利用第三卷积池化函数，对输出的所述混合关联空间特征信息1进行卷积池化处理，得到混合关联空间特征信息2……以此类推，第n次利用第三卷积池化函数，对输出的混合关联空间特征信息n-1进行卷积池化处理，得到混合关联空间特征信息n。这样就完成对周期OD二维图序列数据的多次卷积池化处理。
输入周期OD二维图序列数据为：
Figure PCTCN2020104902-appb-000006
其中，i是出发地编号，j是目的地编号；n、m为正整数，i为0-n之间整数，j为0-m之间的整数。
具体的，所述第三卷积池化函数表述为：
X conv2(k)＝pool(f(w conv2(k)*X conv2(k-1)+b conv2(k)))·
其中，pool是池化层，f是激活函数，w conv2(k)和b conv2(k)分别是第k层网络单元需要学习的权重和偏置。
示例性的，pool池化层可使用且不限于下列方式：Average pool池化或Max pool池化；激活函数f可使用且不限于下列激活函数：Relu激活函数或Tanh激活函数。
具体的，对所述目的地关联空间特征信息、出发地关联空间特征信息、混合关联空间特征信息进行连接处理，得到空间特征数据。
示例性的，输出的空间特征数据表示为：
X＝[X t1 … X ti … X tn]
其中，X ti表示空间关联卷积残差网络输出的ti时刻的空间特征数据。
进一步地，使用提取时间特征模型，提取所述空间特征数据中的时间特征数据。
示例性的，所述空间特征是多重周期下连续时间点的空间特征。
提取时间特征模型可以采用但不限于下列模型：长短期记忆神经网络模型(LSTM)。
长短期记忆神经网络能够学习序列长期的时间依赖性，将空间特征数据X输入长短期记忆神经网络，输出时间特征数据。
进一步地，将所述时间特征数据进行特征提取步骤为：
将所述时间特征数据融合，得到时间特征融合数据；将所述时间特征融合数据进行特征提取，得到想要预测时刻的OD客流预测值。
进一步地，预测方法还包括：利用平均绝对误差和均方误差对基于深度学习的轨道交通客流需求预测值进行评估。
示例性的，平均绝对误差(MAE)为：
Figure PCTCN2020104902-appb-000007
其中，x i和
Figure PCTCN2020104902-appb-000008
分别表示真实值和预测值，n表示预测值个数。
均方误差(MSE)为：
Figure PCTCN2020104902-appb-000009
其中，x i和
Figure PCTCN2020104902-appb-000010
分别表示真实值和预测值，n表示预测值个数。
当MAE≤第一阈值时且MSE≤第二阈值时，则所述预测方法的预测精度满足要求。
示例性的：第一阈值第二阈值为预设值，可根据实际进行修改。
具体的，一种基于深度学习的轨道交通客流需求预测装置，包括：
数据输入模块，用于采集OD数据，将所述OD数据转化为周期OD二维图序列数，并输出所述周期OD二维图序列数据；
空间特征信息提取模块，所述空间特征信息提取模块使用空间复杂关联卷积残差网络模型，用于提取所述周期OD二维图序列数据中 的空间特征数据；
多重周期模式下时间信息提取模块，用于提取所述空间特征数据中的时间特征数据；
数据输出模块，对所述时间特征数据进行特征提取，得到预测时刻的OD客流预测值。
示例性的，所述预测装置可以采用但不限于下列模型：一种基于深度学习的轨道交通客流需求预测装置模型，如图2所示。
示例性的，模型包括四层：数据输入层、空间特征信息提取层、多重周期模式下时间信息提取层、数据输出层。
数据输入层，采集OD数据，输出四种周期OD二维图序列数据。
空间特征信息提取层使用空间复杂关联卷积残差网络模型，接收所述数据输入层传输来的周期OD二维图序列数据，输出空间特征数据。其中空间复杂关联卷积残差网络模型包含若干组空间关联卷积残差网络(CNNRES_net)。
多重周期模式下时间信息提取层使用长短期记忆网络模型，接收所述空间特征信息提取层传输来空间特征数据，输出时间特征数据。
数据输出层包括连接层(Concatenate)和密集层(Dense)，接收所述多重周期模式下时间信息提取层传输来时间特征数据，输出预测时刻的OD客流预测值。
具体的，空间复杂关联卷积残差网络模型包括若干组空间关联卷积残差网络。所述空间关联卷积残差网络包括数据变换模块、卷积残差模块和空间特征信息连接模块。
示例性的，空间关联卷积残差网络采用但不限于下列结构，如图3所示。
空间关联卷积残差网络包括行堆叠变换模块、列堆叠变换模块、三个卷积残差单元和空间特征信息连接模块。三个卷积残差单元分别为第一卷积残差单元、第二卷积残差单元和第三卷积残差单元。
行堆叠变换单元设置在第一卷积残差单元的输入端，所述第一卷积残差单元用于对目的地关联空间特征信息提取。
列堆叠变换单元设置在第二卷积残差单元的输入端，所述第二卷积残差单元用于对出发地关联空间特征信息提取。
周期OD二维图序列数据直接输入到第三卷积残差单元，所述第三卷积残差单元用于对混合关联空间特征信息提取。
第一卷积残差单元、第二卷积残差单元和第三卷积残差单元的输出端连接于空间特征信息连接模块输入端，空间特征信息连接模块用于对第一卷积残差单元、第二卷积残差单元和第三卷积残差单元输出数据融合，然后输出空间特征数据。
具体的，所述第一卷积残差单元包括第一卷积神经单元和残差连接；
所述第二卷积残差单元包括第二卷积神经单元和残差连接；
所述第三卷积残差单元包括第三卷积神经单元和残差连接。
示例性的，第一、第二、第三卷积残差单元可以采用但不限于下列结构，如图4所示，图中为多层卷积神经网络单元叠加结构上加入一个跨越多层到输出层(或接近输出层的其它层)的残差连接。数据由输入端输入至第一层卷积神经网络单元，卷积池化处理后，传入至第二层卷积神经网络单元；卷积池化处理后，传入至第三层卷积神经网络单元……经n次卷积池化处理后输出；同时，从输入层加上一条直接跨越多层到输出层(或接近输出层的其它层)的残差连接，以防止梯度信息消失。经残差处理后的数据再次传入第n+1层卷积神经网络单元，卷积池化处理后输出。
OD数据在通过多层卷积神经网络单元进行信息处理时，会造成的信息流失；残差连接用于防止梯度信息消失。
具体的，所述第一卷积神经单元对一维行堆叠OD数据进行卷积池化处理，为一维卷积单元(Conv1Unit)；
所述第二卷积神经单元对一维列堆叠OD数据进行卷积池化处理，为一维卷积单元；
所述第三卷积神经单元对周期OD二维图序列数据进行卷积池化处理，为二维卷积单元(Conv2Unit)。
示例性的，一维卷积单元结构如图5所示，包括一层一维卷积层(Conv1)和一层一维最大池化层(maxpooling1D)。
示例性的，二维卷积单元结构如图6所示，包括一层二维卷积层(Conv2)和一层二维最大池化层(maxpooling2D)。
具体的，第一卷积神经单元卷积池化处理方式为：
X o(k)＝pool(f(w o(k)*X o(k-1)+b o(k)))
其中，pool是池化层，f是激活函数，w o(k)和b o(k)分别是第k层网络单元需要学习的权重和偏置。
示例性的，pool池化层可使用且不限于下列方式：Average pool池化或Max pool池化；激活函数f可使用且不限于下列激活函数：Relu激活函数或Tanh激活函数。
所述第二卷积神经单元卷积池化处理方式为：
X d(k)＝pool(f(w d(k)*X d(k-1)+b d(k)))·
其中，pool是池化层，f是激活函数，w d(k)和b d(k)分别是第k层网络单元需要学习的权重和偏置。
示例性的，pool池化层可使用且不限于下列方式：Average pool池化或Max pool池化；激活函数f可使用且不限于下列激活函数：Relu激活函数或Tanh激活函数。
所述第三卷积神经单元卷积池化处理方式为：
X conv2(k)＝pool(f(w conv2(k)*X conv2(k-1)+b conv2(k)))·
其中，pool是池化层，f是激活函数，w conv2(k)和b conv2(k)分别是第k层网络单元需要学习的权重和偏置。
示例性的，pool池化层可使用且不限于下列方式：Average pool 池化或Max pool池化；激活函数f可使用且不限于下列激活函数：Relu激活函数或Tanh激活函数。
进一步地，所述多重周期模式下时间信息提取层，接收所述空间特征信息提取层传输来空间特征数据，输出时间特征数据。所述空间特征信息提取层为提取时间特征模型。
示例性的，提取时间特征模型采用但不限于下列模型：长短期记忆神经网络模型。
多重周期模式下的时间特征信息提取是对不同周期输入信息构建一个时间特征信息提取单元，每个提取单元由一系列长短期记忆神经网络层组成。
具体的，所述数据输出层，接收所述多重周期模式下时间信息提取层传输来时间特征数据，输出预测时刻的OD客流预测值。所述数据输出层包括连接层和密集层。
示例性的，连接层设置在密集层的输入端，用于将所述时间特征数据融合。
密集层，用于对所述融合时间特征进行特征提取，得到预测时刻的OD客流预测值。
进一步地，所述预测装置还包括评估装置。
所述评估装置用于利用平均绝对误差和均方误差对基于深度学习的轨道交通客流需求预测方法和装置进行评估。
示例性的，评估装置采用但不限于下列装置：平均绝对误差装置和均方误差装置两种装置；也可以采用但不限于下列装置：同时包含平均绝对误差评估功能和均方误差评估功能的装置。
尽管参照前述实施例对本公开进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替 换，并不使相应技术方案的本质脱离本公开各实施例技术方案的精神和范围。

Claims (16)

  1. 一种基于深度学习的轨道交通客流需求预测方法，其特征在于，所述预测方法包括：
    采集OD数据，将所述OD数据转化为周期OD二维图序列数据；
    使用空间复杂关联卷积残差网络模型，提取所述周期OD二维图序列数据中的空间特征数据；
    使用提取时间特征模型，提取所述空间特征数据中的时间特征数据；
    对所述时间特征数据进行特征提取处理，得到预测时刻的OD客流预测值。
  2. 根据权利要求1所述的基于深度学习的轨道交通客流需求预测方法，其特征在于，
    所述提取所述周期OD二维图序列数据中的空间特征数据包括：
    对所述周期OD二维图序列数据进行行堆叠变换，得到一维行堆叠OD数据；
    对所述周期OD二维图序列数据进行列堆叠变换，得到一维列堆叠OD数据；
    对所述一维行堆叠OD数据进行卷积残差处理，得到目的地关联空间特征信息；
    对所述一维列堆叠OD数据进行卷积残差处理，得到出发地关联空间特征信息；
    对所述周期OD二维图序列数据进行卷积残差处理，得到混合关联空间特征信息；
    对所述目的地关联空间特征信息、出发地关联空间特征信息、混合关联空间特征信息进行连接处理，得到空间特征数据。
  3. 根据权利要求2所述的基于深度学习的轨道交通客流需求预测方法，其特征在于，
    对数据进行所述卷积残差处理包括：
    对所述一维行堆叠OD数据卷积池化处理；
    对所述一维列堆叠OD数据卷积池化处理；
    对所述周期OD二维图序列数据卷积池化处理。
  4. 根据权利要求3所述的基于深度学习的轨道交通客流需求预测方法，其特征在于，
    所述一维行堆叠OD数据使用第一卷积池化函数进行卷积池化处理；
    所述第一卷积池化函数表述为：
    X o(k)＝pool(f(w o(k)*X o(k-1)+b o(k)))   (1)
    式(1)中，pool是池化层，f是激活函数，w o(k)是第k层网络单元需要学习的权重，b o(k)是第k层网络单元需要学习的偏置；
    所述一维列堆叠OD数据使用第二卷积池化函数进行卷积池化处理；
    所述第二卷积池化函数表述为：
    X d(k)＝pool(f(w d(k)*X d(k-1)+b d(k)))   (2)
    式(2)中，pool是池化层，f是激活函数，w d(k)是第k层网络单元需要学习的权重，b d(k)是第k层网络单元需要学习的偏置；
    所述周期OD二维图序列数据使用第三卷积池化函数进行卷积池化处理；
    所述第三卷积池化函数表述为：
    X conv2(k)＝pool(f(w conv2(k)*X conv2(k-1)+b conv2(k)))(3)
    式(3)中，pool是池化层，f是激活函数，w conv2(k)是第k层网络单元需要学习的权重，b conv2(k)是第k层网络单元需要学习的偏 置。
  5. 根据权利要求1所述的基于深度学习的轨道交通客流需求预测方法，其特征在于，
    对所述时间特征数据进行特征提取步骤为：
    对所述时间特征数据进行融合处理，得到时间特征融合数据；
    对所述时间特征融合数据进行特征提取，得到OD客流预测值。
  6. 根据权利要求1所述的基于深度学习的轨道交通客流需求预测方法，其特征在于，
    所述预测方法还包括：
    利用平均绝对误差和均方误差对所述OD客流预测值进行评估；
    获取真实值与预测值的平均绝对误差和均方误差，在所述平均绝对误差小于等于第一阈值时且所述均方误差小于等于第二阈值时，所述预测方法的预测精度满足要求。
  7. 一种基于深度学习的轨道交通客流需求预测装置，其特征在于，所述预测装置包括：
    数据输入模块，用于采集OD数据，将所述OD数据转化为周期OD二维图序列数，并输出所述周期OD二维图序列数据；
    空间特征信息提取模块，所述空间特征信息提取模块使用空间复杂关联卷积残差网络模型，用于提取所述周期OD二维图序列数据中的空间特征数据；
    多重周期模式下时间信息提取模块，用于提取所述空间特征数据中的时间特征数据；
    数据输出模块，对所述时间特征数据进行特征提取，得到预测时刻的OD客流预测值。
  8. 根据权利要求7所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述空间复杂关联卷积残差网络模型包括若干组空间关联卷积 残差网络，所述空间关联卷积残差网络用于提取周期OD二维图序列数据的空间关联特征数据。
  9. 根据权利要求8所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述空间关联卷积残差网络包括数据变换模块、卷积残差模块和空间特征信息连接模块；
    所述数据变换模块包括行堆叠变换模块和列堆叠变换模块；
    所述行堆叠变换模块，接收所述数据输入层传输来的周期OD二维图序列数据，对所述周期OD二维图序列数据进行行堆叠变换，输出一维行堆叠OD数据；
    所述列堆叠变换模块，接收所述数据输入层传输来的周期OD二维图序列数据，对所述周期OD二维图序列数据进行列堆叠变换，输出一维列堆叠OD数据；
    所述卷积残差模块包括第一卷积残差单元、第二卷积残差单元和第三卷积残差单元；
    所述第一卷积残差单元，接收所述行堆叠变换模块传输来的一维行堆叠OD数据，对所述一维行堆叠OD数据进行卷积残差处理，输出目的地关联空间特征信息；
    所述第二卷积残差单元，接收所述列堆叠变换模块传输来的一维列堆叠OD数据，对所述一维列堆叠OD数据进行卷积残差处理，输出出发地关联空间特征信息；
    所述第三卷积残差单元，接收所述数据输入层传输来的周期OD二维图序列数据，对所述周期OD二维图序列数据进行卷积残差处理，输出混合关联空间特征信息；
    所述空间特征信息连接模块，接收所述第一卷积残差单元、所述第二卷积残差单元和所述第三卷积残差单元传输来的空间特征信息，用于对所述目的地关联空间特征信息、出发地关联空间特征信息、混 合关联空间特征信息进行连接处理。
  10. 根据权利要求9所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述第一卷积残差单元包括第一卷积神经单元和残差连接；
    所述第二卷积残差单元包括第二卷积神经单元和残差连接；
    所述第三卷积残差单元包括第三卷积神经单元和残差连接；
    所述第一卷积神经单元对一维行堆叠OD数据进行卷积池化处理，为一维卷积单元；
    所述第二卷积神经单元对一维列堆叠OD数据进行卷积池化处理，为一维卷积单元；
    所述第三卷积神经单元对周期OD二维图序列数据进行卷积池化处理，为二维卷积单元。
  11. 根据权利要求10所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述第一卷积神经单元卷积池化处理方式为：
    X o(k)＝pool(f(w o(k)*X o(k-1)+b o(k)))
    其中，pool是池化层，f是激活函数，w o(k)是第k层网络单元需要学习的权重，b o(k)是第k层网络单元需要学习的偏置；
    所述第二卷积神经单元卷积池化处理方式为：
    X d(k)＝pool(f(w d(k)*X d(k-1)+b d(k)))
    其中，pool是池化层，f是激活函数，w d(k)是第k层网络单元需要学习的权重，b d(k)是第k层网络单元需要学习的偏置；
    所述第三卷积神经单元卷积池化处理方式为：
    X conv2(k)＝pool(f(w conv2(k)*X conv2(k-1)+b conv2(k)))
    其中，pool是池化层，f是激活函数，w conv2(k)是第k层网络单元需要学习的权重，b conv2(k)是第k层网络单元需要学习的偏置。
  12. 根据权利要求7所述的基于深度学习的轨道交通客流需求预 测装置，其特征在于，
    所述多重周期模式下时间信息提取模块，接收所述空间特征信息提取模块传输来空间特征数据，输出时间特征数据；
    所述空间特征信息提取层为提取时间特征模型。
  13. 根据权利要求12所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述提取时间特征模型为长短期记忆神经网络模型。
  14. 根据权利要求7所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述数据输出模块，接收所述多重周期模式下时间信息提取模块传输来时间特征数据，输出预测时刻的OD客流预测值。
  15. 根据权利要求14所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述数据输出模块包括连接模块和密集模块；
    所述连接模块，接收所述多重周期模式下时间信息提取模块传输来时间特征数据，将所述时间特征数据融合，输出融合时间特征；
    所述密集模块，接收所述连接模块传输来融合时间特征，对所述融合时间特征进行特征提取，输出预测时刻的OD客流预测值。
  16. 根据权利要求7所述的基于深度学习的轨道交通客流需求预测装置，其特征在于，
    所述预测装置还包括评估装置；
    所述评估装置用于利用平均绝对误差和均方误差对基于深度学习的轨道交通客流需求预测方法和装置预测精度进行评估。
PCT/CN2020/104902 2020-03-02 2020-07-27 一种基于深度学习的轨道交通客流需求预测方法和装置 WO2021174755A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP20923491.3A EP4099237A4 (en) 2020-03-02 2020-07-27 DEEP LEARNING RAILWAY PASSENGER FLOW DEMAND PREDICTION METHOD AND APPARATUS

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010133886.2 2020-03-02
CN202010133886.2A CN110991775B (zh) 2020-03-02 2020-03-02 一种基于深度学习的轨道交通客流需求预测方法和装置

Publications (1)

Publication Number Publication Date
WO2021174755A1 true WO2021174755A1 (zh) 2021-09-10

Family

ID=70081312

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/104902 WO2021174755A1 (zh) 2020-03-02 2020-07-27 一种基于深度学习的轨道交通客流需求预测方法和装置

Country Status (3)

Country Link
EP (1) EP4099237A4 (zh)
CN (1) CN110991775B (zh)
WO (1) WO2021174755A1 (zh)

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114037160A (zh) * 2021-11-10 2022-02-11 西南交通大学 SEM-Logit旅游铁路客流预测模型的构建方法
CN114141020A (zh) * 2022-02-08 2022-03-04 北京城建设计发展集团股份有限公司 一种基于轨道交通的行人仿真方法、系统及电子设备
CN114282658A (zh) * 2022-03-02 2022-04-05 北京航空航天大学 一种针对流量序列分析及预测的方法、设备和介质
CN114326391A (zh) * 2021-12-13 2022-04-12 哈尔滨工程大学 一种建筑能耗预测方法
CN114331234A (zh) * 2022-03-16 2022-04-12 北京交通大学 基于乘客出行信息的轨道交通客流预测方法和系统
CN114422382A (zh) * 2022-01-28 2022-04-29 湖南大学 网络流量预测方法、计算机装置、产品及存储介质
CN114492886A (zh) * 2022-01-25 2022-05-13 四川大学 一种考虑出发地和目的地的网约车需求预测方法及系统
CN114511112A (zh) * 2022-01-24 2022-05-17 北京通建泰利特智能系统工程技术有限公司 一种基于物联网的智慧运维方法、系统和可读存储介质
CN114912683A (zh) * 2022-05-13 2022-08-16 中铁第六勘察设计院集团有限公司 一种智慧城市轨道交通异常大客流预测系统及方法
CN115564151A (zh) * 2022-12-06 2023-01-03 成都智元汇信息技术股份有限公司 一种基于形态识别的突发大客流形态识别方法及系统
CN116306794A (zh) * 2023-03-09 2023-06-23 河源市广师大研究院 作物需水量预测方法、装置、电子设备及存储介质
CN116542438A (zh) * 2023-03-28 2023-08-04 大连海事大学 一种基于无基准真相的公交乘客起止点估计修补方法
CN116630900A (zh) * 2023-07-21 2023-08-22 中铁第四勘察设计院集团有限公司 基于机器学习的客运站旅客流线识别方法、系统及设备
CN117236492A (zh) * 2023-09-06 2023-12-15 西南交通大学 基于动态多尺度图学习的交通需求预测方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110991775B (zh) * 2020-03-02 2020-06-26 北京全路通信信号研究设计院集团有限公司 一种基于深度学习的轨道交通客流需求预测方法和装置
CN111242395B (zh) * 2020-04-26 2020-07-31 北京全路通信信号研究设计院集团有限公司 用于od数据的预测模型构建方法及装置
CN111242292B (zh) * 2020-04-26 2020-09-08 北京全路通信信号研究设计院集团有限公司 基于深度时空网络的od数据预测方法及系统
CN112598182B (zh) * 2020-12-25 2022-11-29 卡斯柯信号有限公司 一种轨道交通智能调度方法及系统
CN114819366A (zh) * 2022-05-06 2022-07-29 华侨大学 一种od客流短时预测方法、装置、设备及其存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170249534A1 (en) * 2016-02-29 2017-08-31 Fujitsu Limited Method and apparatus for generating time series data sets for predictive analysis
CN107529651A (zh) * 2017-08-18 2018-01-02 北京航空航天大学 一种基于深度学习的城市交通客流预测方法和设备
CN108985475A (zh) * 2018-06-13 2018-12-11 厦门大学 基于深度神经网络的网约车召车需求预测方法
CN109492814A (zh) * 2018-11-15 2019-03-19 中国科学院深圳先进技术研究院 一种城市交通流量预测方法、系统及电子设备
CN110991775A (zh) * 2020-03-02 2020-04-10 北京全路通信信号研究设计院集团有限公司 一种基于深度学习的轨道交通客流需求预测方法和装置

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102169524A (zh) * 2010-02-26 2011-08-31 同济大学 城市轨道交通网络客流分布的分阶段多路径模型算法
US8798897B2 (en) * 2010-11-01 2014-08-05 International Business Machines Corporation Real-time traffic analysis through integration of road traffic prediction and traffic microsimulation models
CN110651310B (zh) * 2017-04-05 2023-09-08 卡内基梅隆大学 估计对象密度和/或流量的深度学习方法及相关方法和软件
US10733877B2 (en) * 2017-11-30 2020-08-04 Volkswagen Ag System and method for predicting and maximizing traffic flow
CN109034449A (zh) * 2018-06-14 2018-12-18 华南理工大学 基于深度学习和乘客行为模式的短期公交客流预测方法
CN109887272B (zh) * 2018-12-26 2021-08-13 创新先进技术有限公司 一种交通人流量的预测方法及装置
CN109711640A (zh) * 2019-01-23 2019-05-03 北京工业大学 一种基于模糊c均值交通流量聚类以及误差反馈卷积神经网络的短时交通流预测方法
CN110264709B (zh) * 2019-05-06 2021-02-12 北京交通大学 基于图卷积网络的道路的交通流量的预测方法
CN110619419B (zh) * 2019-07-30 2022-08-05 北京交通大学 城市轨道交通的客流预测方法
CN110674987A (zh) * 2019-09-23 2020-01-10 北京顺智信科技有限公司 一种交通流量预测系统、方法及模型训练方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170249534A1 (en) * 2016-02-29 2017-08-31 Fujitsu Limited Method and apparatus for generating time series data sets for predictive analysis
CN107529651A (zh) * 2017-08-18 2018-01-02 北京航空航天大学 一种基于深度学习的城市交通客流预测方法和设备
CN108985475A (zh) * 2018-06-13 2018-12-11 厦门大学 基于深度神经网络的网约车召车需求预测方法
CN109492814A (zh) * 2018-11-15 2019-03-19 中国科学院深圳先进技术研究院 一种城市交通流量预测方法、系统及电子设备
CN110991775A (zh) * 2020-03-02 2020-04-10 北京全路通信信号研究设计院集团有限公司 一种基于深度学习的轨道交通客流需求预测方法和装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4099237A4 *

Cited By (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114037160B (zh) * 2021-11-10 2023-04-18 西南交通大学 SEM-Logit旅游铁路客流预测模型的构建方法
CN114037160A (zh) * 2021-11-10 2022-02-11 西南交通大学 SEM-Logit旅游铁路客流预测模型的构建方法
CN114326391A (zh) * 2021-12-13 2022-04-12 哈尔滨工程大学 一种建筑能耗预测方法
CN114511112A (zh) * 2022-01-24 2022-05-17 北京通建泰利特智能系统工程技术有限公司 一种基于物联网的智慧运维方法、系统和可读存储介质
CN114492886A (zh) * 2022-01-25 2022-05-13 四川大学 一种考虑出发地和目的地的网约车需求预测方法及系统
CN114422382B (zh) * 2022-01-28 2023-02-28 湖南大学 网络流量预测方法、计算机装置、产品及存储介质
CN114422382A (zh) * 2022-01-28 2022-04-29 湖南大学 网络流量预测方法、计算机装置、产品及存储介质
CN114141020B (zh) * 2022-02-08 2022-09-16 北京城建设计发展集团股份有限公司 一种基于轨道交通的行人仿真方法、系统及电子设备
CN114141020A (zh) * 2022-02-08 2022-03-04 北京城建设计发展集团股份有限公司 一种基于轨道交通的行人仿真方法、系统及电子设备
CN114282658A (zh) * 2022-03-02 2022-04-05 北京航空航天大学 一种针对流量序列分析及预测的方法、设备和介质
CN114331234A (zh) * 2022-03-16 2022-04-12 北京交通大学 基于乘客出行信息的轨道交通客流预测方法和系统
CN114331234B (zh) * 2022-03-16 2022-07-12 北京交通大学 基于乘客出行信息的轨道交通客流预测方法和系统
CN114912683B (zh) * 2022-05-13 2024-05-10 中铁第六勘察设计院集团有限公司 一种智慧城市轨道交通异常大客流预测系统及方法
CN114912683A (zh) * 2022-05-13 2022-08-16 中铁第六勘察设计院集团有限公司 一种智慧城市轨道交通异常大客流预测系统及方法
CN115564151A (zh) * 2022-12-06 2023-01-03 成都智元汇信息技术股份有限公司 一种基于形态识别的突发大客流形态识别方法及系统
CN116306794A (zh) * 2023-03-09 2023-06-23 河源市广师大研究院 作物需水量预测方法、装置、电子设备及存储介质
CN116306794B (zh) * 2023-03-09 2023-10-31 河源市广师大研究院 作物需水量预测方法、装置、电子设备及存储介质
CN116542438B (zh) * 2023-03-28 2024-01-30 大连海事大学 一种基于无基准真相的公交乘客起止点估计修补方法
CN116542438A (zh) * 2023-03-28 2023-08-04 大连海事大学 一种基于无基准真相的公交乘客起止点估计修补方法
CN116630900A (zh) * 2023-07-21 2023-08-22 中铁第四勘察设计院集团有限公司 基于机器学习的客运站旅客流线识别方法、系统及设备
CN116630900B (zh) * 2023-07-21 2023-11-07 中铁第四勘察设计院集团有限公司 基于机器学习的客运站旅客流线识别方法、系统及设备
CN117236492A (zh) * 2023-09-06 2023-12-15 西南交通大学 基于动态多尺度图学习的交通需求预测方法
CN117236492B (zh) * 2023-09-06 2024-03-12 西南交通大学 基于动态多尺度图学习的交通需求预测方法

Also Published As

Publication number Publication date
CN110991775A (zh) 2020-04-10
CN110991775B (zh) 2020-06-26
EP4099237A1 (en) 2022-12-07
EP4099237A4 (en) 2023-07-05

Similar Documents

Publication Publication Date Title
WO2021174755A1 (zh) 一种基于深度学习的轨道交通客流需求预测方法和装置
CN112767682A (zh) 一种基于图卷积神经网络的多尺度交通流量预测方法
CN113077090A (zh) 客流预测方法、系统及计算机可读存储介质
Liu et al. Fedgru: Privacy-preserving traffic flow prediction via federated learning
CN114495500B (zh) 一种基于对偶动态时空图卷积的交通预测方法
CN114757389A (zh) 一种基于联邦学习的城市交通流量时空预测方法
Zhang et al. Travel time prediction of urban public transportation based on detection of single routes
CN115017809A (zh) 一种船舶到港量预测系统及方法
Xiong et al. Statistical estimation in passenger‐to‐train assignment models based on automated data
Zhang et al. Link speed prediction for signalized urban traffic network using a hybrid deep learning approach
CN116631181A (zh) 一种基于天气感知图形注意力网络的交通预测方法
CN115965163A (zh) 基于时空图生成对抗损失的轨道交通短时客流预测方法
Han et al. Continuity of the path delay operator for LWR-based network loading with spillback
CN112488422B (zh) 一种基于多任务学习的多模式出行需求预测方法
CN114662304A (zh) 一种城市道路交通车流量的动态事件触发滤波估计方法
Du et al. Stochastic synchronization between different networks and its application in Bilayer coupled public traffic network
CN117196086A (zh) 一种基于机器学习的地铁短时客流预测方法
Chen et al. Dynamic graph hybrid system: A modeling method for complex networks with application to urban traffic
CN111507902A (zh) 一种高分辨率图像获取方法及装置
Gu et al. Daily OD Demand Prediction in Urban Metro Transit System: A Convolutional LSTM Neural Network with Multi-factor Fusion Channel-wise Attention
Han et al. Research on Travel Time Prediction of Multiple Bus Trips Based on MDARNN
Li et al. MCINet: Multi-Channel Informer Networks for Long Sequence Time-Series Forecasting in Ocean-Going Ship Satellite Communication Networks
CN113077281B (zh) 一种地铁客流量分布预测方法以及设备
Zhang et al. Short-Term Speed Forecasting of Large-Scale Urban Road Network Based on Transformer
Zhang Gcn-trn: Efficient transformer based electric vehicle charging demand forecasting system

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20923491

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020923491

Country of ref document: EP

Effective date: 20220829

NENP Non-entry into the national phase

Ref country code: DE