KR102678725B1 - Apparatus and method for predicting congestion by subway car using deep learning - Google Patents

Abstract

The method for predicting the congestion level of each subway car of the present invention includes the steps of a data processing unit receiving a message requesting the congestion level of a subway station and subway car from a user device, and the data processing unit receiving the message at a reference point in time. A step of selecting a plurality of dispatches from the past and generating an input vector string including a plurality of input vectors corresponding to each of the selected dispatches, and a prediction unit performing an operation on the plurality of input vectors through a prediction network to determine the standard It includes the step of deriving the congestion level of the subway car at the subway station when the next train is dispatched a predetermined number of times from the time point.

Description

Apparatus and method for predicting congestion by subway car using deep learning}

The present invention relates to congestion prediction technology, and more specifically, to a device and method for predicting congestion for each subway car using deep learning.

The subway consists of several connected rooms. Additionally, instead of using a boarding and disembarking tag when boarding a train, you use a boarding and disembarking tag at the ticket gate when entering or exiting a subway station. This is a major feature that distinguishes it from buses, and due to these structural characteristics, subways have different density distributions for each room. Additionally, in the same context, there are limitations in methods for identifying or predicting congestion for each room.

Recently, attempts to estimate subway congestion are continuing. However, estimating the level of congestion at a given point in time has limitations in terms of service. Potential subway users in the future cannot determine the congestion information in advance and select a boarding compartment, and can only accept the congestion information when they arrive at the subway platform and when the train just before they are to board arrives. As such, there is a lack of appropriate methodology to predict congestion at future times.

Korean Patent Publication No. 2015-0018975 (published on February 25, 2015)

The purpose of the present invention is to provide an apparatus and method for predicting congestion for each subway car using deep learning.

A method for predicting the congestion level of each subway car according to a preferred embodiment of the present invention to achieve the above-described object includes the steps of a data processing unit receiving a message requesting the congestion level of a subway station and subway car from a user device; The data processing unit selects a plurality of dispatches from the past of the reference point in time when the message is received, and generates an input vector string including a plurality of input vectors corresponding to each selected dispatch, and the prediction unit creates a prediction network. and deriving the congestion level of the subway car in the subway station when the next train is dispatched a predetermined number of times from the reference point by performing an operation on the plurality of input vectors.

The step of deriving the congestion of the subway station and the subway car includes the step of receiving, by the input layer of the prediction network, a plurality of input vectors arranged in time order corresponding to the plurality of stages of the prediction network for each of the plurality of stages, A plurality of hidden cells in the hidden layer of the prediction network perform a weighted operation on the state value of the previous stage and the input vector, which is the input value of the current stage, to calculate the state value of the current stage, and then send the calculated state value to the next stage. delivered to the stage, wherein the last hidden cell among the plurality of hidden cells calculates the state value of the current stage and then applies a weight to the calculated state value of the current stage to transmit the subway to the subway station at the time of the next dispatch of the predetermined number of times. It includes calculating an output vector representing the congestion level of the box.

The method includes, before receiving the message requesting the congestion, a learning unit preparing learning data including a learning input vector string containing a plurality of learning input vectors and a learning ground truth vector that is a label corresponding to the learning input vector string. A prediction network with untrained weights performs an operation on the plurality of input vectors for learning to calculate an output vector, and a loss that is the difference between the output vector and the actual learning vector through a loss function. It further includes calculating and performing optimization to modify the weights of the prediction network so that the loss is minimized.

The step of generating an input vector string including the plurality of input vectors includes the data processing unit selecting a dispatch consisting of a predetermined multiple interval of the dispatch interval of the subway station from a past time based on the reference point, and the data A processing unit collecting event information that quantifies the dispatch time corresponding to the selected dispatch time, the congestion level of the subway car corresponding to the dispatch time, weather information indicating the weather situation, and events contributing to whether or not people are crowded, the data; A processing unit includes a step of generating an input vector string by sorting a plurality of input vectors, each of which includes the dispatch time, the congestion level of the subway car, the weather information, and the event information, in order of dispatch time.

Here, the congestion level of the subway car is characterized in that it is collected through the weight of the subway car.

In order to achieve the above-described object, a device for predicting congestion for each subway car according to a preferred embodiment of the present invention receives a message requesting the congestion of subway stations and subway cars from a user device, and receives the message. A data processing unit that selects a plurality of dispatches from the past of the reference point in time and generates an input vector string containing a plurality of input vectors corresponding to each of the selected dispatches, and performs calculations on the plurality of input vectors through a prediction network. It includes a prediction unit that performs the calculation to derive the congestion level of the subway car at the subway station when the next train is dispatched a predetermined number of times from the reference point.

The prediction network has an input layer that receives a plurality of input vectors arranged in time order corresponding to a plurality of stages for each of the plurality of stages, and weights are applied to the state value of the previous stage and the input vector that is the input value of the current stage. After calculating the state value of the current stage by performing an operation, it includes a plurality of hidden cells that transmit the calculated state value to the next stage, and the last hidden cell of the plurality of hidden cells calculates the state value of the current stage. Afterwards, a hidden layer applies a weight to the calculated state value of the current stage to calculate an output vector representing the congestion level of the subway car in the subway station at the time of the next dispatch of the predetermined number of times, and an output layer that outputs the output vector. do.

The device prepares learning data including a learning input vector string including a plurality of learning input vectors and a ground truth vector for learning, which is a label corresponding to the learning input vector string, and a prediction network with untrained weights is configured to use the plurality of learning input vectors. Calculating an output vector by performing an operation on an input vector for learning, calculating a loss that is the difference between the output vector and the actual measurement vector for learning through a loss function, and modifying the weights of the prediction network so that the loss is minimized. It further includes a learning unit that performs optimization.

The data processing unit selects a dispatch consisting of a predetermined multiple interval of the dispatch interval of the subway station from a past point in time based on the reference point, a dispatch time corresponding to the selected dispatch, and a congestion level of the subway car corresponding to the dispatch time, Collects weather information indicating weather conditions and event information that quantifies events that contribute to whether people are crowded, and a plurality of input vectors each including the dispatch time, congestion of the subway car, the weather information, and the event information It is characterized by generating an input vector string by sorting according to the order of dispatch time.

Here, the congestion level of the subway car is characterized in that it is collected through the weight of the subway car.

According to the present invention, it is possible to predict the level of congestion for each subway car according to the dispatch of a specific subway station. Accordingly, users can use the subway while avoiding crowded locations or times.

Figure 1 is a diagram illustrating a system for predicting congestion for each subway car using deep learning according to an embodiment of the present invention.
Figure 2 is a block diagram for explaining the configuration of a prediction server according to an embodiment of the present invention.
Figure 3 is a block diagram for explaining the detailed configuration of a prediction server according to an embodiment of the present invention.
Figure 4 is a diagram illustrating the configuration of a prediction network for predicting congestion according to an embodiment of the present invention.
Figure 5 is a diagram for explaining unit operations of a prediction network for predicting congestion according to an embodiment of the present invention.
Figure 6 is a flowchart illustrating a method of generating a prediction network for predicting the congestion level of a subway car through learning according to an embodiment of the present invention.
Figure 7 will explain a method for predicting congestion for each subway car using deep learning according to an embodiment of the present invention.

Prior to the detailed description of the present invention, the terms and words used in the specification and claims described below should not be construed as limited to their ordinary or dictionary meanings, and the inventor should use his/her invention in the best possible manner. In order to explain, it must be interpreted as meaning and concept consistent with the technical idea of the present invention based on the principle that the term can be appropriately defined as a concept. Therefore, the embodiments described in this specification and the configurations shown in the drawings are only the most preferred embodiments of the present invention, and do not represent the entire technical idea of the present invention, and therefore, various equivalents that can replace them at the time of filing the present application It should be understood that there may be variations and examples.

In the description and claims of embodiments of the invention, “network” is defined as one or more data links that enable the transfer of electronic data between computer systems and/or modules. When information is transmitted or provided to a computer system through a network or a communication connection of the same or heterogeneous type, that is, wired, wireless, or a combination of wired and wireless, this connection may be understood as a computer-readable medium. . Computer-readable instructions include, for example, instructions and data that cause a general-purpose or special-purpose computer system to perform a particular function or group of functions. Computer-executable instructions may be binary, intermediate format instructions, for example, assembly language, or even source code. Moreover, the invention is applicable to personal computers, laptop computers, handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, mobile phones, cell phones, and mobile devices. In a network computing environment with various types of computer system configurations including communication terminals, smartphones, PDAs, pagers, etc., it can be applied to the computer systems.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. At this time, it should be noted that in the attached drawings, identical components are indicated by identical symbols whenever possible. Additionally, detailed descriptions of well-known functions and configurations that may obscure the gist of the present invention will be omitted. For the same reason, in the accompanying drawings, some components are exaggerated, omitted, or schematically shown, and the size of each component does not entirely reflect the actual size.

First, a system for predicting congestion for each subway car using deep learning according to an embodiment of the present invention will be described. Figure 1 is a diagram illustrating a system for predicting congestion for each subway car using deep learning according to an embodiment of the present invention. Referring to Figure 1, a system for predicting congestion for each subway car using deep learning according to an embodiment of the present invention (hereinafter abbreviated as 'prediction system') includes a prediction server 10, a user device 20, and information. Includes a provision server (30).

The prediction server 10, the user device 20, and the information providing server 30 can communicate with each other through a network (NW). For example, the network (NW) may include wireless networks such as Wireless LAN (WLAN), Wi-Fi, Wibro, Wimax, and High Speed Downlink Packet Access (HSDPA), Depending on how the system is implemented, it may include wired networks such as Ethernet, xDSL (ADSL, VDSL), Hybrid Fiber Coaxial Cable (HFC), FTTC (Fiber to The Curb), and FTTH (Fiber To The Home). In addition, the network (NW) of the present invention may include, for example, a network consisting of a plurality of access networks and a core network connecting them. Additionally, the network (NW) according to the present invention may include the Internet. The Internet refers to an ordinary open network, that is, a public network, where information is exchanged according to the TCP/IP protocol. Through this network (NW), the prediction server 10, the user device 20, and the information provision server 30 are interconnected to form a prediction system according to the present invention.

The prediction server 10 is basically used to predict the congestion of subway cars according to the request of the user device 20. To this end, the prediction server 10 collects past data related to the corresponding subway car from the information providing server 30, and uses a prediction network (EN) learned using the collected data to predict the corresponding subway car. Predict congestion. This prediction server 10 is an entity that exists on a network and performs the roles of a web server, database server, and application server.

Although the information providing server 30 is shown as one, it may be a plurality of servers providing corresponding data. For example, the information providing server 30 may be a server operated by the Subway Corporation, a server operated by the Korea Meteorological Administration, or a server operated by the Korea Disease Control and Prevention Agency.

The user device 20 is a device used by the user. The user device 20 connects to the prediction server 10, specifies the subway station and subway car to the prediction server 10, and requests the congestion level, and the prediction server 10 It is possible to receive the predicted congestion level of subway cars distributed to the corresponding subway station. The user device 20 may be a representative example of a smartphone, but is not limited thereto and can be applied to various terminals such as all information and communication devices, multimedia terminals, wired terminals, fixed terminals, and IP (Internet Protocol) terminals. For example, the terminal is a mobile terminal with various mobile communication specifications such as a mobile phone, PMP (Portable Multimedia Player), MID (Mobile Internet Device), tablet PC, phablet PC, and information and communication devices, or a laptop, personal computer, etc. It may be a fixed terminal.

Next, we will look at the configuration of the prediction server 10 according to an embodiment of the present invention in more detail. Figure 2 is a block diagram for explaining the configuration of a prediction server according to an embodiment of the present invention. Figure 3 is a block diagram for explaining the detailed configuration of a prediction server according to an embodiment of the present invention. Figure 4 is a diagram illustrating the configuration of a prediction network for predicting congestion according to an embodiment of the present invention. Figure 5 is a diagram for explaining unit operations of a prediction network for predicting congestion according to an embodiment of the present invention.

First, referring to FIG. 2, the prediction server 10 includes a communication module 11, a storage module 12, and a control module 13.

The communication module 11 is for communicating with the user device 20 and the information provision server 30 through a network. The communication module 11 may include a modem that modulates a transmitted signal and demodulates a received signal to transmit and receive data through a network. This communication module 11 can transmit data received from the control module 13 through a network. Additionally, the communication module 11 can transmit data received through the network to the control module 13.

The storage module 12 serves to store programs and data necessary for the operation of the prediction server 10, and can be divided into a program area and a data area. The program area can store programs that control the overall operation of the prediction server 10, an operating system (OS) that boots the prediction server 10, and applications for prediction. The data area is an area where data generated according to the operation of the prediction server 10 and data necessary for the operation of the prediction server 10 are stored. Various types of data stored in the storage module 120 can be deleted, changed, or added according to user operations.

The control module 13 controls the overall operation of the prediction server 10 and signal flow between internal blocks of the prediction server 10, and can perform a data processing function to process data. This control module 13 may be a central processing unit (CPU), a graphic processing unit (GPU), a digital signal processor (DSP), or the like. Referring to FIG. 3, the control module 13 includes a learning unit 100, a data processing unit 200, and a prediction unit 300.

The learning unit 100 is used to generate a prediction network (EN) through learning using learning data. Different learning data is prepared for each subway station, subway car, and time point for predicting congestion, and the learning unit 100 generates a plurality of prediction networks (EN) using the different learning data. The prediction network (EN) according to an embodiment of the present invention is a deep neural network (DNN) and may be formed as a recurrent neural network. For example, the prediction network (EN) may include Recurrent Neural Network (RNN), Long Short-Term Memory Models (LTSM), and Gated Recurrent Unit (GRU). In an embodiment of the present invention, the prediction network (EN) is described as being formed by a Recurrent Neural Network (RNN), but is not limited thereto, and includes Long Short-Term Memory Models (LTSM), Gated Recurrent Unit (GRU), etc. It can also be formed from other recurrent neural networks.

Referring to FIG. 4, the data processing unit 200 is used to generate a plurality of input vectors (X1 to Xn). The data processing unit 200 is configured to each dispatch a predetermined number of times selected at a predetermined multiple interval (d The dispatch time of the selected train, the congestion level of the corresponding subway car of the dispatched subway, weather information and event information at the dispatch time at which the dispatch was made are collected. For example, assume that the prediction network (EN) consists of 5 stages (S5, n=5). Also, assume that the multiple interval (d×inter) of the dispatch interval (inter) is ‘2×inter’. Then, the data processing unit 200 selects five even-numbered dispatches among the 10 dispatches from the past point in time (P) to the reference point in time (B). And the data processing unit 200 provides the dispatch time of the selected train from the information provision server 30 through the communication module 11, the congestion level of the corresponding subway car of the dispatched subway, weather information and event information at the dispatch time at which the dispatch was made. Collect. At this time, the data processing unit 200 may collect the dispatch time of the selected train from the information providing server 30. Here, the dispatch time includes the month, day, day of the week, and time of the dispatch. A load detection sensor that detects the weight of each subway car is installed, and the data processing unit 200 receives the weight detected in the subway car from the information provision server 30 and normalizes the received weight to determine the congestion level of the subway car. creates . Additionally, the data processing unit 200 may collect weather information at the dispatch time when the selected dispatch was made from the information providing server 30. This weather information may include temperature, humidity, wind speed, weather (whether it is raining or snowing), etc. And the data processing unit 200 can collect event information of the dispatch time at which the selected dispatch was made from the information providing server 30. Event information includes information about various events that may affect the number of floating populations. For example, in a situation where an epidemic is prevalent, such as a pandemic, the number of infected people can be high. As described above, after collecting dispatch time, congestion, weather information, and event information, the data processing unit 200 includes dispatch time, congestion of subway cars, weather information, and event information in response to the plurality of selected dispatches, respectively. A plurality of input vectors are generated, and the generated plurality of input vectors (X1 to ) is generated, and the generated input vector string (Xt: X1 to Xn) is provided to the prediction unit 300.

The prediction unit 300 is used to predict the congestion of the corresponding subway car at the prediction time (F), when trains are dispatched a predetermined number of times from the reference time (B), using the prediction network (EN). The prediction unit 300 inputs an input vector string (Xt: ) Calculate the output vector (Y) representing the congestion level of the corresponding subway car. In other words, one prediction network (EN) is an input vector string corresponding to a dispatch consisting of a predetermined multiple interval (d × inter) of the dispatch interval (inter) from the past time point (P) based on the reference point (B). (Xt: X1 to Calculate an output vector (Y) representing the congestion level of the subway car.

Each of the multiple prediction networks (EN) is created through learning for each subway station, subway car, and prediction point (F). The prediction network (EN) consists of a plurality of stages (S1 to Sn), and each stage includes an input layer (IL), a hidden layer (HL), and an output layer (OL). The length (n) of the stages (S1 to Sn) is set in advance.

The input layer (IL) includes a plurality of input buffers corresponding to each of the plurality of input vectors (X1 to Xn), and inputs the plurality of input vectors (X1 to Xn) to the hidden layer (HL). That is, a plurality of input vectors (X1 to Xn) arranged in time order are input to the input layer (IL).

The hidden layer (HL) includes a plurality of hidden cells (HC) corresponding to a multiple of the number (n) of stages. That is, the hidden layer HL is shown as one layer, but may be formed of two or more layers. A plurality of hidden cells (HC) calculate an output vector (Y) by performing one or more operations in which weights are applied to a plurality of input vectors (X1 to Xn) in a circular manner.

The output layer (OL) outputs the output vector (Y) calculated by the hidden cell (HCn) of the last stage (Sn). The output vector (Y) represents the congestion level of the corresponding subway car at the predicted time (F) when trains are dispatched a predetermined number of times from the reference time (B).

Meanwhile, referring to FIG. 5, the hidden cell (HC) consists of one or more operations to which weights (Weight, Wx, Wh, Wy) are applied. Here, operation refers to an operation applying an activation function. Examples of activation functions include sigmoid, hyperbolic tangent (tanh), Exponential Linear Unit (ELU), Rectified Linear Unit (ReLU), Leakly ReLU, Maxout, Minout, Softmax, etc. In addition, in one hidden cell (HC), the weights are the input weight Wx applied to the input vector (Xt), the state weight Wh applied to the state value Ht-1 of the previous stage, and the output applied to the output value Yt. Includes weight Wy. For example, an operation in which a weight applied to a hidden cell (HC) is applied may be exemplified by Equation 1 below.

Here, b is the threshold or bias. In particular, the tanh function and ReLU function can be changed to other activation functions. Referring to Equation 1 and FIG. 5, each of the plurality of hidden cells (HCt) has the state value (Ht-) of the previous stage (St-1) calculated by the hidden cell (HCt-1) of the previous stage (St-1). 1) and perform an operation where weights (W: Wh, Wx, Wy) are applied to the input value (Xt) of the own stage (St) to obtain the state value (Ht) and output value (Yt) of the current stage (St). can be calculated.

The input layer (IL) of the prediction network (EN) stores each of the plurality of input vectors (X1 to to HCn). Then, each of the first to nth hidden cells (HC1 to HCn) of the hidden layer (HL) has the state value (St-1) of the previous stage (St-1) calculated by the hidden cell (HCt-1) of the previous stage (St-1). Calculate the state value (Ht) of the current stage (St) by performing an operation where weights (state weight and input weight) are applied to the input value (Xt) of Ht-1) and the own (HCt) stage (St). Afterwards, the calculated state value (Ht) is transmitted to the hidden cell (HCt+1) of the next stage (St+1). For example, the first hidden cell (HC1) applies the state weight (Wh) to the initial state value (H0) and the input weight (Wx) to the first input vector (X1) to determine the state of the first stage (S1). Calculate the value (H1). In the case of the first hidden cell (HC1), since there is no previous stage, the initial state value (H0) is used. Subsequently, the second hidden cell (HC2) applies the state weight (Wh) to the state value (H1) of the previous stage, the first stage (S1), and applies the input weight (Wx) to the second input vector (X2). Thus, the state value (H2) of the second stage (S2) is calculated. And the third hidden cell (HC3) applies the state weight (Wh) to the state value (H2) of the previous stage, the second stage (S2), and applies the input weight (Wx) to the third input vector (X3) This is an equation that calculates the state value (H3) of the third stage (S3). In this way, a plurality of hidden cells (HCt) perform operations in which the state and input weights (Wh, Wx) are applied to the state value (Ht-1) of the previous stage and the input vector (Xt), which is the input value of the current stage. The state value (Ht) of the current stage is calculated, and the calculated state value (Ht) is passed to the next stage (St+1).

During the performance of this operation, the last hidden cell, the nth hidden cell (HCn), has a state weight (Wh) applied to the state value (Hn-1) of the previous stage, the n-1th stage (Sn-1), and the current An operation is performed in which the input weight (Wx) is applied to the input vector (Xn), which is the input value of the stage (Sn), to calculate the state value (Hn) of the nth stage (Sn), which is the current stage. Then, the nth hidden cell (HCn) calculates the output vector (Y) by performing an operation that applies the output weight (Wy) to the state value (Hn) of the nth stage (Sn), which is the current stage.

As such, in the prediction network (EN) according to an embodiment of the present invention, the input layer (IL) receives a plurality of input vectors (X1 to Xn) arranged in time order corresponding to each of the plurality of stages (S1 to Sn). Input is received for each stage (S1 to Sn). Then, the plurality of hidden cells (HC1 to HCn-1), excluding the last hidden cell (HCn) of the hidden layer (HL), use the state value (Ht-1) of the previous stage (St-1) and the input of the current stage (St) After calculating the state value (Ht) of the current stage by performing an operation in which the state and input weight (Wx, Wh) are applied to the input vector (Xt), the calculated state value is sent to the next stage (St+1). deliver it to Then, the last hidden cell (HCn) contains the state and input weight (Wx, After calculating the state value (Hn) of the current stage (Sn) by performing an operation to which Wh) is applied, the output vector (Y) is generated by applying the output weight (Wy) to the calculated state value (Hn) of the current stage. Calculate

The output vector (Y) represents the congestion level of the corresponding subway car at the prediction time (F) when dispatch is made a predetermined number of times from the reference point (B). In this way, when the prediction network (EN) calculates the output vector (Y), , the prediction unit 300 derives the congestion level of the corresponding subway car at the corresponding prediction time (F) according to the output vector (Y), and transmits the derived congestion level to the user device 20 through the communication unit 11.

Next, a method of generating a prediction network for predicting the congestion level of a subway car through deep learning according to an embodiment of the present invention will be described. Figure 6 is a flowchart illustrating a method of generating a prediction network for predicting the congestion level of a subway car through learning according to an embodiment of the present invention.

Referring to FIG. 6, the learning unit 100 initializes the prediction network (EN) in step S110. This means initializing the weights of the prediction network (EN), that is, the input weight Wx, state weight Wh, and output weight Wy.

Then, the learning unit 100 prepares learning data in step S120. The learning data includes a learning input vector string containing a plurality of learning input vectors and a learning ground truth, which is a label for the learning input vector string. The learning unit 100 can collect learning data using past data from the information providing server 30 through the communication module 11. A plurality of learning input vectors are prepared according to the size (number of stages) of the input layer (IL) of the initialized prediction network (EN). For example, if the size (number of stages) of the input layer (IL) of the prediction network (EN) is 5, 5 learning input vectors according to time order are required. Accordingly, if the multiple interval (d × inter) of the dispatch interval (inter) is '3 Select 5 as the dispatch interval. Accordingly, the learning unit 100 collects the dispatch time of the selected train, the congestion level of the corresponding subway car of the dispatched subway, weather information and event information at the dispatch time at which the dispatch was made, and normalizes it to generate an input vector for learning. . Additionally, the output vector for learning can be collected depending on what number of times the prediction point (F) is from the reference point (B). For example, if the prediction time (F) is the third dispatch after the reference time (B), the learning unit 100 receives information from the information provision server 30 through the communication module 110 after the reference time (B) of the subway station. The weight of the subway car in the third dispatch is collected, and a ground truth for learning representing the congestion is generated by normalizing the collected weight.

Next, the learning unit 100 inputs a plurality of learning input vectors to the prediction network (EN) initialized in step S130, that is, including weights that have not been learned. Then, the prediction network (EN) calculates an output vector by performing an operation on a plurality of learning input vectors in step S140. In other words, the prediction network (EN) is a learning input vector string (Xt: One or more operations to which weights are applied in the order of a plurality of stages (S1 to Sn) are performed on X1 to Calculate an output vector representing the level of congestion.

Then, the learning unit 100 calculates a loss that is the difference between the output vector corresponding to the plurality of learning input vectors and the learning ground truth, which is the label corresponding to the plurality of learning input vectors, through the loss function in step S150. And, optimization is performed to modify the default value of the prediction network (EN) so that the loss is minimized.

The learning unit 100 may generate a prediction network (EN) by repeating steps S120 to S150 described above using different learning data until the loss becomes less than a predetermined target value.

According to the above-mentioned procedure, a plurality of prediction networks (EN) are created for each subway station, subway car, and prediction point (F). Next, a method for predicting the congestion level of a subway car using a plurality of prediction networks (EN) will be described. Figure 7 will explain a method for predicting congestion for each subway car using deep learning according to an embodiment of the present invention.

Referring to FIG. 7, the data processing unit 200 receives a message requesting the congestion level of the subway station and subway car from the user device 20 through the communication module 11 in step S210. In other words, the message requesting the level of congestion is transmitted with the requested subway station and subway car specified.

Then, in step S220, the data processing unit 200 moves from a past point in time (P) to a predetermined multiple interval (d The created dispatch is selected, and an input vector string (Xt: X1 to Xn) containing a plurality of input vectors (X1 to Xn) corresponding to each selected dispatch is generated. Step S220 is described in more detail as follows. That is, the data processing unit 200 selects a dispatch consisting of a predetermined multiple interval (d Then, the data processing unit 200 receives the dispatch time corresponding to the previously selected dispatch, the congestion level of the subway car corresponding to the dispatch time, and the weather condition at the dispatch time from the information provision server 30 through the communication module 11. Collects event information that quantifies the weather information and events that contribute to whether there are clusters of people corresponding to the dispatch time. Next, the data processing unit 200 sorts a plurality of input vectors (X1 to to Xn).

Next, the prediction unit 300 inputs an input vector string (Xt: X1 to Xn) including a plurality of input vectors (X1 to Xn) into the prediction network (EN) in step S230. That is, a plurality of input vectors (X1 to Xn) sorted according to dispatch time order are input to the prediction network (EN).

Then, in step S240, the prediction network (EN) applies weights to the input vector string (Xt: One or more operations are performed to calculate an output vector (Y) that represents the congestion level of the subway car at the corresponding subway station at the prediction time (F), which is the time when trains are dispatched a predetermined number of times from the reference time (B).

This step S240 will be described in more detail with reference to FIGS. 4 and 5 as follows. The input layer (IL) of the prediction network (EN) receives a plurality of input vectors (X1 to Xn) arranged in time order corresponding to the plurality of stages (S1 to Sn) of the prediction network for each stage (S1 to Sn). Receive input. Then, each of the plurality of hidden cells (HC1 to HCn) of the hidden layer (HL) of the prediction network (EN) contains the state value (Ht-1) of the previous stage (St-1) and the input value of the current stage (St). Calculate the state value (Ht) of the current stage (St) by performing an operation where weights (state weight and input weight) are applied to the input vector (Xt), and then transmit the calculated state value (Ht) to the next stage. do. At this time, the last hidden cell (HCn) among the plurality of hidden cells (HC1 to HCn) applies a weight (output weight) to the calculated state value (Hn) of the current stage to determine the number of subway cars in the subway station at the next dispatch a predetermined number of times. Calculate an output vector representing the level of congestion.

Then, the prediction unit 300 transmits the congestion level according to the output vector (Y) calculated by the prediction network (EN) to the user device 20 through the communication module 11 in step S250. Then, the user device 20 can display the corresponding congestion level so that the user can view it.

Meanwhile, the method according to the embodiment of the present invention described above can be implemented in the form of a program readable through various computer means and recorded on a computer-readable recording medium. Here, the recording medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the recording medium may be those specifically designed and constructed for the present invention, or may be known and available to those skilled in the art of computer software. For example, recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. It includes specially configured hardware devices to store and execute program instructions, such as magneto-optical media and ROM, RAM, and flash memory. Examples of program instructions may include machine language wires, such as those produced by a compiler, as well as high-level language wires that can be executed by a computer using an interpreter, etc. These hardware devices may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

Although the present invention has been described above using several preferred examples, these examples are illustrative and not limiting. As such, those of ordinary skill in the technical field to which the present invention pertains will understand that various changes and modifications can be made according to the theory of equivalents without departing from the spirit of the present invention and the scope of rights set forth in the appended claims.

10: Prediction server 11: Communication module
12: storage module 13: control module
20: User device 30: Information provision server
100: Learning unit 200: Data processing unit
300: prediction unit

Claims (10)

In a method for predicting congestion for each subway car,
A data processing unit receiving a message requesting the congestion level of a subway car at a subway station from a user device;
The data processing unit selecting a plurality of dispatches from the past of a reference point in time when the message is received, and generating an input vector string including a plurality of input vectors corresponding to each of the selected dispatches;
A prediction unit performing an operation on the plurality of input vectors through a prediction network to derive the congestion level of the subway car at the subway station when the next train is dispatched a predetermined number of times from the reference point in time;
Includes,
The step of deriving the congestion of the subway car in the subway station is
The input layer of the prediction network receiving a plurality of input vectors arranged in time order corresponding to the plurality of stages of the prediction network for each of the plurality of stages; and
A plurality of hidden cells in the hidden layer of the prediction network perform a weighted operation on the state value of the previous stage and the input vector, which is the input value of the current stage, to calculate the state value of the current stage, and then the calculated state value is Pass it on to the next stage,
The last hidden cell among the plurality of hidden cells calculates the state value of the current stage and then applies a weight to the calculated state value of the current stage to indicate the congestion level of the subway car in the subway station at the time of the next dispatch of the predetermined number of times. calculating an output vector;
Characterized by including
A method for predicting congestion. delete According to paragraph 1,
Before receiving the message requesting the congestion level,
A learning unit preparing learning data including a learning input vector string including a plurality of learning input vectors and a learning ground truth vector that is a label corresponding to the learning input vector string;
A prediction network with untrained weights performs an operation on the plurality of learning input vectors to calculate an output vector;
Calculating a loss that is the difference between the output vector and the actual learning vector through a loss function, and performing optimization to modify the weights of the prediction network to minimize the loss;
Characterized by further comprising
A method for predicting congestion. According to paragraph 1,
The step of generating an input vector string containing the plurality of input vectors is
The data processing unit selecting a train dispatch at an interval that is a predetermined multiple of the dispatch interval of the subway station from a past time based on the reference point;
The data processing unit collecting event information that quantifies the dispatch time corresponding to the selected dispatch time, the congestion level of the subway car corresponding to the dispatch time, weather information indicating the weather situation, and events contributing to whether or not people are crowded; and
The data processing unit generating an input vector string by sorting a plurality of input vectors each including the dispatch time, the congestion level of the subway car, the weather information, and the event information according to the order of dispatch time;
Characterized by including
A method for predicting congestion. According to paragraph 4,
The congestion level of the subway car is
Characterized in that it is collected through the weight of the subway car.
A method for predicting congestion. In a device for predicting congestion for each subway car,
Receive a message from the user device requesting the congestion level of the subway car at the subway station,
a data processing unit that selects a plurality of dispatches from the past of the reference point in time when the message is received, and generates an input vector string including a plurality of input vectors corresponding to each of the selected dispatches; and
a prediction unit that performs an operation on the plurality of input vectors through a prediction network to derive a congestion level of the subway car in the subway station when the next train is dispatched a predetermined number of times from the reference point;
Includes,
The prediction network is
an input layer that receives a plurality of input vectors arranged in time order corresponding to the plurality of stages for each of the plurality of stages;
It calculates the state value of the current stage by performing a weighted operation on the state value of the previous stage and the input vector, which is the input value of the current stage, and then includes a plurality of hidden cells that transmit the calculated state value to the next stage. However,
The last hidden cell among the plurality of hidden cells calculates the state value of the current stage and then applies a weight to the calculated state value of the current stage to indicate the congestion level of the subway car in the subway station at the time of the next dispatch of the predetermined number of times. a hidden layer that calculates the output vector; and
an output layer that outputs the output vector;
Characterized by including
A device for predicting congestion. delete According to clause 6,
Prepare learning data including a learning input vector string containing a plurality of learning input vectors and a learning ground truth vector that is a label corresponding to the learning input vector string,
A prediction network with untrained weights performs an operation on the plurality of learning input vectors to calculate an output vector,
Calculate a loss that is the difference between the output vector and the actual learning vector through a loss function, and perform optimization to modify the weights of the prediction network so that the loss is minimized.
Learning Department;
Characterized by further comprising
A device for predicting congestion. According to clause 6,
The data processing unit
Selecting a dispatch consisting of a predetermined multiple interval of the dispatch interval of the subway station from a past time based on the reference point,
Collect event information that quantifies the dispatch time corresponding to the selected train, the congestion level of the subway car corresponding to the dispatch time, weather information indicating the weather situation, and events contributing to whether or not people are crowded,
Characterized by generating an input vector string by sorting a plurality of input vectors, each containing the dispatch time, the congestion level of the subway car, the weather information, and the event information, in the order of dispatch time.
A device for predicting congestion. According to clause 9,
The congestion level of the subway car is
Characterized in that it is collected through the weight of the subway car.
A device for predicting congestion.