CN106022541B - An Arrival Time Prediction Method - Google Patents

Abstract

A system and method for predicting arrival time based on a random neural network group. The vehicle-mounted mobile device periodically obtains positioning information (latitude and longitude coordinates) and compares it with the site information (a polygon or circular area) in the vehicle-mounted mobile device. Compare and determine whether it has arrived at the site (or left the site), and then reports the arrival (or departure) information and time point to the cloud server. The cloud server is responsible for collecting the arrival (or departure) information reported by the vehicle-mounted mobile devices and analyzing the travel time between each site, storing this data in the cloud database server, and then using the travel time data set to train random neural networks Parameter values of the network swarm algorithm.

Description

An Arrival Time Prediction Method

The application of the present invention is a divisional application of the invention application whose filing date is March 3, 2015, the application number is 201510094269.5, and the invention name is "A system and method for predicting arrival time".

technical field

The present invention relates to the technical field of wireless communication, in particular to a method for predicting arrival time.

Background technique

At present, the existing technology of public transport arrival time prediction is: using historical data for statistics and averaging, obtaining the average speed and travel time between each station This is to estimate the arrival time. However, these methods cannot reflect the real-time changes in road conditions between stations, thus causing a large error in arrival time information.

The Taiwan patent with publication number TW201137803 mainly proposes to collect the arrival information reported by the bus in the past to estimate the average speed and travel time between stations and stations, and to make statistics according to different weeks and time periods. The historical average speed and travel time can be obtained. Although this method can quickly provide the estimated arrival time, it mainly uses the average value of historical data, so it is impossible to predict the arrival time based on real-time road conditions, so it may cause a large error in the arrival time prediction.

The Taiwan patent with the publication number TW201344647, after predicting the arrival time, provides the driver with speed adjustment suggestions based on the real-time location information of the bus, so as to improve the punctuality rate of the station. Although this method can estimate the arrival time and provide punctuality control, this method mainly uses the average value of historical data, and cannot predict the arrival time according to real-time road conditions, so it may cause arrival at the station. Larger error in time prediction.

The Taiwan Patent Publication No. TW201405497 mainly proposes that when the vehicle-mounted mobile device passes through each road segment, the travel time of each road segment is reported to the back-end monitoring center in real time through the vehicle-mounted mobile device, and then the monitoring center reports the shortest travel time of each road segment. Times and maximum travel times are published to all in-vehicle mobile devices. If the travel time of the in-vehicle mobile device is between the shortest travel time and the longest travel time, it will not be returned. Although this method can effectively grasp the travel time of each road section and reduce the number of transmissions, it does not propose a method for predicting the bus arrival time, so the bus arrival information cannot be predicted.

The Taiwan patent with publication number TW201117146 mainly provides a method for inquiring about the travel time of a bus, allowing users to inquire about the real-time location and travel time of the bus they want to take. Although this method allows users to query the real-time location and travel time of buses, it does not propose a method for predicting bus arrival time, so it is impossible to predict bus arrival information.

The Taiwan patent with publication number TW200828190 mainly proposes to use the user's mobile device to receive station information, and when the user arrives at the station, a notification will be issued to remind the user. Although this method can alert the user when arriving at the station and can provide real-time arrival information, it cannot provide predictive information.

The Taiwan patent with the announcement number TWI252441 mainly proposes to receive satellite positioning signals through the bus, and transmit the position information back to the monitoring center in real time, and then the prediction module of the monitoring center will predict the arrival time according to the real-time position of the bus. Although this method can provide the arrival time prediction, the patent only mentions the reference empirical value, and the future explicitly mentions the prediction method of the bus arrival time.

The Taiwan patent with the announcement number TWI341998 mainly proposes to predict the travel time according to the real-time speed of the bus and the distance from the bus to each station; and calculate the walking time according to the user's walking speed and the distance from the user to each station . Finally, based on travel time and walking time to estimate suitable stops. Although this method can provide a prediction method for bus travel time, this method mainly considers the current real-time speed of the bus and the distance to the station. However, the traffic information between the vehicle and the station is not considered, so it may cause the arrival of the station. Larger error in time prediction.

The Taiwan patent with the publication number TW201232489 proposes to use the Hilbert-Huang Transform (HHT) empirical mode decomposition method combined with the gray model to predict the driving speed, and then convert the travel time and arrival time according to the estimated speed. Although this method can effectively use mathematical and statistical models for speed prediction, because this method uses all data for analysis, it cannot avoid the influence of extreme values, which may cause a large error in the arrival time prediction.

SUMMARY OF THE INVENTION

In view of the above-mentioned problems of the prior art, the purpose of the present invention is to provide a method for predicting the arrival time, by collecting the travel time between stations and stations of each road segment and time period, and propose a novel random neural network group to analyze For the above travel time data set, multiple neural network models are established to avoid the influence of extreme values, and the prediction results of multiple neural network models are comprehensively considered to improve the accuracy of the prediction, so as to predict the bus that the user wants to take. The arrival time of the car is provided to the user as a reference.

The arrival time prediction system of the present invention includes multiple station signs, multiple vehicle terminal devices, multiple cellular network base stations, cloud computing servers, cloud historical databases and multiple arrival time prediction system client devices. Wherein, each station sign has a latitude and longitude coordinate information. When each of the in-vehicle terminal devices approaches the plurality of station signs, each of the in-vehicle terminal devices senses the plurality of longitude and latitude coordinate information, thereby generating arrival information. The arrival information is transmitted through the plurality of cell network base stations, and the cloud computing server calculates the travel time after receiving the arrival information from the cell network base stations, and then predicts the remaining travel time according to the travel time and the query site, and converts it. For the arrival time, the arrival time is then transmitted through the cellular network base station. The cloud history database stores latitude and longitude coordinate information and travel time between station signs. The client device of the arrival time prediction system sends the query site, and receives the arrival time transmitted through the cellular network base station, and then displays the arrival time.

The arrival time prediction method of the present invention includes the following steps: setting random neural network group algorithm parameter values; reading the travel time between stations in the historical database; randomly generating m neural network models; filtering out correct After the neural network model whose rate is lower than the threshold value, there are k neural network models remaining; obtain real-time station-to-station travel time or test data in the test phase; input travel time or test data into the filtered and predict the travel time from station to station; and after obtaining the predicted travel time from station to station, convert it to the arrival time of the target station.

To sum up, the arrival time prediction system and method of the present invention has one or more of the following advantages:

1. The present invention collects the real-time travel time between stations of each road segment and time period to estimate the travel time of the current vehicle position to the target station.

2. The present invention proposes a novel random neural network group to analyze the above-mentioned travel time data set, establish a plurality of neural network models, and then comprehensively consider the prediction results of the multiple neural network models to improve the prediction accuracy, so as to predict The arrival time of the bus that the user wants to take is provided to the user as a reference.

3. In the learning stage of the random neural network group algorithm, the present invention randomly takes out multiple pieces of data from the data set for each neural network model as training data, and uses the remaining data as the test data in the training stage. , and then input the training data into each neural network model for learning, thus avoiding the influence of extreme values.

4. In the testing stage and the implementation stage of the random neural network group algorithm, the present invention uses the travel time predicted by each neural network model and the weight learned in the training stage to perform a weighted average, and finally the travel time after the weighted average is calculated. As the travel time prediction value of this random neural network swarm algorithm, the travel time is converted into the arrival time, so as to predict the arrival time.

Description of drawings

1 is a schematic structural diagram of an arrival time prediction system according to Embodiment 1 of the present invention;

2 is a schematic flowchart of a method for predicting arrival time according to Embodiment 2 of the present invention;

3 is a schematic flowchart of a method for predicting arrival time according to Embodiment 3 of the present invention;

4 is a schematic diagram of four types of neural network models according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of predicting travel time according to Embodiment 5 of the present invention.

Detailed ways

Referring to FIG. 1, the present invention relates to a system for arrival time prediction based on stochastic neural network swarms. The system can mainly predict the arrival time of vehicles, and is suitable for passenger transport operators, logistics operators, or other related operators who have the demand for arrival time prediction, and provides the predicted arrival time to the client device, so that customers or users can Real-time grasp of vehicle information and arrival information to save waiting time, which mainly includes the following six modules: (1) Multiple station signs 100: This station sign equipment mainly contains a set of latitude and longitude coordinate information, and this information can be stored in advance in In the on-board terminal device and the cloud computing server, when the on-board terminal device approaches the station sign, the on-board terminal device can perceive the station information. In addition, this stop sign device can also be embedded with an RFID (Radio Frequency IDentification, radio frequency identification) tag, which can sense the stop sign when the vehicle is approaching, and can use this to determine the arrival of the stop. (2) Multiple in-vehicle terminal devices 101: This device mainly includes a GPS (Global Positioning System, global positioning system) module, a cellular network module, and a database module (not shown in FIG. 1 ), which can collect the current position of the vehicle (including latitude and longitude coordinates), and determine whether the current position is close to the station sign 100, and if it is within the range near the station sign 100, it is determined to arrive at the station, and the arrival information and time point are sent back to the cloud through the cellular network base station 102 The computing server side 103 . In addition, in the part of arrival judgment, the vehicle terminal device 101 can also be embedded with an RFID reader, which can sense the stop sign when the vehicle is approaching, and can receive the RFID tag signal from the stop sign device to determine whether to arrive at the station. (3) Multiple cell network base stations 102: each cell network base station 102 provides data transmission and data reception functions, and is responsible for the vehicle terminal equipment 101, the cloud computing server 103, and the arrival time prediction system client equipment 106 data transfer between. (4) Cloud computing server 103: This server can mainly collect and analyze the arrival information and arrival time point from the vehicle terminal device 101, and calculate the travel time between each station and the station according to each arrival time point, Then, the data set of travel time between the first multiple stations on the driving route of the target station queried by the user is input into the class trained by the random class neural network group arrival time prediction method proposed by the present invention. Neural network group, analyze and calculate to obtain the remaining travel time prediction to the target site, and then convert it into the arrival time to the target site. (5) Cloud historical database 105: This database can mainly store the travel time between each station in history, and can be used as a training data set of random neural network groups to train each neural network model. (6) Multiple client devices 106 of the arrival time prediction system: this device can be a mobile device with a human-computer interaction interface and a network transmission module, allowing users to query and display the target site they want to obtain through this device arrival time forecast. The station and time that the user wants to take can be preset by the user, and then the device can actively update and judge, and automatically send reminder information and sound to the user when the vehicle is about to arrive.

Referring to FIG. 2 and FIG. 3 , the present invention further provides a method for predicting arrival time based on random neural network groups. This method will mainly consist of 2 phases: (a) training phase and (b) implementation and testing phase. Among them, the training phase mainly includes 4 steps, namely:

Step S201: setting the parameter values of the random-like neural network group algorithm;

Step S202: read the travel time between each station in the historical database;

Step S203: randomly generating m neural network models;

Step S208 : after filtering out the neural network-like models whose accuracy rate is lower than the threshold value, k neural network-like models remain.

The implementation and testing phase mainly consists of 3 steps, namely:

Step S301: Obtain real-time travel time between each station or station or test data in the test phase;

Step S302: Input the data into the filtered k neural network models, and predict the travel time between stations;

Step S306: After obtaining the predicted station-to-station travel time, convert it into the arrival time of the target station.

In step S201, the developer of the arrival time prediction system first sets the relevant parameter values of the random neural network group algorithm. The maximum number of hidden layers in the network model (hmax will be used as an example to be explained later), the maximum number of neurons in each hidden layer in the neural network model (cmax will be used as an example to be explained later), and the number of training neural network models. The ratio of the number of training data to the total number of data in the training phase (the r% will be used as an example for description in the following), and the threshold value of the correct rate (the wthreshold will be used as an example for description in the following).

In step S202, the time when the vehicle arrives at each station is obtained from the cloud historical database 103, and converted into the travel time between stations The station time is time point t2, then the travel time from station 1 to station 2 is |t2-t1|. This travel time set is then used as the input and output data of the neural network-like model for subsequent learning. Taking Figure 1 as an example, to predict the time to arrive at station n when the vehicle travels to station n-2 (that is, the target output travel time is |tn-tn-2|), the input travel time data set may include {|t2- t1|,|t3-t2|,...,|tn-2-tn-3|}.

In step S202, according to the parameter values of the random neural network group algorithm set by the developer of the arrival time prediction system, m neural network models are randomly generated, and each neural network model uses the total training data obtained randomly. r% of the data is used for training and learning, and the remaining data (i.e. 100%-r% of the data) is used as the validation of each class of neural network model, each class of neural network model will achieve different data for training and validation. In addition, each neural network-like model will generate 0-hmax hidden layers according to the parameter setting value, and generate 0-cmax neurons for each hidden layer, wherein the hidden layer and neural network of each neural network-like model The combination of elements will all be different. Then input the aforementioned r% data into the neural network-like model for training and learning. After reaching convergence, input the 100%-r% data (that is, the test data in the training phase) into the trained neural network. model, and obtain the predicted travel time, and compare it with the correct travel time to obtain the correct rate of each neural network model, and use the correct rate as the weight value in the implementation and testing phases.

In step S208, after filtering out the neural network-like models whose accuracy rate is lower than the threshold value, the remaining k neural network-like models: compare the accuracy of the randomly generated m neural network-like models with the accuracy rate threshold wthreshold For comparison, after excluding the neural network models below this threshold (that is, the accuracy rate is too low), k neural network models are left; if there is no neural network model whose accuracy is higher than the threshold, Returning to step S201, the developer of the arrival time prediction system resets the threshold value and retrains the random neural network group.

In step S301, in the implementation and test phases, firstly obtain the real-time travel time between stations of the vehicle, for example: when the vehicle moves to station n-2 in Fig. 1, the user wants to query Arrival time prediction for station n (i.e. the travel time of the target output is |t _n -t _n-2 |). At this point, the travel time data set {|t ₂ -t ₁ |,|t ₃ -t ₂ |,...,|t _n-2 -t _n-3 | }, as the input data of the neural network-like model.

In step S302, the acquired real-time travel time data set {|t ₂ -t ₁ |, |t ₃ -t ₂ |,...,|t _n-2 -t _n-3 |} is input to the filter After k neural network models, each neural network model will predict a predicted travel time of |t _n -t _n-2 |, and then multiply it by the neural network of each class obtained by training The weight value of the model (that is, the correct rate of each class of neural network models during the training phase), and the sum of the weighted values is divided by the sum of the weight values (that is, weighted average).

In step S301, after obtaining the predicted travel time |t _n -t _n-2 | obtained by comprehensively considering k neural network models, add the predicted travel time |t to the real-time time point t _n-2 of the vehicle _n -t _n-2 | Get the arrival time prediction at station n, and provide the prediction result to the user.

The present invention collects and analyzes the arrival (departure) station information (including station information and time point, etc.) returned from the vehicle terminal device 101, converts this data set into the travel time between stations and stores, and stores it in the cloud historical database In 105, and in the cloud computing server 103, an arrival information prediction method module based on random neural network swarm algorithm is designed and implemented, which can access the travel time set in the cloud historical database 105 and input it into the random In the arrival information prediction method module of the neural network swarm algorithm, the neural network model is trained to predict the travel time. When the client of the arrival information prediction system predicts the arrival time of a station, the information of the first multiple stations reported by the current vehicle terminal device 101 on the route can be input into the trained neural network group to travel to the target station The time prediction, reconversion and provision of the arrival time to the target site to the arrival time prediction system client device 106. The technical feature of the present invention is mainly to propose and design a random-like neural network group algorithm, and apply it to the arrival information prediction method, which will be described by way of embodiments below.

The present invention provides a system for predicting arrival time based on a random neural network group, the system architecture of which is shown in FIG. 1 . The system includes multiple station signs 100 , multiple vehicle terminal devices 101 , multiple cellular network base stations 102 , a cloud computing server 103 , a cloud historical database 105 , and multiple arrival time prediction system client devices 106 . In this embodiment, taking the station sign 100 of the same route as an example, there are n stations in this route, and each station has location information (including longitude and latitude). As shown in Table 1, Route 1 contains a total of 12 stations (that is, n in Figure 1 is 12), and the corresponding latitude and longitude can be stored in the vehicle terminal device; when vehicle number 1 travels from station 1 to station 2 , at 2014/4/1 14:53, the GPS module of the vehicle terminal device detects that the longitude of the vehicle is 120.97839 and the latitude is 24.808658, and it is estimated that the vehicle is close to station 2 (for example, within a straight-line distance of 30 meters). station, and the arrival information (including the station number and time point) is sent back to the cloud computing server 103 through the cellular network base station 102 .

In addition, the station sign 100 may also be provided with an RFID tag, and the vehicle-mounted terminal device 101 may be provided with an RFID reader. When the vehicle-mounted terminal device 101 is close to the station stop sign 100, it can detect the RFID tag of the station stop sign 100, and use the This is judged as the arrival, and then the arrival information (including the station number and time point) is sent back to the cloud computing server 103 through the cellular network base station 102 . The vehicle arrival information report data set is shown in Table 2, which can mainly record the route number, vehicle number, station number, and time point, etc., and the cloud computing server 103 can convert the vehicle arrival information into the station arrival travel time information ( As shown in Table 3), and store the information in the cloud history database 105. For example, vehicle number 1 departs from station 1 at 2014/4/1 14:46:28 and arrives at station 2 at 2014/4/1 14:53:31, so the travel time from station 1 to station 2 is 423 seconds; and the time when vehicle number 2 departs from station 1 is 2014/4/1 19:32:22, and arrives at station 2 at 2014/4/1 19:40:13, so the trip from station 1 to station 2 The time is 471 seconds.

When the vehicle with number 10001 travels to station 6 (that is, the cloud server 103 knows the travel time between station 1 and station 6), there is an arrival time prediction system client device to the cloud computing server. 103 Query the arrival time of the station 12 of the route number 1 (ie, predict the travel time from the station 6 to the station 12, and convert it into the arrival time of the station 12). At this time, the cloud computing server 103 can use the data in the cloud historical database 105 (that is, the station-to-station travel time information of the route numbers 1 and 2, as shown in Table 4) as the data of the random neural network group algorithm in the training phase, To establish a random neural network group, and use this algorithm to predict the arrival time.

Table 1 Station location information

Table 2 Vehicle Arrival Information

Table three station arrival travel time information

Table 4. Data of training phase of random class neural network swarm algorithm

The method for predicting the arrival time based on the random neural network group of the present invention has a method flow as shown in FIG. 2 and FIG. 3 . This method mainly consists of 2 phases: (a) training phase and (b) implementation and testing phase.

The training phase mainly includes 4 steps, namely step S201: setting the parameter value of the random class neural network group algorithm; S202: reading the travel time between each station in the historical database; S203: randomly generating m classes A neural network model; and S208 : after filtering out the neural network-like models whose accuracy rate is lower than the threshold value, remaining k neural network-like models.

The implementation and testing phases mainly include three steps, namely S301: obtaining the real-time travel time between each station or station or the test data in the testing phase; S302: inputting the data into the filtered k-class neural network models , and predict the travel time from station to station; and S306 : after obtaining the predicted travel time from station to station, convert it into the arrival time of the target station.

In the training phase, the relevant parameter values of the random neural network swarm algorithm will be set by the developer of the arrival time prediction system (step S201 ). For example, set a total of 10 neural network models (that is, m is 10), the maximum number of hidden layers in the neural network model is 5 (that is, hmax is 5), and the maximum number of neurons in each hidden layer in the neural network model is 7 (that is, cmax is 7), the ratio of the training data of the training neural network model to the total number of training data is 60% (that is, r% is 60%), and the correct rate threshold is 0.945 (that is, wthreshold is 0.945 = 94.5%), and then 10 neural network models will be generated according to this parameter value to predict the arrival time.

In this step S202, the time when the vehicle arrives at each station in the history obtained from the cloud history database is converted into the travel time between stations, as shown in Table 4. In this embodiment, the vehicle to be predicted travels to station 6, and the arrival time of station 12 is to be predicted, and the data set of arrival time between station 1 and station 6 is known { t ₁ , t ₂ , t ₃ , t ₄ , t ₅ , t ₆ }, converted to station-to-station travel time data set {| t ₂ -t ₁ |,| t ₃ -t ₂ |,| t ₄ -t ₃ |,| t ₅ -t ₄ |,| t ₆ -t ₅ |}, and used to predict the travel time from station 6 to station 12 (that is, the travel time of the target output is | t ₁₂ -t ₆ |). In this embodiment, the travel time data set {| t ₂ -t ₁ |, | t ₃ -t ₂ |, | t ₄ -t ₃ |, | t ₅ -t ₄ |, | t ₆ -t ₅ | } are named as parameter names {x1, x2, x3, x4, x5}, respectively, and the travel time of the target output | t ₁₂ -t ₆ | is named as parameter name y.

In step S203, m random neural network models are generated, and step S204 is further included: generating training data and verification data. Specifically, the present invention randomly generates 10 neural network models according to the parameter values of the random neural network group algorithm set by the developer of the arrival time prediction system, and sets the maximum number of hidden layers in the neural network model to 5, The maximum number of neurons in each hidden layer in the neural network model is 7, that is, the number of hidden layers in each neural network model will be between 0 and 5, and the number of neurons in each hidden layer will be between 0 and 7. Examples of resulting results are shown in Table 5 (step S205). The hidden layer of the neural network-like model 1 is 1 layer, and the number of neurons in the hidden layer of this layer is 2 (as shown in Figure 4); the hidden layer of the neural network-like model 2 is 2 layers, and the neurons of the first hidden layer are The number of elements is 3, and the number of neurons in the second hidden layer is 4; by analogy, 10 neural network models can be obtained. In addition, since the number of training data for training neural network models accounts for 60% of the total number of data in the training phase, taking Table 4 as an example, the total number of data in the training phase is 10,000, so each neural network model will randomly 6000 samples were taken out as training neural network models for learning, and the remaining 4000 TDTRS (Testing Data in TRaining Stage, test data in the training stage) were used for verification of each neural network model in the training stage. In this step, the sets of 6000 pieces of data obtained by each neural network model are randomly generated, and each neural network model will obtain different data sets for training and learning.

Table 5 Random class neural network group

Step S206: training and learning of the neural network-like model. In this embodiment, 6000 pieces of data are input to 10 neural network models respectively for training and learning. The following uses neural network model 1 (as shown in FIG. 4 ) as an example for description. The 6000 pieces of data are a data combination that includes the route number 1 and does not include the route number 10000, and the training and learning of the neural network model 1 is described as follows.

Step i: Randomly generate the weights of each neuron and the constant terms of the neurons in the hidden layer and the output layer, as shown in Table 6.

The weights of each neuron in the six types of neural network model 1, and the constant terms of the neurons in the hidden layer and the output layer

w_1,6 w_2,6 w_3,6 w_4,6 w_5,6 w_1,7 w_2,7 w_3,7 w_4,7 w_5,7 w_6,8 w_7,8 6 7 8 0.7 0.7 0.2 0.1 0.6 0.1 0.8 0.5 0.3 1.0 0.6 0.6 0.8 0.7 0.3

)的方式计算输出信号，计算方式如下所示。Step ii: Input the 6000 pieces of data into the neural network-like model 1 one by one. The following takes the route number 1 as an example. First, the data is normalized to a value between 0 and 1. Therefore, the data in the embodiments are all less than 5000, so the normalization is performed by dividing by 5000. The results are shown in Table 7. Then, according to the input signal, the output signal of each hidden layer neuron is calculated, wherein in this embodiment, Logistic distribution (ie,

) to calculate the output signal, and the calculation method is as follows.

Table 7 Normalized route number 1 value

Neuron 6:

Total input signal:

Convert the output signal:

Neuron 7: Total input signal:

Convert the output signal:

Step iii: Calculate the output signal of the neurons in the output layer according to the output signal of the hidden layer.

Neuron 8:

Total input signal:

Convert the output signal:

Step iv: Compare the error term of the output value (ie 0.759554) with the true value (ie 0.7796).

Neuron 8 error term:

Step v: Feed the error term back to the hidden layer, and calculate the error term of the neurons in the hidden layer respectively.

Neuron 6 error term:

Neuron 7 error term:

Step vi: According to the neuron error term, update the weight and constant term of each neuron, and set the learning rate σ to 0.8 in this embodiment.

Step vii: Repeat steps ii to vi, and input each data into the neural network-like model for learning, until the difference between the output signal of this round and the output signal of the previous round is lower than the threshold value othreshold (in this example). othreshold is set to 0.01), then convergence is achieved and learning is completed, and each neuron weight and constant term of this type of neural network model are determined.

The above is the training and learning process of the neural network-like model 1, and other neural-like network models (ie, the neural network-like model 2 to the neural network-like model 10) are simultaneously trained, which can support parallel operations. After the training is completed, steps ii to iii can be repeated when predicting the travel time between station 6 and station 12, using the test data or real-time data as the input signal, and the output signal as the travel prediction value. Among them, the predicted value of travel time produced by the neural network model needs to be normalized and restored before the travel time in seconds can be obtained. For example, if the output signal is 0.759554, it needs to be multiplied by 5000 to obtain the travel time of 3797.769233 seconds.

Step S207: Verification and weight of the neural network-like model. After the training and learning of all neural network models are completed, the remaining 4000 pieces of data can be used to verify each neural network model, and the average correct rate is calculated as the weight of each neural network model. Taking the neural network-like model 1 as an example, all the test data in the training phase are input into the trained neural network-like model 1 and repeat steps ii to iii to calculate the correct rate. For example, when the route number 10000 is the input signal, the normalized value is shown in Table 8, the predicted value is 0.75986369, and then the predicted value is multiplied by 5000 to be 3799.318449, and the correct rate is 1-(|true value- predicted value|/true value)=1-(|3939-3799.318449|/3939)=96.45%; and so on, the average correct rate of test data (TDTRS) in 4000 training stages can be calculated, in this case 93.23 %. In this embodiment, the average correct rates corresponding to 10 neural network models are 93.23%, 94.90%, 94.03%, 93.57%, 94.61%, 93.52%, 94.93%, 95.21%, 94.48%, 94.45%, respectively. As shown in Table 9.

Table 8 Normalized route number 10000 value

Table 9 Average correct rate of each class of neural network model

Step S208 : after filtering out the neural network-like models whose accuracy rate is lower than the threshold value, k neural network-like models remain. In this step, the average correct rate of each neural network model will be analyzed, and the threshold value wthreshold (that is, 94.5% set in this embodiment) that is lower than the correct rate will be filtered out. The neural network model 1, the neural network Model 3, neural network model 4, neural network model 6, neural network model 9, neural network model 10, etc. 6 will be filtered out, leaving 4 neural network models and their weight values for implementation and testing stage use.

Table 10. Filtered neural network-like models and their weights

In step S301, in the implementation and testing stages, the real-time vehicle arrival information is taken and input into the trained random neural network group to predict the arrival time. For example, if the client device of the arrival time prediction system wants to query the arrival time of station 12 at 2014/5/311:59:00, it will take the arrival time of station 1 to station 6 and the travel time between station to station (as shown in Table 11), as the input data of the random neural network group (as shown in Table 12), the target predicted value of travel time from station 6 to station 12 is obtained.

Table 11 Vehicle Arrival Information

Table 12 Station-to-station travel time information

In addition, the developers of the arrival time prediction system can also collect historical data at this stage as the test data (TDTES) in the test phase, and obtain the travel time between each station of each route number as the input of the random neural network group value to analyze and optimize swarms of random neural network-like networks.

In step S302, data is input into the filtered k-class neural network models, and travel time between stations is predicted. As shown in Figure 5, after the input data is obtained, the data can be used as each filtered neural network model (ie neural network model 2, neural network model 5, neural network model 7, neural network model 8). , as shown in Table 10), and the travel times predicted by neural network model 2, neural network model 5, neural network model 7, and neural network model 8 are 3766.607 seconds, 3857.98 seconds, 3661.828 seconds, 3724.095 seconds (step S303), as shown in Table 13. Finally, the weighted average of the weights of each neural network model (steps S304 to S305) is used to obtain a travel time prediction value of 3752.516552 seconds (ie [94.90%*3766.607+94.61%*3857.98+94.93%*3661.828+95.21%*3724.095 ]/[94.90%+94.61%+94.93%+95.21%]=3752.516552).

Table 13 Filtered neural network models and their weights

In step S306, after obtaining the predicted station-to-station travel time, it is converted into the arrival time of the target station. After obtaining the predicted value of travel time from station to station, it can be converted into the arrival time to the target site according to the current arrival information and combined with the predicted value of travel time from station to station. In this embodiment, the time point when the route number 10001 arrives at station 6 is 2014/5/3 11:58:46, and the predicted travel time from station 6 to station 12 is 3752.516552 seconds, so the predicted arrival time at station 12 is 2014/ 5/3 13:01:19, and then send this information back to the client device of the arrival time prediction system.

From the example of the actual application in the passenger transport industry, the data of the passenger transport operator A is used as an empirical test. The data of the whole month of March 2014 are collected, including a total of 2956 trips. The experimental environment covers a total of 40 road sections, and different road sections are used respectively. Data mining algorithm to test its accuracy, including Logistic Regression (LR), traditional Back-Propagation Neural Network (BPNN), and random neural network group proposed by the present invention (Random Neural Networks, RNN), confirming that this method is indeed superior, and the experimental results are shown in Table 14.

Table 14 Comparison of performance between the present invention and other data mining methods

method Correct rate historical data average method 73.79% Logis Returns 77.43% backward pass neural network 77.88% this invention 78.22%

To sum up, the system and method for predicting arrival time based on random neural network group of the present invention collects the travel time between stations and stops for each road segment and time period, and proposes a novel random neural network group to predict the arrival time. Analyze the above travel time data set, establish multiple neural network models to avoid the influence of extreme values, and comprehensively consider the prediction results of multiple neural network models to improve the prediction accuracy, so as to predict the bus that the user wants to take The arrival time of the train is provided to the user as a reference.

The above description is exemplary only, not limiting. Any equivalent modifications or changes that do not depart from the spirit and scope of the present invention shall be included in the appended patent application scope.

【Symbol Description】

100: Station sign

101: Vehicle terminal equipment

102: Cell Network Base Station

103: Cloud computing server

104: Cloud computing room

105: Cloud History Database

106: Arrival time prediction system client device

S201～207, S301～S306: Steps

1 to 8: class neural network model

Based on the same inventive concept, an embodiment of the present invention also provides a mobile terminal. Since the method corresponding to the mobile terminal in FIG. 3 is a method for starting a mobile terminal in an embodiment of the present invention, the implementation of the method in the embodiment of the present invention can refer to The implementation of the system will not be repeated here.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each process and/or block in the flowchart illustrations and/or block diagrams, and combinations of processes and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

Although the preferred embodiments of the present invention have been described, additional changes and modifications to these embodiments may occur to those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be construed to include the preferred embodiment and all changes and modifications that fall within the scope of the present invention.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit and scope of the invention. Thus, provided that these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims (3)

1. a method for predicting arrival time, characterized in that the method comprises: Set multiple random neural network group algorithm parameter values; Read multiple travel times between stations in the historical database, and use the set of these travel times as training data; Randomly generate m neural network models, wherein each of the neural network models randomly obtains r% of the training data for training and learning, and uses the remaining data of the training data as verification, wherein the remaining data (100%-r%) of the training data, and the r% of the training data obtained by each of the neural network models that each receive r% of the training data are different from each other , after reaching convergence, a predicted travel time is generated based on the remaining data, and the predicted travel time is compared with the correct travel time to obtain the correct rate of each type of neural network; After filtering out the neural network-like models whose accuracy rate is lower than the threshold value, there are k neural network-like models remaining, where the threshold value of the accuracy rate is 0.945; Obtain real-time station-to-station travel times or multiple test data in the test phase, wherein when the user wants to query the arrival time of the vehicle arriving at station n, and the vehicle is located at station ni, the The multiple travel times or test data include |t ₂ -t ₁ |, |t ₃ -t ₂ |,...,|t _ni -t _ni-1 |, where t _k represents the arrival time of the vehicle to station k station time; Inputting the plurality of travel times or test data into the filtered k neural network models, and predicting a plurality of travel times from station to station by each of the k neural network models; After the plurality of predicted station-to-station travel times are obtained from each of the k types of neural network models, the predicted station-to-station travel times are determined according to the accuracy of each of the k types of neural networks. The travel time is weighted and averaged to convert to the arrival time at the destination site. 2. The method according to claim 1, wherein the multiple random neural network group algorithm parameter values include the number of neural network models, the maximum number of hidden layers in the neural network model, and the number of neural network models. The maximum number of neurons in each hidden layer, the ratio of the number of training data for training the neural network model to the total number of data in the training phase, and the threshold for the correct rate. 3. The method according to claim 2, wherein after filtering out the neural network model whose correct rate is lower than the threshold value, the steps of remaining k neural network models include: Comparing the accuracy rates of the randomly generated m neural network-like models with the accuracy rate threshold, and excluding the neural network-like models below the accuracy rate threshold.

Images