CN107945534A - A kind of special bus method for predicting based on GMDH neutral nets - Google Patents

Abstract

The invention relates to a traffic traffic flow prediction method based on a GMDH neural network, which includes offline vehicle flow training of the GMDH neural network and real-time prediction of the online vehicle flow of the GMDH neural network. The beneficial effects of the present invention are: the method uses the GMDH neural network algorithm to predict the traffic flow at traffic intersections. The general prediction method has the disadvantages of long time and low accuracy in the process of processing large volumes of data, and it is difficult to realize traffic flow. Requirements for real-time prediction; due to the strong approximation ability of the GMDH neural network, the prediction of traffic flow can be divided into two parts: offline learning and online prediction: in the offline learning link, a large amount of data is combined, and the training of the neural network is used to learn the traffic flow. The law of change; in the online prediction part, by calling the neural network that has been learned, it can quickly and effectively predict the traffic status of the vehicle in real time.

Description

A Traffic Flow Forecasting Method Based on GMDH Neural Network

technical field

The invention relates to a traffic vehicle flow prediction method, in particular to a traffic vehicle flow prediction method based on a GMDH neural network.

Background technique

Most of the existing traffic flow forecasting methods use time series models, or use linear models similar to time series models. This type of method regards the traffic flow at a certain moment as a non-stationary random sequence, and analyzes and calculates it in the dimension of time. Take methods based on time series models as an example. This type of method is based on a large amount of uninterrupted data and has high prediction accuracy, but requires complex parameter estimation, and the calculated parameters are not portable. In practical applications, due to various reasons, it is easy to cause data omission, which can easily lead to a reduction in the prediction accuracy of the model. In addition, it also relies on a large amount of historical data, and the data cost is relatively high.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art, and provide a traffic flow prediction method based on GMDH neural network with high prediction accuracy, little dependence on historical data, and low data cost.

A traffic flow prediction method based on GMDH neural network, comprising the steps:

1), GMDH neural network offline traffic flow training

1.1), read the historical data of vehicle traffic; the historical data is collected according to the intelligent camera on the vehicle road, including the speed of the passing vehicle, the license plate number of the vehicle and other information;

1.2) Determine whether the smart camera is working normally and whether the collected vehicle data is correct; remove the data such as excessive speed, incorrect recognition of the vehicle license plate number, discrepancies in the front and rear driving trajectories of the vehicle, and ghost vehicles.

1.3), initialize the GMDH neural network; set the structure of the neural network according to the frequency of data collection and the total number of days of data collection, and the dimension of the neural network is generally proportional to the size of the data volume to ensure the accuracy of the prediction of the neural network;

1.4), carry out the training of neural network; Use gradient descent method to adjust the weight and bias of neural network, realize the training of neural network by traversing each layer of neural network multiple times;

1.5), save the optimal training result; after the neural network training is finished, the structure, weight and bias of the neural network should be saved, so as to be used for the real-time prediction of online traffic flow in step 2);

2), GMDH neural network online traffic flow real-time prediction

2.1), read the historical data of vehicle traffic; the historical data is collected according to the intelligent camera on the vehicle road, including the speed of the passing vehicle, the license plate number of the vehicle and other information;

2.2) Determine whether the smart camera is working normally and whether the collected vehicle data is correct; exclude data such as excessive speed, incorrect identification of the vehicle license plate number, discrepancies in the front and rear driving trajectories of the vehicle, and ghost vehicles;

2.3), initialize the GMDH neural network according to the neural network structure determined in step 1;

2.4), carry out the prediction of neural network; Use the structure of the neural network obtained in previous step 1), weight and bias to predict the real-time traffic flow;

2.5), the output result is used to control the traffic situation.

As preferably: step 1.2) specifically comprises the following steps:

1.2.1), judging the number n of vehicles passing per unit time, should satisfy the constraint: n _nin <n<n _max (n _nin , n _max is the minimum and maximum number of vehicles passing per unit time), if this constraint cannot be satisfied, explain There was a problem with the data collection process.

1.2.2), judging that the number n ₁ of vehicles at this intersection meets the constraints with the number n _i of vehicles at other intersections at the current time: a·n _i <n ₁ <b·n _i (a, b is the corresponding coefficient), if not Satisfying the current constraints indicates that there is a problem in the data collection process.

1.2.3), other methods for verifying the collected data are also included in the scope of this patent requirement.

1.2.4), if it is judged that the data is inaccurate, use the corresponding data restoration plan to complete the data, and speculate the corresponding missing data from the historical data.

1.2.5), if it cannot be inferred from the data, an error will be reported and the system will be required to maintain the sensor module.

As preferably: step 1.3) specifically comprises the following steps:

1.3.1), the number of days D to read the training data is used to measure the dimension of the data and determine the scale of the neural network.

1.3.2), read the sampling period T of the training data, and calculate the number N of data collected every day, and determine the scale of the neural network.

1.3.3), determine the number of input neurons C _in =N of the neural network, and the number of output neurons C _out =N of the neural network, and determine the number of neurons of the middle layer of the neural network C _min =2N.

1.3.4), determine the number of neuron layers of the neural network, L=2+(C _min +C _out ) ^0.5 .

1.3.5), save the corresponding neural network information for use in initializing the neural network in 1.4).

As preferably: step 1.4) specifically comprises the following steps:

1.4.1), initialize the neural network; according to the neural network information saved in 1.3), establish the neural structure, initialize the weight W of the neural network, the bias equivalent B, initialize the layer number pointer P _f =1, and learn the number of times P _s =1.

1.4.2), obtain the output of the current layer; use I, B to calculate the output matrix O of the current layer, the formula is O=I*W+B.

1.4.3), record the output data O of this layer, for the study of step 1.3.4).

1.4.4), judge whether the operation reaches the last layer, if not, add 1 to the number of training layers, return to step 1.4.2), if reached, then enter step 1.4.5).

1.4.5), judge whether the data reaches the number of training times, if not, add 1 to the number of training times, and at the same time, the layer number pointer returns to 1, and returns to step 1.4.2); if it reaches step 1.4.6).

1.4.6), save the training results.

As preferably: step 2.2) specifically comprises the following steps:

2.2.1), judging the number n of vehicles passing per unit time, should meet the constraints: n _nin <n<n _max (n _nin , n _max is a selected coefficient), if the current constraints cannot be satisfied, it means that there is a problem in the data collection process .

2.2.2), judging that the number n ₁ of vehicles at this intersection should satisfy the constraint relationship with the number n _i of vehicles at other intersections at the current time: a·n _i <n ₁ <b·n _i (a, b is the corresponding coefficient), if the current constraints cannot be satisfied, it means that there is a problem in the data collection process.

2.2.3), other methods for verifying the collected data are also included in the scope of this patent requirement.

2.2.4), if it is judged that the data is inaccurate, use the corresponding data restoration scheme to infer the corresponding missing data from the historical data.

2.2.5), if it cannot be inferred from the data, then report an error to the use of this method and require the sensor module to be maintained.

As preferably: step 2.3) specifically comprises the following steps:

2.3.1), the continuous number of days D to read the training data.

2.3.2), read the sampling period T of the training data, and calculate the number N of data collected every day through the formula, which is used to measure the scale of the dimension of the data read into the neural network.

2.3.3), the number of input neurons C _in =N read into the neural network, and the number of output neurons C _out =N of the neural network, the number of intermediate layer neurons read into the neural network C _min =2N.

2.3.4), the number of neuron layers read into the neural network, L=2+(C _min +C _out ) ^0.5 .

2.3.5), save the corresponding neural network information for use in initializing the neural network in 1.4).

As preferably: step 2.4) specifically comprises the following steps:

2.4.1), read the network structure, read the weight W of the neural network saved in 1.4), the bias equivalent value B, and set the layer number pointer to 1.

2.4.2), obtain the output of the current layer; use the input matrix I of the current layer as raw material to calculate the output matrix O of the current layer, the formula is O=I*W+B.

2.4.3), record the output data O of this layer, for the prediction of step 2.3.4).

2.4.4), judge whether the operation reaches the last layer, if not, add 1 to the number of training layers, return to step 1.4.2), if reached, then enter step 1.4.5).

2.4.5), save the prediction result.

The beneficial effects of the present invention are: the method uses the GMDH neural network algorithm to predict the traffic flow at the traffic intersection, and the general prediction method has the shortcomings of long time and low accuracy in the process of processing large volumes of data, and it is difficult to realize the traffic flow Requirements for real-time forecasting. Due to the strong approximation ability of the GMDH neural network, the prediction of traffic flow can be divided into two parts: offline learning and online prediction: in the offline learning link, a large amount of data is combined to learn the law of traffic flow changes through neural network training; The online prediction part quickly and effectively predicts the traffic status of vehicles in real time by calling the learned neural network. The learning of offline data can be carried out by rolling, so that the neural network can conform to the transformation characteristics of real traffic flow in real time.

Description of drawings

Figure 1 is a patent flow chart

Figure 2 is the basic processing unit of the GMDH network

Figure 3 is a schematic diagram of the formation of a 5-input GMDH network

Figure 4 is the structure diagram of the first layer of the GMDH network

Figure 5 is the screening diagram of the first layer intermediate model of the GMDH network

Figure 6 is a structural diagram of the second layer of the GMDH network

Figure 7 is the screening diagram of the second-layer intermediate model of the GMDH network

Figure 8 is an example 1 of GMDH prediction results

Figure 9 is an example 2 of GMDH prediction results

Figure 10 is an example 3 of GMDH prediction results

Figure 11 is an example 4 of GMDH prediction results

Figure 12 is Example 5 of GMDH prediction results.

Detailed ways

The present invention will be further described below in conjunction with the examples. The description of the following examples is provided only to aid the understanding of the present invention. It should be pointed out that for those skilled in the art, without departing from the principle of the present invention, some improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

1 Algorithm principle description

The main idea of the GMDH algorithm is to simulate the "genetic-variation-selection-evolution" process of organisms: starting from a simple initial model set, the elements in the model set are combined with each other according to certain prescribed rules to generate new intermediate candidates model (heredity, variation), and then screen (select) the intermediate candidate model through a certain strategy or scheme, and repeat this process of inheritance, variation, selection and evolution, so that the complexity of the generated intermediate model continues to increase, until The complexity of the newly generated model is no longer increased to obtain the optimal complexity model.

The GMDH algorithm starts from the sample of factors that have an impact on the system, and divides the sample data into a training set, a test set and a prediction set. The data in the training set is used to estimate the parameters of each intermediate model generated in the modeling process (generally, the least square method can be used), and the sample data in the test set is used to combine the external criteria (mainly the error sum of squares criterion, the minimum information criterion, average relative error criterion) to screen the generated intermediate candidate models. The network termination rule modeled by the GMDH algorithm is given by the optimal complexity principle.

Before GMDH modeling, it is generally necessary to select an appropriate initial model set as the initial layer variable. The initial model is generally generated by a reference function, and the reference function generally uses the Kolmogorov-Gabor (K-G) polynomial of formula (1):

2GMDH network construction

Schematic diagram of the basic processing unit of the GMDH network. This is a double-input single-output structure, and the transfer function (also known as the reference function) can be in various forms, such as:

y=ax ₁ +bx ₂ +c

y＝a ₀ +a ₁ x ₁ +a ₂ x ₂ +a ₃ x ₁ ² +a ₄ x ₂ ² +a ₅ x ₁ x ₂

In the formula, x ₁ , x ₂ are the two inputs of the dual-input system, and y is the output of the system.

Figure 2 is a schematic diagram of the formation of a GMDH network containing 5 modeling variables, where vi is the input,

y _ij is the j-th intermediate model of the i-th layer, w _i is the input of the second layer, z _i is the input of the third layer, s ₁ and s ₂ are the inputs of the penultimate layer, and y is the optimal model. In the modeling process, the structure of the GMDH network is determined by self-organization, and the number of layers of the network is not fixed.

The construction steps of the GMDH network are as follows:

(1) Determination of initial layer variables. The GMDH modeling process is completed by self-organization, and the user only needs to give the model input. The initial model is generated by combining the input variables in pairs according to the transfer function, and the transfer function is generally a Kolmogorov-Gabor (referred to as K-G) polynomial shown in the formula.

The initial model set is generally selected as:

V={v ₁ =x ₁ ,v ₂ =x ₂ ,v ₃ =x ₃ ,v ₄ =x ₁ ² ,v ₅ =x ₂ ² ,v ₆ =x ₃ ² ,v ₇ =x ₁ x ₂ , v ₈ ＝x ₁ x ₃ , v ₉ ＝x ₂ x ₃ }

(2) Starting from the initial model, the variables are recombined with each other to obtain a new first-level intermediate model (as shown in Figure 4).

If the number of input variables is five,

(3) According to the outer criterion, the intermediate model of the first layer is selected, and the remaining intermediate model is used as the input variable of the next layer, as shown in Figure 5.

(4) Regeneration and reselection. Repeat step (2), and use the output of the processing unit retained in the first layer as the input variable of the next layer to continue to construct the processing unit of the second layer (Fig. 6). Repeat step (3) to screen the second layer of processing units, and the remaining units continue to construct the next layer of units (Fig. 7).

3GMDH algorithm modeling process

The GMDH algorithm is a modeling method based on sample division, which divides the modeling data sample set W (n data samples) into a training set A (training set A), the number of samples: n _A , and a testing set B (testing set B) , Number of samples: n _B , the modeling data mentioned here does not include the prediction verification data. The data storage is shown in Table 1, and the input and output data forms of the model are as follows:

In the above formula, XAXB are the training data input and test data input respectively, m is the number of system modeling variables, y _A , y _B are the actual output of the training sample and the actual output of the test sample. And n=n _A +n _B , W=A∪B. (x _A , y _A ) is the training set A, which is used to estimate the parameters of the intermediate model. (x _B , y _B ) is the test set B, which is used to screen the intermediate model. The data storage form is shown in Table 1.

Table 1

4GMDH Algorithm Prediction Result Analysis

Figure 8-12 shows the results of five sets of data using the GMDH algorithm for data prediction. The GMDH neural network is trained using data from the east intersection of Xixi Road, Tianmushan Road, Hangzhou City for 60 consecutive days. Then use the trained neural network to predict the traffic flow of the next day with continuous 30-day data as input. The traffic flow conditions in Figures 8-12 are the comparisons between the traffic flow data and the forecast data from November 15 to 15, 2016. Overall, the error is less than 3%, which has a strong use value.

Claims (7)

1. a traffic flow prediction method based on GMDH neural network, is characterized in that, comprises the steps: 1), GMDH neural network offline traffic flow training 1.1), read the historical data of vehicle traffic; the historical data is collected according to the smart camera on the vehicle road, including the speed of the passing vehicle and the license plate number information of the vehicle; 1.2) Determine whether the smart camera is working normally and whether the collected vehicle data is correct; exclude excessive speed, wrong vehicle license plate number recognition, discrepancies between the front and rear driving tracks of the vehicle, and ghost vehicle data. 1.3), initialize the GMDH neural network; set the structure of the neural network according to the frequency of data collection and the total number of days collected by the data, and the dimension of the neural network is proportional to the size of the data volume to ensure the accuracy of the prediction of the neural network; 1.4), carry out the training of neural network; Use gradient descent method to adjust the weight and bias of neural network, realize the training of neural network by traversing each layer of neural network multiple times; 1.5), save the optimal training result; after the neural network training is finished, the structure, weight and bias of the neural network should be saved, so as to be used for the real-time prediction of online traffic flow in step 2); 2), GMDH neural network online traffic flow real-time prediction 2.1), read the historical data of vehicle traffic; the historical data is collected according to the smart camera on the vehicle road, including the speed of the passing vehicle and the license plate number of the vehicle; 2.2) Determine whether the smart camera is working normally and whether the collected vehicle data is correct; exclude excessive speed, incorrect recognition of the vehicle license plate number, discrepancies between the front and rear driving trajectories of the vehicle, and ghost vehicle data; 2.3), initialize the GMDH neural network according to the neural network structure determined in step 1); 2.4), carry out the prediction of neural network; Use the structure of the neural network obtained in previous step 1), weight and bias to predict the real-time traffic flow; 2.5), the output result is used to control the traffic situation. 2. the traffic vehicle flow prediction method based on GMDH neural network according to claim 1, is characterized in that: described step 1.2) specifically comprises the steps: 1.2.1), judging the number n of vehicles passing per unit time, should meet the constraints: n _nin <n<n _max , n _nin and n _max are the minimum and maximum numbers of vehicles passing per unit time, if this constraint cannot be satisfied, then the data There is a problem in the collection process; 1.2.2), judging that the number n ₁ of vehicles at this intersection meets the constraints with the number n _i of vehicles at other intersections at the current time: a·n _i <n ₁ <b·n _i , where a and b are corresponding coefficients, if not Satisfying the current constraints indicates that there is a problem in the data collection process; 1.2.3), other methods of verifying the collected data are also included in the scope of this patent requirement; 1.2.4), if it is judged that the data is inaccurate, use the corresponding data restoration plan to complete the data, and infer the corresponding missing data from the historical data; 1.2.5), if it cannot be inferred from the data, an error will be reported and the system will be required to maintain the sensor module. 3. the traffic vehicle flow prediction method based on GMDH neural network according to claim 1, is characterized in that: described step 1.3) specifically comprises the steps: 1.3.1), the number of days D to read the training data is used to measure the dimension of the data and determine the scale of the neural network; 1.3.2), read the sampling period T of the training data, and calculate the number N of data collected every day, and determine the scale of the neural network; 1.3.3), determine the input neuron number C _in =N of the neural network, and the output neuron number C _out =N of the neural network, determine the intermediate layer neuron number C _min =2N of the neural network; 1.3.4), determine the number of neuron layers of the neural network, L=2+(C _min +C _out ) ^0.5 ; 1.3.5), save the corresponding neural network information for use in initializing the neural network in 1.4). 4. the traffic vehicle flow prediction method based on GMDH neural network according to claim 1, is characterized in that: described step 1.4) specifically comprises the steps: 1.4.1), initialize the neural network; according to the neural network information saved in 1.3), establish the neural structure, initialize the weight W of the neural network, the bias equivalent value B, initialize the layer number pointer P _f =1, and the number of learning times P _s = 1; 1.4.2), obtain the output of current layer; Use I, B to calculate the output matrix O of current layer, formula is O=I*W+B; 1.4.3), record this layer output data O, be used for the study of step 1.3.4); 1.4.4), judge whether the operation reaches the last layer, if not, add 1 to the number of training layers, return to step 1.4.2), if reached, enter step 1.4.5); 1.4.5), judge whether the data reaches the number of training times, if not, add 1 to the number of training times, and at the same time, the layer pointer returns to 1, and returns to step 1.4.2); if it reaches step 1.4.6); 1.4.6), save the training results. 5. the traffic vehicle flow prediction method based on GMDH neural network according to claim 1, is characterized in that: described step 2.2) specifically comprises the steps: 2.2.1), judging the number n of vehicles passing per unit time, should meet the constraints: n _nin <n<n _max , the selected coefficients of n _nin and n _max , if the current constraints cannot be satisfied, it means that there is a problem in the data collection process; 2.2.2), judging that the number n ₁ of vehicles at this intersection should satisfy the constraint relationship with the number n _i of vehicles at other intersections at the current time: a·n _i <n ₁ <b·n _i , a and b are corresponding coefficient, if the current constraint cannot be satisfied, it means that there is a problem in the data collection process; 2.2.3), other methods of verifying the collected data are also included in the scope of this patent requirement; 2.2.4), if it is judged that the data is inaccurate, use the corresponding data restoration plan to speculate the corresponding lost data from the historical data; 2.2.5), if it cannot be inferred from the data, then report an error to the use of this method and require the sensor module to be maintained. 6. the traffic vehicle flow prediction method based on GMDH neural network according to claim 1, is characterized in that: described step 2.3) specifically comprises the steps: 2.3.1), the continuous number of days D to read the training data; 2.3.2), read the sampling period T of the training data, and calculate the number N of data collected every day through the formula, which is used to measure the dimension of the data read into the scale of the neural network; 2.3.3), the number of input neurons C _in =N read into the neural network, and the number of output neurons C _out =N of the neural network, the number of intermediate layer neurons C _min =2N read into the neural network; 2.3.4), the number of layers of neurons read into the neural network, L=2+(C _min +C _out ) ^0.5 ; 2.3.5), save the corresponding neural network information for use in initializing the neural network in 1.4). 7. the traffic vehicle flow prediction method based on GMDH neural network according to claim 1, is characterized in that: described step 2.4) specifically comprises the steps: 2.4.1), read the network structure, read the weight W of the neural network saved in 1.4), the bias equivalent value B, and set the layer number pointer to 1; 2.4.2), obtain the output of the current layer; take the input matrix I of the current layer as raw material to calculate the output matrix O of the current layer, the formula is O=I*W+B; 2.4.3), record this layer output data O, for the prediction of step 2.3.4); 2.4.4), judge whether the operation reaches the last layer, if not, add 1 to the number of training layers, return to step 1.4.2), if reached, enter step 1.4.5); 2.4.5), save the prediction result.