CN105701571A - Short-term traffic flow prediction method based on nerve network combination model - Google Patents

Abstract

The invention provides a short-term traffic flow prediction method based on a neural network combination model. The method constructs a backpropagation neural network combination prediction model, and proposes a short-term traffic flow prediction method based on the model. According to the characteristics of the traffic flow, the present invention first uses the fuzzy C-means clustering algorithm to cluster the traffic flow, builds a backpropagation neural network prediction model for each cluster generated by the clustering, and calculates the prediction of each prediction model according to the degree of membership The weighted sum of the results is used as the final prediction result. In order to improve the prediction accuracy, the present invention adopts the Taguchi method for experimental design to test the influence of different structural parameters on the prediction accuracy of the prediction model, and uses the optimal structural parameters as the initial structure of the prediction model. The method of the invention can effectively improve the prediction accuracy of short-term traffic flow, reduce the influence of noise in training data on the prediction accuracy, and the running time is reasonable.

Description

A Short-term Traffic Flow Forecasting Method Based on Neural Network Combination Model

technical field

The invention relates to a short-term traffic flow prediction method, which uses a neural network combination model to improve the prediction accuracy of short-term traffic flow and reduces the influence of noise in training data on the prediction accuracy, and belongs to the cross-technical application field of traffic flow and neural network.

Background technique

Intelligent transportation system is a real-time, accurate and efficient transportation management system, which effectively integrates advanced information technology, data communication technology, electronic control technology and computer processing technology. The core technology of intelligent transportation system——traffic control technology and traffic guidance technology is the most effective way to solve urban traffic congestion and improve the efficiency of road network traffic, and it is a research hotspot in recent years. The basis for realizing traffic control and traffic guidance is real-time and accurate short-term traffic flow forecasting.

Traffic flow forecasting can be divided into long-term traffic forecasting and short-term traffic forecasting according to the length of forecasting time. Long-term flow forecasting provides traffic flow forecasts for months or even years, and is mainly used to construct long-term planning for transportation systems. Short-term traffic forecasting focuses on predicting short-term changes in traffic flow, generally predicting the traffic in the next 5 minutes. Short-term traffic flow is highly nonlinear, time-varying, and uncertain. This uncertainty may come from environmental factors, such as road conditions, weather changes, etc., or from emergencies, such as traffic accidents, large gatherings, etc. These factors make it difficult to accurately predict short-term traffic flow.

For the prediction of short-term traffic flow, researchers have proposed a variety of models. Early forecasting models, such as exponential smoothing models, time series models, and historical average models, were mainly based on calculus and mathematical statistics, which had problems such as large demand for historical data and low forecasting accuracy. Afterwards, some modern scientific techniques and methods were gradually introduced into the prediction, and some models with better prediction effect appeared, such as Kalman filter model, non-parametric regression model, gray system prediction model and neural network model, etc.

Artificial neural network is a highly nonlinear dynamical system, and it has strong nonlinear fitting ability, so it has specific advantages in traffic flow prediction. The backpropagation neural network (BP neural network) model is one of the most widely used prediction models in traffic flow forecasting. The backpropagation neural network with a hidden layer can approximate any continuous function with arbitrary precision. However, the backpropagation neural network prediction model also has many shortcomings, such as slow convergence speed, easy to fall into local minimum, and prediction accuracy affected by its structure, etc. Therefore, the prediction accuracy of the backpropagation neural network prediction model still needs to be improved. In order to further improve the predictive ability of the neural network, researchers combine it with other intelligent methods or statistical methods to construct a combined forecasting model. These combined models tend to have better predictive accuracy than single models.

Clustering is a process of dividing a set of data objects into subsets, and each subset is a cluster, so that objects in a cluster are similar to each other, but not similar to objects in other clusters. Clustering methods can be roughly divided into four categories: partition methods, hierarchical methods, density-based methods, and grid-based methods. K-means is a typical division method, which needs to specify the number of clusters in advance, and the division of clusters is mutually exclusive. However, not all objects definitely belong to one cluster, and can also belong to several clusters at the same time, so the concept of fuzzy set can be used in clustering, which gave birth to the fuzzy clustering method. Fuzzy c-means is a typical fuzzy clustering method. Each object belongs to different clusters with a certain degree of membership. It is also based on division, and the number of clusters needs to be specified.

Taguchi method is a low-cost, high-efficiency quality engineering method, which emphasizes that the improvement of product quality is not through inspection, but through design. The idea is mainly to build quality into the product by selecting design parameters around the set target value in the initial stage of product development and design, and minimizing variation through experiments, so that all products produced have the same, stable High quality, greatly reducing costs and losses. In the field of quality control, the Taguchi method has been successfully applied to design reliable and high-quality products at low cost, such as automobiles and consumer electronics.

Contents of the invention

Technical problem: the purpose of this invention is to provide a kind of short-term traffic flow forecasting method based on neural network combination model, this method solves the prediction accuracy of short-term traffic flow is low, the noise in the training data has a great impact on the prediction accuracy, and the running time is not long. Reasonable and other issues.

Technical solution: The present invention will build a backpropagation neural network combination forecasting model, and on the basis of the backpropagation neural network combination forecasting model, a short-term traffic flow forecasting algorithm TFBCM is proposed. First, for the characteristics of traffic flow, use The fuzzy C-means clustering algorithm clusters the traffic flow, builds a backpropagation neural network prediction model for each cluster generated by clustering, and calculates the weighted sum of the prediction results of each prediction model according to the degree of membership as the final prediction result. In order to further improve the prediction accuracy, the Taguchi method was used for experimental design to test the influence of different structural parameters on the prediction accuracy of the prediction model, and the optimal structural parameters were used as the initial structure of the prediction model.

The short-term traffic flow prediction method based on the neural network combination model of the present invention, the method constructs a kind of neural network combination prediction model, utilizes the characteristics of the traffic flow, and uses the fuzzy C-means clustering algorithm to divide the traffic flow into a small number of patterns, And build a neural network prediction model for each traffic pattern.

The short-term traffic flow prediction method based on the neural network combination model includes the following steps:

Step 1) Set the structural parameters of the short-term traffic flow forecast, the specific steps are as follows:

Step 1.1) Use the number of clusters as factor A, make factor A represent the classification of traffic flow, set the level of factor A, take the lowest level of factor A as 3, the highest level as 5, and the middle level as 4. The level represents the amount taken by each factor or the state value it is in.

Step 1.2) The number of input nodes is used as factor B, and the three levels of factor B are set to 1, 3, and 5 respectively. These three levels successively represent three situations: fewer nodes, normal nodes, and more nodes;

Step 1.3) Use the number of hidden layer nodes as factor C, and factor C is determined by the empirical formula It is determined that l is the number of hidden layer nodes, n is set as factor B, that is, n is the number of input nodes, m is the number of output nodes, and δ is a constant with a value between 1 and 10. The three levels of Factor C are set to 2, 3 and 7.

Step 1.4) The learning rate is used as the factor D, the value of the factor D is between 0.01 and 0.1, and its three levels are set as 0.02, 0.04 and 0.07.

Step 1.5) Use the Taguchi method to conduct experiments to obtain the levels of factor A, factor B, factor C, and factor D when the effect value is the highest, and use these levels as the initial values of the structural parameters. In the Taguchi method, using To evaluate the effect value of each design factor, the k is the number of times each group of experiments is performed, e _MRE is the average relative error, is the predicted value of traffic flow, θ(T) is the actual value of traffic flow input by the user, T is a certain forecast time point, N is the total number of forecast time points, p and q are variables. The Taguchi method is a quality engineering method that makes all products produced have the same, stable quality by selecting design parameters around set target values and minimizing variation through experiments.

Step 2) Cluster the traffic flow. The user inputs the data set X={x ₁ ,x ₂ ,...x _n }, the threshold ε and the maximum number of iterations maxt composed of data objects; set the cluster center composed of categories as V={v ₁ , v ₂ ,...v _c }, u _ij is the degree of membership that the j-th data object x _j belongs to the i-th category v _i , all u _ij form the membership degree matrix U, and the objective function J is set as the weighted degree of membership The sum of squared distances, n is the total number of data, c is the number of clusters, m is a weighting parameter not less than 1, d _ij is the distance between the object x _j and the cluster center v _i , i and j are variables, Set the number of iteration steps t. The specific steps of step 2) are as follows:

Step 2.1) Set c as the level of factor A obtained in step 1), the user inputs m, assigns a random number between 0 and 1 to u _ij , and all u _ij form the initial membership matrix U, so that it satisfies Set the number of iteration steps t to 0.

Step 2.2) Use Calculate the current value of each element of the cluster center V, the is the value of u _ij at the tth iteration, Indicates the value of v _i at the tth iteration.

Step 2.3) Use Calculate J ^t , the J ^t is the value of J at the t-th iteration, d _ij is the Euclidean distance between the data object x _j and the classification v _i , and the Euclidean distance is in the space The true distance between two points.

Step 2.4) utilize The membership degree matrix U is corrected according to the distance from the object to the cluster center, the is the value of u _ij after the t-th iteration, and d _kj is the Euclidean distance between the data object x _j and the classification v _k .

Step 2.5) Use Calculate J ^t+1 , where J ^t+1 is the value of J after the t iteration.

Step 2.6) Judging whether |J ^t+1 -J ^t |≤ε or t is equal to maxt, if satisfied, go to step 3), otherwise t=t+1, go to step 2.2).

Step 3) Use the backpropagation neural network for training. The backpropagation neural network is also called BP neural network, which is a multi-layer feedforward neural network trained according to the error backpropagation algorithm. The specific steps of step 3) are as follows:

Step 3.1) Use the level value of factor B obtained in step 1) as the number n of network input layer nodes, the level value of factor C as the number l of nodes in the hidden layer, use the number m of nodes in the output layer, and the level value of factor D as the learning rate λ, the user Given the connection weights ω _rs and ω _su between the input layer, hidden layer and output layer neurons to set the initial value, the user then enters the initial hidden layer threshold value a ₀ and output layer threshold value b ₀ , and resets The maximum number of iterations maxt, set the number of iterations t to 0.

Step 3.2) According to the data set variable X input by the user, use the formula Calculate hidden layer output H _s , s=1,2,...,l;

Step 3.3) utilize Calculate the predicted output O _u of the backpropagation neural network, u=1, 2, . . . , m.

Step 3.4) The user inputs the desired output Y _u , and uses e _u =Y _u -O _u to calculate the prediction error e _u , where u=1,2,...,m.

Step 3.5) utilize ${\mathrm{\ω}}_{r\mathrm{the\; s}}={\mathrm{\ω}}_{r\mathrm{the\; s}}+{\mathrm{\λH}}_{\mathrm{the\; s}}(1-h_{\mathrm{the\; s}})x_r\underset{k=1}{\overset{m}{\Σ}}{\mathrm{\ω}}_{ju}{e}_u,$ and ${\mathrm{\ω}}_{\mathrm{su}}={\mathrm{\ω}}_{\mathrm{su}}+{\mathrm{\λH}}_{\mathrm{the\; s}}{e}_u,$ Bring prediction error _u into update network connection weights ω _rs and ω _su , r=1,2,...,n,s=1,2,...,l,u=1,2,.. ., m.

Step 3.6) utilize and b _u = b _u + e _u , bring in the prediction error e _u to update the network node thresholds a _s and b _u , s=1,2,...,l, u=1,2,...,m .

Step 3.7) Judging whether iteration t has reached the maximum number of iterations maxt, if so, then go to step 4), otherwise t=t+1, go to step 3.2).

Step 4) Carry out short-term traffic flow prediction. The user inputs the prediction time point T in sequence, and from the prediction output of the backpropagation neural network obtained in step 3), take c prediction values O ₁ , O ₂ ,...,O _c ; according to step 3) the membership degree matrix U For the membership degree of each cluster at the prediction time point, use the membership degree as the weight, and find the weighted sum of the predicted values of the backpropagation neural network corresponding to each cluster as the predicted value O _T at the prediction time point T, that is The T is a prediction time point, and T=1, 2, . . . , n.

Beneficial effects: the present invention solves the problem of short-term traffic flow prediction. The prediction results of each prediction model use the weighted sum of the membership degrees as the final prediction value, which alleviates the influence of noise in the training data on the prediction accuracy to a certain extent, and has the following benefits Effect:

(1) The present invention uses the Taguchi method to select appropriate structural parameters, and sets the structure of the prediction model according to these parameters, further improving the prediction accuracy.

(2) In the case of artificially specifying the structural parameters, the prediction accuracy of the present invention is significantly higher than that of the traditional backpropagation neural network algorithm and the CITFF algorithm. Compared with the genetic algorithm, using the Taguchi method to select the structural parameters can be achieved in a shorter time. reduce the prediction error more in a short period of time.

(3) The present invention has high prediction accuracy for short-term traffic flow, and the calculation time is reasonable.

Description of drawings

Figure 1 is a framework diagram of short-term traffic flow forecasting scheme.

detailed description

The present invention is aimed at short-term traffic flow prediction at crossroads, and considers the diversity of traffic flow distribution forms, model training adequacy and prediction accuracy in specific implementation, and reduces random generation or artificial designation of these parameter values. The present invention will be described in more detail below.

In the specific implementation of the present invention, a backpropagation neural network combination prediction model is first constructed, which is composed of a backpropagation neural network prediction model, a fuzzy C-means clustering model and a structural parameter selection model.

The backpropagation neural network is a multi-layer feedforward neural network, and its main feature is that the signal propagates forward and the error propagates backward. In the forward propagation, the input signal is processed layer by layer from the input layer through the hidden layer to the output layer. The state of neurons in each layer only affects the state of neurons in the next layer. If the output layer cannot get the desired output, it will switch to backpropagation, and adjust the network weights and thresholds according to the prediction error, so that the predicted output of the backpropagation neural network will continue to approach the desired output. In the specific implementation, the present invention uses a single hidden layer reverse propagation neural network to predict short-term traffic flow.

Fuzzy C-means clustering provides a way to group data points in multi-dimensional space into a specific number of groups. This method first randomly selects several cluster centers, and all data points are given a certain degree of fuzzy membership to the cluster centers, and then through The iterative method continuously corrects the cluster centers, and the optimization goal is to minimize the weighted sum of the distances from all data points to each cluster center and the membership value in the iterative process. In the specific implementation, the present invention adopts fuzzy clustering method to divide the traffic flow, so that each time point can belong to different modes with a certain degree of membership. Since the clustering accuracy hardly affects the accuracy of the prediction model, the fuzzy C-means algorithm with a relatively small amount of calculation is used for clustering to improve the overall efficiency of the algorithm.

The selection of prediction model parameters is crucial to its prediction accuracy, and experiments should be conducted to verify the influence of different parameters on the prediction model accuracy. In the experimental design, the factors represent the independent variables in the experiment, and are related factors and conditions that affect the experimental indicators, while the level represents the dosage or state of each factor. At present, there are mainly four experimental design methods: trial and error method, one-factor experiment at a time, full factorial experiment and Taguchi method. Taguchi method considers the advantages of one-time one-factor test method and full-factor test method, selects typical and representative combinations from all possible combinations, makes the test combinations evenly distributed in the global range, and can reflect the overall situation. The number of test combinations is as small as possible, so an orthogonal table is used. The present invention adopts the Taguchi method to design and conduct tests in the specific implementation, calculates the effect value of each design factor according to the test results, and selects reasonable parameters of the combined forecasting model to further improve the forecasting accuracy.

In the specific implementation of the present invention, the Taguchi method is first used to determine the structural parameters of the clustering model and the prediction model, and then the fuzzy C-means clustering algorithm is used to cluster the traffic flow, and the training data is divided according to the generated clusters, corresponding to each clusters to train a backpropagation neural network prediction model. When forecasting, all forecasting models are used to predict the forecasting time point, and according to the membership degree of each forecasting model at the forecasting time point, the weighted sum of each forecasting value is obtained as the final forecasting value. The framework of the model is shown in Figure 1.

In the specific implementation of the present invention, the short-term traffic flow prediction is carried out on the basis of the combined prediction model of the backpropagation neural network. Considering the diversity of traffic flow distribution forms, the present invention roughly divides the flow dimension, uses different prediction models to describe different flow patterns, and determines the optimal number of divisions by experiments. In order to avoid problems caused by too many patterns, the present invention generally limits the number of clusters to no more than 5 in specific implementation. Due to the uncertainty of short-term traffic flow, the present invention uses the degree of membership to indicate the possibility of each time point belonging to a certain mode, and calculates the weighted sum of the prediction results of each prediction model according to the degree of membership as the final prediction value. In order to further improve the prediction accuracy, the present invention adopts the Taguchi method to design experiments, and determines the reasonable structural parameters of the clustering model and the prediction model with a small number of experiments. Using these parameters to initialize the structure of the model can make the prediction results more accurate.

1. Selection of structural parameters

The structure of the backpropagation neural network and its initial weights and thresholds will have a certain impact on its accuracy, convergence speed and whether it will fall into a local minimum. The present invention considers the selection of the structural parameters of the backpropagation neural network in the specific implementation.

In the specific implementation of the present invention, the design factor A directly affects the output of the clustering model, and the design factors B, C and D determine the structure of the neural network prediction model. The selection of each design factor and its level is described as follows:

1) Design factor A: The number of clusters determines how many categories the fuzzy C-means algorithm needs to cluster the traffic flow into, which also determines how many prediction models need to be established. This parameter has a great influence on the prediction accuracy, and it is necessary to judge its optimal value through experiments. According to common sense, traffic flow can be divided into at least three categories according to its size, namely peak, trough and ordinary flow, so its minimum level is determined to be 3. Considering the efficiency and the problems of the CITFF algorithm, the number of clusters should not be too many, so the highest level is determined to be 5, and the second level criterion takes the middle value of 4.

2) Design factor B: The number of input nodes determines the amount of historical traffic flow information provided to the backpropagation neural network prediction model. If the number of input nodes is too small, it will not be able to provide enough information to the prediction model, resulting in a decrease in prediction accuracy; and if the number of input nodes is too large, some noise information will be introduced, which will also affect the prediction accuracy. In the specific implementation of the present invention, the three levels of the number of input nodes are set to 1, 3 and 5, which respectively cover the three situations of less input nodes, general and more input nodes.

3) Design factor C: The number of nodes in the hidden layer affects the learning ability of the neural network. If it is too small, it will not be able to generate enough connection weight combinations to satisfy the learning of several samples; if it is too large, the generalization ability of the network after learning will be affected. worse. For the determination of the number of hidden layer nodes, it is determined by the following formula: Said l is the number of hidden layer nodes, set n as factor B, that is, n is the number of input nodes, m is the number of output nodes, and δ is a constant with a value between 1 and 10. Since the number of input nodes is set to 1, 3 and 5, and the number of output nodes is fixed at 1, according to the formula, the three levels of the number of hidden layer nodes can be set to 2, 3 and 7 respectively.

4) Design factor D: The BP algorithm learns by backpropagating errors and correcting the weight threshold. The learning rate has a great influence on the convergence speed and training results. If the learning rate is too small, it will make the learning too slow; if the learning rate is too large, it may cause oscillation or divergence. According to experience, the value of the learning rate should be between 0.01 and 0.1, so the three levels of the learning rate can be set to 0.02, 0.04 and 0.07.

Since there are 4 design factors, and the level number of each design factor is 3, the orthogonal table described in Table 1 is used to design the experiment. In the specific implementation, the present invention adopts the following average relative error e _MRE to represent the error of prediction, said is the predicted value of traffic flow, θ(T) is the actual value of traffic flow input by the user, T is a certain forecast time point, and N is the total number of forecast time points. The error of the experiment conforms to the expected small characteristic, and the smaller the better, in the specific implementation, the effect value of each design factor is evaluated by using the expected small characteristic η in the Taguchi method, namely Said k is the number of executions of each group of experiments in the orthogonal table, and both p and q are variables.

In the specific implementation, the η of each group of experiments is calculated by performing experiments, and the η of the experiments corresponding to each level of a certain factor is added to obtain the average value, and the effect value of each level of the factor can be obtained. The level with the highest effect value of each factor is selected as the best combination of initial values of structural parameters.

2. Clustering of traffic flow

The present invention uses the fuzzy C-means clustering algorithm to cluster the traffic flow in the specific implementation, and the cluster number c is determined in advance by the Taguchi method design test. Given a data set X={x ₁ ,x ₂ ,...x _n }, fuzzy C-means clustering is to divide X into c classes, and the c cluster centers are V={v ₁ ,v ₂ ,. .. v _c }. In fuzzy clustering, unlike the hard clustering method, each object is not strictly divided into a certain class, but the degree of membership of each object corresponding to all classes is calculated. If u _ij is used to represent the membership degree of the j-th object belonging to the i-th class, then u _ij ∈ [0,1], and there are: That is, the sum of the membership degrees of each object for all classes is 1. The goal of fuzzy C-means clustering is still to minimize the intra-class distance and maximize the inter-class distance, so its objective function can be set to the sum of squared distances weighted by membership, that is The n is the total number of objects, c is the number of clusters, m is a weighting parameter not less than 1, and d _ij is the distance between the object x _j and the cluster center v _i , which is calculated by Euclidean distance here.

Fuzzy C-means clustering is to iteratively adjust the membership matrix U and cluster center V to minimize the objective function. The specific steps are as follows:

(2.1) Select the number of clusters c and the weighting parameter m, and initialize the membership matrix U ⁰ with a random number between 0 and 1, so that it satisfies constraints. Set the number of iteration steps t to 0.

(2.2) use Calculate the current value of each element of the cluster center V.

(2.3) use Calculate J ^t , the J ^t is the value of J at the t-th iteration, d _ij is the Euclidean distance between the data object x _j and the classification v _i , and the Euclidean distance is in the space The true distance between two points.

(2.4) Use The membership degree matrix U is corrected according to the distance from the object to the cluster center, the is the value of u _ij after the t-th iteration, and d _kj is the Euclidean distance between the data object x _j and the classification v _k .

(2.5) use Calculate J ^t+1 , where J ^t+1 is the value of J after the t iteration.

(2.6) Determine whether |J ^t+1 -J ^t |≤ε or t is equal to maxt, if satisfied, continue the following operation, otherwise t=t+1, turn to (2.2).

According to the principle of maximum membership, each object should belong to its corresponding class with the largest membership degree.

3. Back propagation neural network training

The traffic flow data is clustered to form c clusters, corresponding to c patterns of traffic flow. In order to build corresponding predictive models for each mode, the training set for training these models should first be divided. Divide the data in the original training set according to the time dimension, and for each cluster, divide all the training data whose time dimension is the same as the object in the cluster into a sub-training set, so that a total of c sub-training sets are formed, and each sub-training set is used Train a backpropagation neural network. The training process of the backpropagation neural network includes the following steps:

(3.1) Network initialization. Using the previously obtained network input layer node number n, hidden layer node number l, output layer node number m, and learning rate λ, the connection weights ω _rs and ω _su , and given the hidden layer threshold a ₀ and the output layer threshold b ₀ , reset the maximum number of iterations maxt, and set the number of iterations t to 0.

(3.2) Hidden layer output calculation. According to the input variable X, calculate the hidden layer output s=1,2,...,l.

(3.3) Output layer output calculation. Calculate the predicted output of the backpropagation neural network based on the output of the hidden layer $o_u=\underset{\mathrm{the\; s}=1}{\overset{l}{\mathrm{\Σ}}}{h}_{\mathrm{the\; s}}{\mathrm{\ω}}_{\mathrm{su}}-b_u,$ u=1,2,...,m

(3.4) Error calculation. The user inputs the expected output Y _u , and calculates the prediction error e _u =Y _u −O _u , u=1,2,...,m.

(3.5) Weight update. Bring the prediction error e _u into the updated network connection weights ω _rs and ω _su , namely ω _su =ω _su +λH _s e _u , r=1,2,...,n,s=1,2,...,l,u=1,2,...,m.

(3.6) Threshold update. According to the prediction error e _u , update the network node thresholds a _s and b _u , namely b _u =b _u +e _u , the s=1,2,...,l, u=1,2,...,m.

(3.7) Judging whether iteration t has reached the maximum number of iterations maxt, if so, then go to step 4), otherwise t=t+1, go to (3.2).

In the above-mentioned concrete implementation of the present invention, use Taguchi method to determine suitable structural parameter, for example, utilize the orthogonal table of table 1 to design test, every group of test is carried out 5 times, obtain 3 levels corresponding to four kinds of design factors respectively The effect value of each factor is selected, and the level with the highest effect value of each factor is selected as the best combination of the initial value of the structural parameters. The test results of the design factors and their levels are shown in Table 2. In the following implementation, select the traffic flow of a certain day in the training set (any day is acceptable), cluster it using the fuzzy C-means clustering algorithm, and form multiple sub-training sets, and use each sub-training set to train a counter To propagate the neural network, which in turn corresponds to different traffic patterns. In the final stage of specific implementation, according to the prediction output results of the backpropagation neural network and the membership degree matrix U, the prediction value at a specific prediction time point is calculated.

Table 1 is an example of an orthogonal table,

Table 1

Table 2 is an example of design factors.

Table 2

Claims (4)

1. the short-term traffic flow forecast method based on neural network ensemble model, it is characterised in that the step that the method comprises is:

Step 1) structural parameters of short-term traffic flow forecast are set,

Step 2) traffic flow is clustered, user inputs the data set X={x being made up of data object₁,x₂,...x_n, threshold values ε and maximum iteration time maxt；The cluster centre being set with classification composition is V={v₁,v₂,...v_c, u_ijIndicate that jth data object x_jBelong to i-th classification v_iDegree of membership, all u_ijComposition subordinated-degree matrix U, target setting function J be fuzzy set theory square distance and, n is data count, c for cluster number, m be not less than 1 weighting parameters, d_ijFor object x_jWith cluster centre v_iBetween distance, i and j is variable, arranges iterative steps t；

Step 3) use reverse transmittance nerve network to be trained, described reverse transmittance nerve network, also referred to as BP neutral net, is the multilayer feedforward neural network by Back Propagation Algorithm training；

Step 4) carry out short-term traffic flow forecast, user inputs predicted time point T successively, from step 3) the prediction output of the reverse transmittance nerve network that obtains, take c predictive value O₁,O₂,...,O_c；According to step 3) predicted time point is for the degree of membership of each bunch in subordinated-degree matrix U, by degree of membership as weights, asks the weighted sum of each bunch corresponding reverse transmittance nerve network predictive value as the predictive value O at predicted time point T_T, namelyDescribed T is predicted time point, T=1,2 ..., n。

2. a kind of short-term traffic flow forecast method based on neural network ensemble model according to claim 1, it is characterised in that described step 1) structural parameters that arrange short-term traffic flow forecast specifically comprise the following steps that

Step 1.1) number will be clustered as factors A, make factors A represent the classification of traffic flow, the level of factors A is set, the minimum level of factors A is taken as 3, the highest level is taken as 5, and middle level is taken as 4, and described level represents the consumption or the state in which value that refer to that each factor takes；

Step 1.2) using input number of nodes as factor B, three levels of factor B respectively 1,3,5 are set, these three level represents node three kinds of situations less, general, more successively；

Step 1.3) using node in hidden layer as factor C, factor C is by empirical equationDetermining, described l is node in hidden layer, and arranging n is factor B, and namely n is input number of nodes, and arranging m is output node number, and δ is value constant between 1 to 10, and three levels of factor C are set to 2,3 and 7；

Step 1.4) using learning rate as factor D, factor D value is between 0.01 to 0.1, and its three levels are set to 0.02,0.04 and 0.07；

Step 1.5) use Taguchi's method to test, it is thus achieved that factors A, factor B, factor C, the factor D level when effect value peak, using these levels initial value as structural parameters。In Taguchi's method, useAssessing the effect value of each design factor, described k is the number of times that often group test performs, e_MREIt is average relative error, For the predictive value of traffic flow, the actual value of the traffic flow that θ (T) inputs for user, T is a certain predicted time point, and N is the sum of predicted time point, and p and q is variable；Described Taguchi's method is a kind of quality engineering method, and the method is by selecting design parameter around set desired value, and reduces variation through overtesting bottom line, makes all over products produced have identical, stable quality。

3. a kind of short-term traffic flow forecast method based on neural network ensemble model according to claim 1, it is characterised in that described step 2) specifically comprise the following steps that

Step 2.1) c is set as step 1) level of factors A that obtains, user inputs m, is assigned to u with the random number between 0 to 1_ij, all u_ijForm initial subordinated-degree matrix U so that it is meetIterative steps t is set to 0；

Step 2.2) useCalculate the current each element value of cluster centre V, described inIt is u_ijValue when the t time iteration,Represent v_iValue when the t time iteration；

Step 2.3) useCalculating Jt, described Jt is the J value when the t time iteration, d_ijFor data object x_jWith classification v_iBetween Euclidean distance, described Euclidean distance is the actual distance in space between two points；

Step 2.4) utilizeSubordinated-degree matrix U is modified by the distance according to object to cluster centre, described inIt is u_ijValue after the t time iteration, d_kjFor data object x_jWith classification v_kBetween Euclidean distance；

Step 2.5) useCalculate J^t+1, described J^t+1It it is J value after the t time iteration；

Step 2.6) judge whether to meet | J^t+1-J^t|≤ε or t is equal to maxt, if it is satisfied, then forward step 3 to), otherwise t=t+1, turn to step 2.2)。

4. a kind of short-term traffic flow forecast method based on neural network ensemble model according to claim 1, it is characterised in that described step 3) specifically comprise the following steps that

Step 3.1) using step 1) the factor B level value that obtains is as network input layer nodes n, factor C level value is as node in hidden layer l, use output layer nodes m, factor D level value is as learning rate λ, and user provides the connection weights ω between input layer, hidden layer and output layer neuron_rsAnd ω_suArranging initial value, user inputs initial hidden layer threshold values a again₀, output layer threshold values b₀, reset maximum iteration time maxt, iterative steps t be set to 0；

Step 3.2) according to user input data set variable X, utilize formulaCalculate hidden layer output H_s, s=1,2 ..., l；

Step 3.3) utilizeCalculate the prediction output O of reverse transmittance nerve network_u, u=1,2 ..., m；

Step 3.4) user inputs desired output Y_u, utilize e_u=Y_u-O_uCalculate forecast error e_u, u=1,2 ..., m；

Step 3.5) utilizeWithBy forecast error e_uBring renewal network into and connect weights ω_rsAnd ω_su, r=1,2 ..., n, s=1,2 ..., l, u=1,2 ..., m；

Step 3.6) utilizeAnd b_u=b_u+e_u, bring forecast error e into_uUpdate network node threshold values a_sAnd b_u, s=1,2 ..., l, u=1,2 ..., m；

Step 3.7) judge whether iteration t has reached maximum iteration time maxt, if reaching, then forward step 4 to), otherwise t=t+1, turn to step 3.2)。