CN110070715A - A kind of road traffic flow prediction method based on Conv1D-NLSTMs neural network structure - Google Patents

Abstract

A kind of road traffic flow prediction method based on Conv1D-NLSTMs neural network structure, comprising the following steps: 1), construct the traffic flow data matrix of associated road, and data are pre-processed；2) road traffic flow space-time characteristic, is extracted based on traffic flow data matrix；3) road traffic flow prediction model, is constructed based on road traffic flow space-time characteristic: regression forecasting is made to obtained road traffic flow space-time characteristic using full articulamentum, future time instance is obtained without the prediction result of the road traffic flow of renormalization, and model parameter is continued to optimize using back-propagation algorithm according to the result of mean square error, result is mapped as actual traffic flow value eventually by renormalization；4) it, verifies road traffic flow prediction model: the road traffic flow data in test set being predicted using the model that training is completed, compare prediction result and actual value to test model performance.Prediction result of the present invention is more acurrate according to the experimental results.

Description

A Road Traffic Flow Prediction Based on Conv1D-NLSTMs Neural Network Structure method

technical field

The invention belongs to the field of traffic prediction, and relates to a road traffic flow prediction method based on a Conv1D-NLSTMs neural network structure.

Background technique

With the rapid development of social economy, the domestic car ownership has increased significantly, but the accompanying traffic jam problem has also brought great inconvenience to people's travel. The emergence of intelligent transportation systems analyzes various traffic data to realize real-time, accurate and efficient management and regulation of ground traffic, which relieves the pressure of ground traffic to a certain extent. As a part of the intelligent transportation system, road traffic flow prediction can predict the traffic state in the future and play an important role in the real-time management of traffic.

The existing road traffic flow prediction methods are mainly based on mathematical statistics, machine learning prediction models, or a single combination of several types of models. Although certain prediction effects can be achieved, there are still some limitations, and traffic flow data are often ignored. some of the features in .

SUMMARY OF THE INVENTION

In order to overcome the shortcomings of the low precision of the existing road traffic prediction methods, the present invention provides a road traffic flow prediction method based on the Conv1D-NLSTMs neural network structure. At the same time, the traffic flow data from a long time ago will also have a certain impact on the traffic flow in the future. The NLSTMs neural network in this method can enhance the feature extraction in this aspect.

The technical scheme adopted by the present invention to solve its technical problems is:

A road traffic flow prediction method based on Conv1D-NLSTMs neural network structure, including the following steps:

1), construct the traffic flow data matrix of the associated road, and preprocess the data: select the traffic flow data of the predicted road and its associated road, construct the traffic flow data matrix, and normalize it;

2) Extract the spatiotemporal features of road traffic flow based on the traffic flow data matrix: use a one-dimensional convolutional network to extract spatial features from the traffic flow data of different road sections at the same time in the traffic flow data matrix to obtain sequence data with spatial features, and then use NLSTMs The neural network extracts the time series features in the sequence data, so as to obtain the spatiotemporal features of road traffic flow;

3) Construct a road traffic flow prediction model based on the spatio-temporal characteristics of road traffic flow: use the fully connected layer to perform regression prediction on the obtained spatio-temporal characteristics of road traffic flow, and obtain the prediction result of road traffic flow without de-normalization in the future time. According to the result of the mean square error, the model parameters are continuously optimized by the back-propagation algorithm, and finally the result is mapped to the actual traffic flow value through inverse normalization;

4) Verify the road traffic flow prediction model: use the trained model to predict the road traffic flow data in the test set, and compare the predicted results with the actual values to test the performance of the model.

The technical idea of the present invention is: mainly use one-dimensional convolutional network (Conv1D) and Nested LSTMs (NLSTMs) neural network to extract features from the spatial relationship of road traffic flow and the time series of traffic data. Conv1D extracts road traffic The spatial features of flow data, and NLSTMs neural network is similar to embedding one or more LSTM neural networks inside the LSTM neural network structure, which can extract time series features more effectively than LSTM neural network, thereby improving the prediction accuracy of road traffic flow

The beneficial effects of the present invention are as follows: the one-dimensional convolutional neural network can effectively extract spatial features from the traffic flow data of multiple associated roads, and the NLSTMs neural network strengthens the consideration of the influence of the traffic flow of a longer time ago on the traffic flow in the future, Mining the changing trend of road traffic flow from multiple angles makes the prediction result more accurate.

Description of drawings

Figure 1 is the structure diagram of NLSTMs neural network;

Figure 2 is the structure diagram of the Conv1D-NLSTMs neural network model;

Figure 3 shows the traffic flow prediction results based on the Conv1D-NLSTMs neural network model.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Referring to Figures 1 to 3, a road traffic flow prediction method based on the Conv1D-NLSTMs neural network structure includes the following steps:

1), construct the traffic flow data matrix of the associated road, and preprocess the data, the process is as follows:

First, assuming that the traffic flow of road k is predicted, the traffic flow data of this road and its associated roads are selected to construct a traffic flow data matrix X', in the following form:

Among them, the column vector represents the traffic flow sequence data of m different road sections respectively, the row vector represents the traffic flow data of different roads at n times, then _x'it is the original traffic flow data of road i at time t;

According to the time series of road traffic flow, the data is divided into training data and test data, and the division ratio is 8:2.

Then, the maximum and minimum normalization method is used to normalize the road traffic flow data, and the process is as follows:

Among them, x′ _max , x′ _min are the maximum and minimum values in the original traffic flow data respectively, x _it is the traffic flow of road i at time t after preprocessing, let P _t =[x _1t ,x _2t ,...,x _mt ], representing the traffic flow state of different roads at time t, then the preprocessed traffic flow data matrix X is:

2) Extract the spatiotemporal characteristics of road traffic flow based on the traffic flow data matrix;

One-dimensional convolution operation is performed on P _t at each moment to extract the spatial characteristics of road traffic flow data. The calculation process is as follows:

s _t =f(W _p *P _t +b _p ) (4)

Among them, W _p represents the weight matrix, b _p represents the bias term, * represents the convolution operation, f represents the activation function relu:max{x,0}, and s _t represents the result obtained by the one-dimensional convolution operation, then the extracted The spatial feature of road traffic flow data is S=[s ₁ , s ₂ ,...,s _n ] ^T ;

Before the spatial features are input into the NLSTMs neural network, the spatial feature S needs to be formally transformed, and the transformation form is as follows:

Among them, d represents taking the first d continuous data of road traffic flow to predict the traffic flow at the next moment, S′=[x _d ,x _d+1 ,…,x _n-1 ] ^T ,n≥d, represents the spatial feature The transformed result, the input of the NLSTMs neural network is x _t , d≤t≤n-1, and the number of samples is nd;

The NLSTMs neural network is divided into two parts, the inner and outer parts, and the update and gating mechanism of the state of the outer LSTM neural network unit is expressed as the following equation:

i _t =σ(x _t W _xi +h _t-1 W _hi +b _i ) (6)

f _t =σ(x _t W _xf +h _t-1 W _hf +b _f ) (7)

o _t =σ(x _t W _xo +h _t-1 W _ho +b _o ) (9)

h _t =o _t ·σ(c _t ) (10)

Among them, · represents the dot product, σ() represents the sigmoid function, W _xf , W _xi , W _xo represent the input weight matrix of the external forget gate, input gate, and output gate, W _hf , W _hi , W _ho represent the external forget gate, The output weight matrix of the input gate and the output gate at the previous moment, b _f , b _i , and b _o represent the bias matrix of the external forget gate, input gate, and output gate, i _t , f _t , c _t , o _t , h _t represents the output of the external input gate, forget gate, cell state, output gate, and memory cell, represents the output of the internal memory unit;

The calculation formula of the internal embedded LSTM neural network unit is similar to the calculation formula of the state update and gating mechanism of the LSTM neural network, and the expression is as follows:

Among them, · represents the dot product, σ(·) represents the sigmoid function, represents the internal input, W _xc represents the weight matrix of the internal input, W _hc represents the weight matrix of the state unit at the previous moment of the internal input, b _c represents the bias matrix of the internal input, represents the input weight matrix of the internal forget gate, input gate, state unit, and output gate, represents the previous output weight matrix of the internal forget gate, input gate, state unit, and output gate, represents the bias matrix of the internal forget gate, input gate, state unit, and output gate, Represents the output of internal input gates, forget gates, cell states, output gates, and memory cells. Therefore, the final output of the NLSTMs neural network, that is, the spatiotemporal characteristics of road traffic flow is H=h _t ;

3) Build a road traffic flow prediction model based on the spatiotemporal characteristics of road traffic flow. The process is as follows:

First, take the spatiotemporal feature H of road traffic flow as the input of the fully connected layer, and predict the next moment of traffic flow y _t (without denormalization) based on the input traffic flow data. The fully connected expression is as follows:

y _t = W _h · H (18)

Among them, W _h is the weight matrix of the fully connected layer, and the true value corresponding to y _t Y _t =x _{k, t+1} , d≤t≤n-1;

Then define the mean squared error as the loss function L:

Calculate the loss function L of the model, and then use the back-propagation algorithm to continuously optimize the model parameters. The gradient calculation and parameter update in the back-propagation algorithm are implemented by the Adam optimizer;

Finally, by de-normalizing the output y _t of the fully connected layer, the actual traffic flow prediction value can be obtained;

4) Verify the road traffic flow prediction model

Validate the model with test data and compare the predicted results with the actual values. In this experiment, absolute mean square error (MAE) and root mean square error (RMSE) are selected as the indicators of road traffic flow prediction accuracy. The calculation formulas are as follows:

Among them, b is the number of samples, Y _o is the actual traffic flow, and Y _o ′ is the predicted flow output by the model.

Example: The data in the actual experiment, the prediction process is as follows:

1) Select experimental data

The original road traffic flow data includes the traffic flow data of 3 roads for 29 days. The road traffic flow data is the traffic flow data of some sections of the Second Ring Road in Beijing, and the sampling interval T is 2min. The road traffic flow data of the first 23 days of these three roads are used as training data to train model parameters, and the road traffic flow data of the last 6 days are used as test data to verify the model performance.

2) parameter determination

The experiment of the present invention is implemented based on the tensorflow environment, and keras is used to complete the construction of the entire experimental model framework. The one-dimensional convolution process is implemented by the Conv1D function in keras, the NLSTMs neural network is implemented by the NestedLSTM layer, and the fully connected layer is implemented by the Dense function. . Therefore, the parameters of the whole experiment are set as follows: the number of layers of one-dimensional convolution is 1, the size of the input matrix is 10x3 (the total number of roads is 3, and the traffic data of the previous 10 moments is predicted, that is, d=10, the length of the convolution kernel is 1, The number of convolution kernels is 3, the filling method is "padding" and the activation function is relu:max{x,0}; the output units of the NestedLSTM layer are all 64, and the number of layers is set to 2; the number of output units of the fully connected layer is 1, that is, the prediction traffic flow at the next moment.

3) Experimental results

In the experiment, the traffic flow of these three roads was predicted respectively, and the method was compared with LSTM neural network and NLSTMs neural network method. The statistical analysis of the results is shown in Table 1:

Table 1.

Claims (5)

1. a kind of road traffic flow prediction method based on Conv1D-NLSTMs neural network structure, which is characterized in that the side Method the following steps are included:

(1), the traffic flow data matrix of associated road is constructed, and data are pre-processed: selection prediction road and its correlation Join the traffic flow data of road, constructs traffic flow data matrix, and it is normalized；

(2), road traffic flow space-time characteristic is extracted based on traffic flow data matrix: using one-dimensional convolutional network to traffic flow data The traffic flow data of synchronization different sections of highway extracts space characteristics in matrix, obtains the sequence data with space characteristics, then The temporal aspect in the sequence data is extracted using NLSTMs neural network, to obtain road traffic flow space-time characteristic；

(3), road traffic flow prediction model is constructed based on road traffic flow space-time characteristic: using full articulamentum to obtained road Traffic flow space-time characteristic makees regression forecasting, obtains future time instance without the prediction result of the road traffic flow of renormalization, and Model parameter is continued to optimize using back-propagation algorithm according to the result of mean square error, maps result eventually by renormalization For actual traffic flow value；

(4), verify road traffic flow prediction model: using training complete model to the road traffic flow data in test set into Row prediction compares prediction result and actual value to test model performance.

2. a kind of road traffic flow prediction method based on Conv1D-NLSTMs neural network structure as described in claim 1, It is characterized in that, the process of the step (1) is as follows:

First, it is assumed that predicting the magnitude of traffic flow of road k, then the magnitude of traffic flow number of the road and its associated road is chosen According to for constructing traffic flow data matrix X ', form is as follows:

Wherein, column vector respectively represents the traffic flow sequence data of m different sections of highway, and row vector represents the n moment of different roads Traffic flow data, then x '_itFor the original traffic flow data of t moment road i；

According to the time series of road traffic flow, training data and test data, division proportion 8:2 are splitted data into；

Then road traffic flow data is normalized using maxmin criterion method, process is as follows:

Wherein, x '_max, x '_minMaximin respectively in original traffic flow data, x_itFor t moment road i after pretreatment The magnitude of traffic flow enables P_t=[x_1t,x_2t,…,x_mt], indicate the traffic flow modes of t moment difference road, then by pretreated Traffic flow data matrix X are as follows:

3. a kind of road traffic flow prediction side based on Conv1D-NLSTMs neural network structure as claimed in claim 1 or 2 Method, which is characterized in that in the step (2), to the P at each moment_tMake one-dimensional convolution operation, extracts road traffic flow data Space characteristics, calculating process are as follows:

s_t=f (W_p*P_t+b_p) (4)

Wherein, W_pIndicate weight matrix, b_pIndicating bias term, * indicates that convolution algorithm, f indicate activation primitive relu:max { x, 0 }, s_tIndicate that one-dimensional convolution algorithm obtains as a result, the space characteristics of the road traffic flow data then extracted are S=[s₁,s₂,…, s_n]^T；

It before space characteristics are input to NLSTMs neural network, needs to make space characteristics S formal transformation, converts Form is as follows:

Wherein, d indicates to take d continuous data before road traffic flow to predict the magnitude of traffic flow of subsequent time, S '=[x_d, x_d+1,…,x_n-1]^T, n >=d, it is after representation space eigentransformation as a result, then the input of NLSTMs neural network be x_t,d≤t≤n- 1, and its sample number is n-d；

NLSTMs neural network is divided into inside and outside 2 parts, the update of external LSTM neural network location mode and door control mechanism It is expressed as following equation:

i_t=σ (x_tW_xi+h_t-1W_hi+b_i) (6)

f_t=σ (x_tW_xf+h_t-1W_hf+b_f) (7)

o_t=σ (x_tW_xo+h_t-1W_ho+b_o) (9)

h_t=o_t·σ(c_t) (10)

Wherein, dot product is indicated, σ () indicates sigmoid function, W_xf、W_xi、W_xoIt indicates external and forgets door, input gate, out gate Input weight matrix, W_hf、W_hi、W_hoIndicate the external previous moment output weight matrix for forgeing door, input gate, out gate, b_f、b_i、b_oIndicate the external bias matrix for forgeing door, input gate, out gate, i_t、f_t、c_t、o_t、h_tIt indicates external input door, forget The output of door, location mode, out gate, memory unit,Then indicate the output of memory internal unit；

The state of its calculation formula for being internally embedded LSTM neural network unit and LSTM neural network updates and door control mechanism meter Calculation formula is similar, and expression formula is as follows:

Wherein, dot product is indicated, σ () indicates sigmoid function,Indicate internal input, W_xcIndicate the internal weight inputted Matrix, W_hcIndicate the internal previous moment state cell weight matrix inputted, b_cIndicate the internal bias matrix inputted,Indicate the internal input weight matrix for forgeing door, input gate, state cell, out gate, Indicate the internal previous moment output weight for forgeing door, input gate, state cell, out gate Matrix,Indicate the internal bias matrix for forgeing door, input gate, state cell, out gate,The output for indicating internal input gate, forgeing door, location mode, out gate, memory unit, therefore The final output of NLSTMs neural network, i.e. road traffic flow space-time characteristic are H=h_t。

4. a kind of road traffic flow prediction side based on Conv1D-NLSTMs neural network structure as claimed in claim 1 or 2 Method, which is characterized in that the process of the step (3) is as follows:

Firstly, inputting road traffic flow space-time characteristic H as full articulamentum, lower a period of time based on input traffic flow data is predicted Carve magnitude of traffic flow y_t, it is as follows that expression formula is connected entirely:

y_t=W_h·H (18)

Wherein, W_hFor the weight matrix of full articulamentum, and y_tCorresponding true value Y_t=x_{K, t+1},d≤t≤n-1；

Then defining mean square error is loss function L:

Then the loss function L of computation model continues to optimize model parameter using back-propagation algorithm realization, backpropagation Gradient in algorithm is calculated to be realized by Adam optimizer with parameter update；

Finally, by the output y of full articulamentum_tMake renormalization operation, actual traffic flow forecasting value can be obtained.

5. a kind of road traffic flow prediction side based on Conv1D-NLSTMs neural network structure as claimed in claim 1 or 2 Method, which is characterized in that in the step (4), choose absolute value mean square deviation MAE, root-mean-square error RMSE as road traffic flow The index of precision of prediction, calculation formula difference are as follows:

Wherein, b is sample number, Y_oFor actual traffic flow, Y_o' the predicted flow rate exported for model.