CN110503833A - A linkage control method for on-ramps based on deep residual network model - Google Patents

Abstract

The invention discloses an on-ramp linkage control method based on a deep residual network model. Firstly, the historical data of traffic flow characteristics is collected, and digital map conversion is performed after preprocessing; secondly, image data is input, and the characteristic value of traffic flow is established and trained. Forecasting model; third, collect real-time traffic flow characteristic data, convert the input model after preprocessing, and output short-term trend prediction graphs, and convert the predicted trend graphs into text data by graph-to-digital conversion; fourth, convert the Text data, use the trained prediction model to predict the road traffic characteristic value in a short time, and carry out linkage control on the traffic flow entering the main line in advance; finally, use the VB+VISSIM program to carry out the simulation evaluation and analysis of the ALINEA algorithm, and release the road conditions information. In the control method of the present invention, digital map conversion is performed after data preprocessing, and digital map conversion can extract more detailed features of two-dimensional images, reduce model training and prediction time, and improve prediction accuracy and real-time information processing speed.

Description

A linkage control method for on-ramps based on deep residual network model

technical field

The invention belongs to the field of traffic data prediction, and relates to a short-term traffic flow characteristic data prediction and traffic flow state division method, data prediction based on a deep residual network model, and a method for on-ramp linkage control using the prediction data.

Background technique

The characteristic data of road traffic flow mainly include traffic volume, traffic speed, traffic density and vehicle occupancy rate. Road traffic flow characteristic data prediction can predict the data of the next period and divide the traffic flow state, which is a necessary prerequisite for traffic management and control. Traffic management and control based on traffic flow forecasting not only facilitates travelers to make better travel plans, but also facilitates traffic management departments to make better management decisions.

Among the existing road traffic flow characteristic data prediction methods, shallow model and time series model are the most widely used at present. The shallow model cannot mine the information in the traffic flow data well, and the time series model only considers the characteristics of the traffic flow in time and ignores the influence of space. The deep residual network can not only extract temporal features based on the data trend image, but also extract spatial features. Therefore, the present invention proposes a traffic flow feature data prediction method based on the deep residual network. The temporal and spatial characteristics of the traffic flow characteristic data change trend graph are analyzed and nonlinear regression is performed, and finally the prediction of road traffic flow characteristic data is realized, and traffic management and control are carried out based on the predicted data.

With the rapid development of deep learning and artificial intelligence technology, the accuracy of road traffic flow prediction is increasing day by day. Road traffic flow prediction can assist the traffic management department to make more reasonable traffic control strategies, provide reference for vehicle owners to respond to traffic control strategies, alleviate traffic congestion, and reduce the waste of traffic resources. Road traffic flow prediction based on deep residual network model provides basic data for intelligent transportation systems and promotes the development and application of intelligent transportation systems.

purpose of invention

In order to improve the accuracy of the existing short-term traffic flow characteristic data prediction and traffic flow state division methods, and to solve the shortcomings of the traditional on-ramp control method that does not have a feedforward mechanism and a prediction mechanism, the present invention provides a deep residual network based Model short-term traffic flow characteristic data prediction and on-ramp linkage control method.

A method for on-ramp linkage control based on a deep residual network model is realized, which mainly includes the following steps:

Step 1: Collect the historical data of traffic flow characteristics, perform digital map conversion after data preprocessing, and convert the data into two-dimensional images with time sequence;

Step 2: Input image data, establish and train a prediction model based on the characteristic value of traffic flow in the deep residual network, set hyperparameters and network levels, perform parameter tuning through forward propagation and back propagation, complete model training, and deep learning The change trend of historical traffic flow characteristic value;

Step 3: Collect real-time traffic flow characteristic data, convert the preprocessed data into digital images, and convert the data into two-dimensional images with time sequence; after conversion, input the image data into the trained deep residual network prediction model for Short-term forecasting, output short-term trend forecasting graphs, and convert the forecasting trend graphs into text data by using graph number conversion;

Step 4: Using the converted text data, use the trained deep residual network prediction model to make short-term predictions of road traffic characteristic values, input the predicted data and identified traffic flow status into the on-ramp control ALINEA algorithm, and calculate the next control Periodic vehicle occupancy rate and the maximum number of queuing vehicles, and the linkage control of the traffic flow into the main line in advance;

Step 5: Use the VB+VISSIM program to perform simulation evaluation and analysis of the ALINEA algorithm, and release road condition information.

Step 1 specifically includes the following steps:

1.1 Historical data collection

The historical traffic flow data can be collected by two methods: UAV shooting video and vehicle detector detection. UAV shooting video can collect road section traffic flow data and road section traffic state changes; vehicle detector can collect road section lanes through detection and processing The characteristic value of traffic flow;

The traffic flow characteristic values are defined as traffic flow Q, average traffic speed _v , traffic density km and vehicle occupancy rate _Rt . The traffic flow characteristic values can be obtained through direct collection or indirect calculation. The vehicle occupancy rate uses the time occupancy rate index. The formula is as follows:

T is the time length of each control cycle; t _i takes the i-th vehicle to pass the section, in seconds; n is the number of vehicles passing the section within the measurement time.

1.2 Data preprocessing

The traffic flow feature values directly collected by drones and vehicle detectors inevitably have problems such as data loss, error, invalidity, and time drift. If the problem data is directly fed back to the deep residual network model application, it will cause the model Forecast error. Therefore, the collected data must be preprocessed before digital map conversion, that is, data cleaning (dataclean), including abnormal data identification, abnormal data repair, data denoising and normalization;

1.2.1 Abnormal data identification

The collected abnormal traffic flow data mainly includes three situations:

(1) Data such as traffic flow, speed and occupancy rate exceed the reasonable threshold range;

(2) The relationship between traffic flow, speed and occupancy data does not conform to the traffic flow theory;

(3) Data such as traffic flow, speed and occupancy rate are missing;

For the above types of traffic flow data, it can be directly deleted or data repaired.

1.2.2 Abnormal data repair

Traffic flow anomaly feature data mainly includes data error and data missing. For the identified data outliers, they are directly deleted; for the missing data, the K-means clustering method is used to complete the data. First, according to the Euclidean distance or correlation analysis, the K samples closest to the sample with missing data are determined, and then the K values are weighted and averaged to estimate the missing data of the sample.

1.2.3 Data denoising

Traffic flow data often contains relatively more Gaussian noise when the acquisition period is relatively short, which affects the analysis and modeling of traffic data to a certain extent. Therefore, it is necessary to perform simple denoising processing on the sampled data. ; The data denoising adopts an exponential smoothing algorithm to preserve the short-term trend of traffic flow data to the greatest extent without increasing the complexity of the algorithm; the calculation formula of an exponential smoothing method is as follows:

In the formula, and X(m) are the smoothed data and actual data at time m respectively, and k is the smoothing index, which is taken as 0.1.

1.2.4 Data normalization

When constructing a neural network model for traffic flow data, it is usually necessary to normalize the data to avoid saturation of neurons, and use the Logistic/Softmax transformation method to transform all the data to be processed into the [0,1] interval Inside.

1.3 Digitmap conversion

The preprocessed traffic flow data is converted into a two-dimensional image in time sequence, and the real-time time-traffic flow characteristic data change trend graph under different road conditions is obtained. The processed data is saved in the form of a picture and processed as a deep residual Training and testing of network models.

Step 2 specifically includes the following steps:

2.1 Data input and initialization

Image sample data input based on TensorFlow, feature set selection, different types of sample selection, sample vector map and feature map conversion, TFRecords sample data set generation, data set input 5 parts; TFRecords is a binary file format, occupying memory Small, easy to copy, move and store, no need for a separate label file;

Read the training samples in TFRecords format, and encode the data in a One-Hot way according to the sample labels;

In order to avoid over-fitting and enhance the robustness of the model, the original image data can be enhanced by image saturation, contrast conversion and other data enhancement methods, while keeping the features unchanged, by changing the pixel position and other information to obtain a more accurate image. for a sufficient sample dataset.

2.2 Hyperparameter Settings

Hyperparameters need to be set before training the deep residual network model. The main setting parameters include batch training size (Batch-size), learning rate (Learning rate), weight decay rate (weight-decay-rate), optimizer selection (optimizer )Wait;

The batch training size value determines the direction of decline, which is inversely proportional to the size of the data set; when the data set is large enough, appropriate reduction can reduce the amount of calculation; if the amount of data is small and there is noise data, it should be set larger The value of the batch training size to reduce the interference of noisy data;

The learning rate determines the magnitude of the weight update. Setting the learning rate in an appropriate range will help the model gradient drop to the optimal value; first set a larger initial learning rate setting, and gradually adjust it to Minimum learning rate for faster training and model accuracy;

Overfitting will occur during the training process of the deep residual network model. The larger the network weight, the higher the corresponding degree of overfitting. Therefore, the weight decay rate is used to set the L2 regularization parameter. _The main function is Adjust the influence of model complexity on the loss function to prevent model overfitting;

Choose the Momentum optimizer, which is mainly based on the gradient-based moving exponential weighted average. During network optimization, the loss function converges faster and the swing is smaller.

2.3 Network model level setting

The deep residual network model (ResNet) uses a 3x3 small convolution kernel mode, and replaces one large convolution kernel with multiple small convolution kernels, which reduces model parameters, increases the number of nonlinear activation functions, and reduces the amount of model calculations. The recognition error is lower. For a convolutional layer with the same input and output feature map size, the number of filters remains the same. When the size of the feature map is halved, the number of filters is doubled, and the feature map pooling step is 2 to maintain the relationship between layers. time complexity.

In the present invention, it is only necessary to identify the changing trend of the image for prediction, and there is no need to set too many network layers, so as to avoid unnecessary waste of resources. The first layer of the network is a convolutional layer, which is responsible for extracting low-level features, and the second layer is a maximum pooling layer, which reduces the estimated mean offset and retains image texture information; layers 3-6 are convolutional layers, which are responsible for extracting high-level features; The seventh layer is the average pooling layer, which suppresses the increase in the variance of the estimated value, calculates the features extracted by the convolutional layer and inputs them to the fully connected layer; the fully connected layer is responsible for the output of the classification probability.

The deep residual network consists of a set of residual blocks, each residual block contains several stacked convolutional layers, and the rectified linear unit (Relu) and batch normalization layer (BN) are attached as convolutional layers, avoiding the gradient disappears or an overflow condition occurs.

The Relu activation function introduces nonlinear characteristics into the model network, converts the input features of the model nodes into output features, and passes them to the next operation unit. Using the Relu function to make part of the output data zero, the model network can introduce sparsity by itself, and the performance of the segmented line can effectively overcome the problem of gradient disappearance.

The deep residual network model rewrites the optimal mapping as H(x)=F(x)+x, and approximating the residual function F(x) is also equivalent to approximating the optimal mapping H(x). The rewritten residual map is easier to optimize than the original optimal solution map.

By adding a "Shortcut Connections" in the feed-forward network to realize the network residual, the low-level error can be propagated to the upper layer through the shortcut during the training process, which reduces the phenomenon of gradient disappearance caused by the number of layers, and the calculation amount increases less. In this case, the model training accuracy is improved.

The shortcut skips one or more layers and merges with the main path with different step sizes, and the structural output can be expressed as

m _l+1 ＝Re lu(m _l +F(m _l ,w _l ))

In the formula, m _l and m _l+1 are the input and output of the first residual block respectively, Re lu() is the modified linear unit function, F represents the residual mapping function, and w _l is the parameter of the residual learning unit.

If the input and output dimensions are different, you need to add a linear projection To match the dimension size, after adding the linear projection, the formula is further transformed into

After batch normalization is used, as the depth of the model accelerates or the parameters change during the training process, the input data can still be distributed in a standard interval, avoiding the disappearance of the gradient, accelerating the convergence of the model, reducing the dependence of the model on the initial network weight, batch normalization The operating formula is expressed as

In the formula, x ^k is the activation degree of the neuron, E[x ^k ] represents the average value of x ^k obtained from a batch of training data sets, is the standard deviation of each batch of training data x ^k .

At the same time, the batch normalization operation introduces two learnable parameters (γ, β), which are used to reconstruct the transformed activation and restore the feature distribution learned by the original network. This operation will not destroy the features learned by the data in the previous layer operation, and will not affect the expressive ability of the network.

2.4 Forward Propagation

Forward propagation can extract the high-level features of the input image to obtain more abstract semantic features. Considering the differences between the training set and actual traffic flow data and the expressive ability of deep features, the convolutional layer features in the present invention are extracted . In the forward propagation training process, the expected learning objective function should be set first, and the function setting is:

where x is the input feature value, is the model prediction result probability, w and b are the parameters obtained from model training;

Through the feature sparse extraction of multiple convolutional layers, the mean pooling operation is used to calculate the extracted sparse convolutional features. Each batch of input sample images is converted into sparse features and entered into the fully connected layer. After logits calculation, we get The batch of sample data for each type of [batch training size × classification number] classification probability matrix;

After the softmax operation, all outputs are guaranteed to be positive, and the values of all rows of the matrix are stretched to the [0,1] interval, and the sum of the probabilities of any row is equal to 1. Softmax operates the stretched matrix, and the maximum value of each row is the value with the highest output probability, which is the prediction result of this training.

2.5 Back propagation and parameter tuning

During the training process of the deep residual model, the convolutional layer extracts each batch of sample data and calculates the sparse features layer by layer and records the corresponding parameter values. The sparse features extracted at the bottom layer are input to the logits layer to calculate the sample classification value;

The loss function of each training is calculated as the cross entropy between the true type of the sample and the model prediction result, and the training loss function of each batch of samples is calculated as follows

In the formula, t _ki is the probability that sample k belongs to category i, and y _ki is the model prediction probability that sample k belongs to category i;

By comparing the real and model prediction and recognition classification results, the model loss function is calculated, the model fitting error is backpropagated, and each parameter is continuously adjusted during the iterative process of the deep residual model, which effectively increases the robustness of the model and reduces the Probability of overfitting;

Input the training set data, tune the parameters, and select the corresponding optimizer. The commonly used mini-batch SGD training algorithm is easy to fall into local optimum and is greatly affected by the learning rate. Therefore, the Momentum optimizer based on the gradient-based moving exponential weighted average is selected to smooth the network parameters, which can solve the problem of excessive swings in the update range of the mini-batch SGD optimization algorithm, and at the same time accelerate the convergence speed of the network;

Assuming that the current iteration step is t, the calculation formula based on the Momentum optimization algorithm is as follows:

v _dw =βv _dw +(1-β)dW

v _db =βv _db +(1-β)db

W=W-αv _dw

b=b-αv _db

In the above formula, v _dw and v _db are the gradient momentum accumulated by the loss function in the previous t-1 round of iterations. β is an exponential value of gradient accumulation, which is set to 0.9; The obtained gradient, W and b are the update formulas of the network weight vector and bias vector, and α is the learning rate of the network;

After parameter tuning is completed, the last step of model training is to enter the verification set data, test the performance of the model, and manually fine-tune the learning rate and other hyperparameter values.

Step 3 specifically includes the following steps: After completing the training of the deep residual network prediction model, the short-term prediction of road traffic characteristic values can be performed based on the deep residual network model; real-time data can be collected by two methods: video shot by drone and vehicle detector. Traffic flow characteristic data, and set the phase transition judgment condition for the vehicle trajectory moving in the traffic flow: in the time interval exceeding the given threshold, the speed of the vehicle traveling along the track becomes lower or higher than the given threshold speed of the phase transition point ;

Set the data input time window, input the real-time traffic characteristic values collected by the vehicle detector, after the input, the deep residual network model preprocesses the data and converts it into a time series change graph, and outputs the traffic characteristic prediction data trend graph for the next period, The two-dimensional image of the predicted output can be converted into text data through the digit map;

Combined with the spatio-temporal characteristics of road traffic flow, the traffic flow is divided into three phases: free flow (F), synchronous flow (S) and wide motion jam (J); The step-by-step phase transition process of motion blockage (F → S → J); referring to the three-phase traffic flow theory, combined with the current situation of roads in my country, set the transition points of each phase with the speed as the threshold.

The designed model of the present invention can also divide each phase interval automatically according to the speed threshold value that is set besides carrying out short-term prediction to traffic characteristic value. The division of the phase interval of the prediction data is beneficial to the implementation of the road control in step 4.

The short-term prediction of the road traffic characteristic value described in step 4, the short-term prediction includes the prediction data of traffic flow speed, traffic flow density, and vehicle occupancy rate.

On-ramp control is a widely used and effective traffic control method to alleviate road congestion. The on-ramp control strategy based on ALINEA is simple, efficient and easy to implement. Using the ALENIA algorithm, the adjustment rate of the ramp can be controlled by controlling the duration of the red light, that is, the number of vehicles released per minute can be adjusted to achieve the purpose of controlling the flow of the entrance ramp.

When the traditional ALENIA algorithm performs queuing control, the turn-in adjustment rate of the current cycle is calculated according to the vehicle occupancy rate of the downstream section of the main line and the turn-in adjustment rate of the previous cycle, without a feedforward mechanism and a prediction mechanism. However, the present invention substitutes the traffic characteristic prediction data into the algorithm, uses the short-term prediction time window as the control period of the ALENIA algorithm, can realize more accurate road control in advance, and makes up for the shortcomings of the traditional algorithm. It not only improves the adjustment efficiency, but also effectively reduces the probability of congestion. At the same time, it can feed back the changes after road control to the model to realize re-optimization control and improve the adjustment accuracy.

The on-ramp control status is set using three indicators: the upstream section vehicle speed, the downstream lane occupancy rate, and the downstream flow of the main line. We stipulate that the speed of the upstream section is consistent with the division of traffic flow in the previous section, that is, the three-phase traffic flow theory and the on-ramp control method are linked, and the division of traffic flow phase is one of the indicators of on-ramp control.

The on-ramp control method described in step 4 specifically includes the following steps:

Before adopting the ALINEA algorithm, it is defined that the downstream vehicle occupancy rate O _out (k-1) in the k-1 control cycle is collected by the vehicle detector, and the downstream vehicle occupancy rate O _out (k) and the on-ramp vehicle arrival rate d( k) Predicted by a deep residual network model;

r(k) is calculated from O _out (k-1) data in k-1 period, then the predicted value of ramp regulation rate for k+1 period can be obtained by means of r(k) and O _out (k);

The ALINEA algorithm has a fixed green light duration, and controls the flow of vehicles entering the main line by adjusting the time interval between adjacent green lights every minute;

In a control cycle, the regulation rate calculation formula is

In the formula, r(k+1) is the ramp regulation rate calculated in the k+1th control period; r(k) is the ramp regulation rate r(k) in the kth control period, which is determined by the The adjustment rate is obtained by calculating the measured data of the rate, and the adjustment rate is the length of the green light in a control cycle, and the unit is s; the K _R parameter adjusts the fixed external disturbance in the feedback control; is the expected occupancy rate of the downstream of the main line; O _out (k) is the predicted value of the vehicle occupancy rate of the downstream of the main line in the kth control cycle.

The simulation evaluation and analysis described in step 5 is specifically to use VB+VISSIM 4.3COM to develop the program, and realize the ramp control based on the ALINEA algorithm; load the road network model in the program for simulation, and the simulation running time is 3600s, which corresponds to the actual peak hour. During the simulation process, each control cycle will return the simulation status data, including the end time of the green light, the green signal ratio and the occupancy rate;

Assuming that the signal control period is t, and the unit is s, the program obtains the occupancy rate returned by the data detector at the t-1th moment of each cycle, and the detector returns the occupancy rate every interval t-2, and according to the adjustment coefficient and the optimal occupancy Calculate the green letter ratio of the next cycle;

When the green signal ratio is greater than or equal to 0.8, the ramp signal control will continue to be green; if it is less than 0.2, the ramp signal control will continue to be red; the green signal ratio between 0.8 and 0.2 will be controlled periodically, and the ramp will be calculated according to the optimized green signal ratio Controlled green light duration.

The present invention aims at predicting the traffic characteristic value, identifying the state of road traffic flow and adopting the on-ramp control method to alleviate traffic congestion, so it is necessary to evaluate the on-ramp control effect. Because the process of embedding the vehicle detector is relatively complicated, the damage to the road surface is relatively large, before the present invention regulates and controls the road, its effect should be evaluated by simulation software.

The deep residual network has good robustness and can effectively alleviate the problem of gradient disappearance. The entrance ramp linkage control method based on the deep residual network model of the present invention collects the historical data and real-time data of traffic flow characteristics, performs digital map conversion respectively after data preprocessing, and digital map conversion can extract more detailed features of two-dimensional images, reducing Model training and prediction time, improve prediction accuracy and real-time information processing speed. It also extracts the spatio-temporal features in the traffic flow characteristic data change trend graph through convolution and performs nonlinear regression, finally realizes the prediction of road traffic flow characteristic data, and conducts traffic management and control based on the predicted data, providing basic data for intelligent transportation systems , and promote the development and application of intelligent transportation systems.

Description of drawings

Fig. 1 is the general flowchart of control method of the present invention;

Fig. 2 is the training flow diagram of deep residual network model of the present invention;

Fig. 3 is an analysis diagram of the variation trend of the characteristic value of traffic flow and the prediction error in the embodiment.

Detailed ways

The content of the present invention will be further described below in conjunction with the embodiments and the accompanying drawings, but the present invention is not limited thereto.

Referring to Figure 1-3, the on-ramp linkage control method based on the deep residual network model includes the following steps:

1. Data collection and processing

1.1 Data collection and screening

The traffic flow data collection method mainly includes UAV video shooting and vehicle detector collection. The traffic flow characteristic data collected by the vehicle detector is set as the training set for model training and parameter optimization, and the video data taken by the UAV is used as the verification set. Manual tuning of model hyperparameters.

Choose a road section with a high incidence of congestion in Nanjing. This road section has a ramp into it. It is a working day and the weather is good. According to the real-time data collected by the vehicle detector, we intercept the data with a duration of 350s, which reflects the formation process of traffic congestion and has obvious characteristics of changes.

1.2 Data preprocessing

Preprocess the intercepted data, delete and repair the traffic flow characteristic data that exceed the reasonable threshold range, do not conform to the traffic flow theory and the missing traffic flow characteristic data, use an exponential smoothing algorithm to denoise the data, and retain the short-term change trend of the traffic flow characteristic data And the Logistic/Softmax transformation method was used to normalize the data. The preprocessed traffic flow characteristic data are shown in the following table:

Table 1 Data table of real-time traffic flow characteristics

1.3 Digitmap conversion

After the data preprocessing is completed, the text information of the data is converted into the "time-vehicle number" change trend image, the "time-vehicle flow speed" change trend image, the "time-vehicle flow density" change trend image and the "time-vehicle flow density" change trend image in time series. Vehicle occupancy rate" change trend image, set the four groups of image change trend time windows to 50s, that is, predict the change trend of traffic flow characteristic data in the next 50s according to the traffic flow characteristic data in the current 50s cycle.

After the conversion of the digit map is completed, the training of the deep residual network traffic flow characteristic data prediction model is started.

2. Deep residual network traffic flow characteristic data prediction model training

2.1 Hyperparameter Settings

Hyperparameters need to be set before training the deep residual network model. The main setting parameters include batch training size (Batch-size), learning rate (Learning rate), weight decay rate (weight-decay-rate), optimizer selection (optimizer )Wait. The specific hyperparameter settings are shown in the table below.

Table 2 Hyperparameter settings of deep residual network

During model training, reducing the Batch will improve the convergence speed far less than the performance degradation caused by introducing a large amount of noise, and the GPU can perform better for Batches to the power of 2, so the Batch value used is 256.

Set a larger initial learning rate setting, and gradually adjust to the minimum learning rate as the number of model iterations increases to obtain faster training speed and model accuracy. Using the Momentum optimizer, which is based on gradient-based moving exponential weighted average, the loss function converges faster and the swing is smaller during network optimization. Set the weight decay rate of 0.0001 to adjust the influence of model complexity on the loss function to prevent model overfitting.

2.2 Data input and initialization

2.2.1 Training set and test set division

The division of training set, verification set and test set is based on the means and types of data collection.

The large amount of data collected by the vehicle detector can be used as a training set for parameter tuning of the deep residual model. The parameters can be updated by gradient descent to minimize the objective function.

The amount of data collected by drones is small and has reference value, so this data set is used as a verification set for the model to manually adjust hyperparameters to achieve model re-optimization.

The test set, as a data set for testing the prediction accuracy of the model, is collected in real time by the vehicle detector, processed and uploaded, and the deep residual network model can predict the next period of data.

2.2.2 Data conversion and enhancement

Image sample data input based on TensorFlow, feature set selection, different types of sample selection, sample vector map and feature map conversion, TFRecords sample dataset generation and dataset input.

Read the training samples in TFRecords format, encode the data according to the sample label with one-hot (One-Hot) method, and perform data enhancement such as image saturation and contrast conversion on the original image data.

2.3 Deep Residual Network Hierarchy Configuration

Based on the training task and complexity of the present invention, the depth of 18-layer residual network model is set to improve the prediction accuracy while not occupying a large amount of training resources. The following table shows each level and its characteristics.

Table 3 Deep residual network level configuration

The model adopts the 3x3 small convolution kernel mode, and replaces one large convolution kernel with multiple small convolution kernels, which reduces the model parameters and increases the number of nonlinear activation functions. For a convolutional layer with the same input and output feature map size, the number of filters remains the same. When the size of the feature map is halved, the number of filters is doubled, and the feature map pooling step is 2 to maintain the relationship between layers. time complexity.

The deep residual network consists of a set of residual blocks, each residual block contains several stacked convolutional layers, and the rectified linear unit (Relu) and batch normalization layer (BN) are attached as convolutional layers, avoiding the gradient disappears or an overflow condition occurs.

The deep residual network model rewrites the optimal mapping as H(x)=F(x)+x, and approximating the residual function F(x) is also equivalent to approximating the optimal mapping H(x). The rewritten residual map is easier to optimize than the original optimal solution map. The network residual is realized by adding a "Shortcut Connections" in the feedforward network. The shortcut skips one or more layers and merges with the main path with different step sizes, and the structure output can be expressed as

m _l+1 ＝Re lu(m _l +F(m _l ,w _l ))

In the formula, m _l and m _l+1 are the input and output of the first residual block respectively, Re lu() is the modified linear unit function, F represents the residual mapping function, and w _l is the parameter of the residual learning unit.

If the input and output dimensions are different, you need to add a linear projection To match the dimension size, after adding the linear projection, the formula is further transformed into

2.4 Forward Propagation

Forward propagation can extract the high-level features of the input image to obtain more abstract semantic features. Considering the differences between the training set and actual traffic flow data and the expressive ability of deep features, the convolutional layer features in the present invention are extracted . In the forward propagation training process, the desired learning objective function should be set first, and the function setting is:

where x is the input feature data, is the probability of the predicted result of the model, and w and b are the parameters obtained from the model training.

Through the feature sparse extraction of multiple convolutional layers, the extracted sparse convolutional features are calculated using the mean pooling operation. Each batch of input sample images is converted into sparse features and entered into the fully connected layer. After logits calculation, we get For each type of [batch training size × classification number] classification probability matrix of this batch of sample data, and then undergo a softmax operation to ensure that all outputs are positive values, and stretch all row values of the matrix to the interval [0,1] , and the sum of any row probabilities is equal to 1. The matrix stretched by the softmax operation, the maximum value of each row is the value with the maximum output probability, which is the prediction result of this training.

2.5 Back propagation and parameter tuning

During the training process of the deep residual model, the convolutional layer extracts each batch of sample data and calculates the sparse features layer by layer and records the corresponding parameter values. The sparse features extracted from the bottom layer are input to the logits layer to calculate the sample classification value.

The loss function of each training is calculated as the cross entropy between the true type of the sample and the model prediction result, and the training loss function of each batch of samples is calculated as follows:

In the formula, t _ki is the probability that sample k belongs to category i, and y _ki is the model prediction probability that sample k belongs to category i.

By comparing the real and model prediction and recognition classification results, the model loss function is calculated, the model fitting error is backpropagated, and each parameter is continuously adjusted during the iterative process of the deep residual model, which effectively increases the robustness of the model and reduces the Probability of overfitting occurring.

Input the training set data, tune the parameters, and select the corresponding optimizer. The commonly used mini-batch SGD training algorithm is easy to fall into local optimum and is greatly affected by the learning rate. Therefore, the Momentum optimizer based on the gradient-based moving exponential weighted average is selected to smooth the network parameters, which can solve the problem of excessive swings in the update range of the mini-batch SGD optimization algorithm, and at the same time accelerate the convergence speed of the network.

Assuming that the current iteration step is t, the calculation formula based on the Momentum optimization algorithm is as follows:

v _dw =βv _dw +(1-β)dW

v _db =βv _db +(1-β)db

W=W-αv _dw

b=b-αv _db

In the above formula, v _dw and v _db are the gradient momentum accumulated by the loss function in the previous t-1 round of iterations. β is an exponential value of gradient accumulation, which is set to 0.9; The obtained gradient, W and b are the update formulas of the network weight vector and bias vector, and α is the learning rate of the network;

After the parameter tuning is completed, proceed to the last step of model training, input the verification set data, test the performance of the model, and manually fine-tune the learning rate and other hyperparameter values set in Table 2.

3. Prediction of traffic characteristic value in the next period

After the tuning of each parameter is completed, the trained deep residual network model is used to predict the change trend of road traffic flow characteristic data in a short period of time. The input data time point is from 1s to 351s, and the output data time point is from 51s to 401s.

The predicted two-dimensional trend image is transformed into text data again through digit-map conversion. The traffic speed and traffic density prediction data are used for three-phase traffic flow division, and the vehicle arrival rate and vehicle occupancy data are used for on-ramp control. The forecast data is shown in the table below:

Table 4 Predicted traffic flow characteristic data table

According to the theory of three-phase traffic flow, combined with the actual road conditions of the road section, the threshold speed is set as the judgment condition for the phase transition. Identify and calibrate the corresponding traffic flow phase interval according to the set speed threshold, and the division of the phase interval also provides a judgment condition for the on-ramp control in step 4. The specific threshold speed and interval division are given in Table 5:

Table 5 Threshold speed interval division table

That is, when the speed of the traffic flow is less than or equal to 50km/h, the free flow turns into a synchronous flow, and when the speed is less than 22km/h, the synchronous flow turns into a wide motion jam; when the traffic speed is greater than 25km/h, the wide motion jam turns into a synchronous flow, and the speed is greater than 70km /h, the synchronous flow is converted to a free flow.

4 ramp control and information release

4.1 Ramp linkage control method

Take the speed threshold set in the previous step as one of the judgment conditions for judging ramp opening, ramp adjustment and ramp closing. The traffic flow in the free flow, synchronous flow and wide motion congestion controls the opening and closing of the ramp. When the road is in the state of synchronous flow and wide motion congestion, the corresponding on-ramp control method is adopted.

The on-ramp control status is set using three indicators: the upstream section vehicle speed, the downstream lane occupancy rate, and the downstream flow of the main line. We take the speed threshold for dividing traffic flow phases as one of the indicators for on-ramp control, and combine the two theories of three-phase traffic flow theory and on-ramp control method to realize road linkage control. The specific on-ramp control status is shown in the table below.

Table 6 Entry ramp control condition table

When carrying out ramp regulation, we use ALINEA algorithm, this algorithm has been improved in the present invention, utilizes the next control cycle vehicle occupancy rate of prediction to calculate ramp regulation rate and makes ALINEA algorithm possess feedforward mechanism and prediction mechanism. We set the adjustment cycle and the short-term forecast window to the same value, namely 50s. This assignment method can improve the fit among the deep residual model, three-phase traffic flow and ALINEA algorithm, making it possible for the three algorithms to carry out linkage control of on-ramps.

Assuming that the current cycle is k-1 cycle, O _out (k-1) and other data are collected by the vehicle detector in real time, O _out (k) is the predicted value of vehicle occupancy in the next cycle k, and r(k) is determined by k-1 cycle The internal O _out (k-1) data is calculated, then the predicted value of the ramp regulation rate for the k+1 period can be obtained with the help of r(k) and O _out (k). The green light duration of ALINEA algorithm is fixed, and the flow of vehicles entering the main line is controlled by adjusting the time interval between adjacent green lights every minute.

The ALINEA algorithm is as follows:

Before calculating the ramp regulation rate, it is necessary to calibrate the three parameters of downstream expected occupancy rate, regulation cycle and _KR . According to relevant research experience, it is recommended to set the expected occupancy rate to 0.3 and _KR to 70 in order to obtain the best control effect. Vehicle occupancy rate O _out (k-1)=0.37, O _out (k)=0.44.

Calculated from the measured data of the k-1 period vehicle collector, the periodical ramp adjustment rate r(k) is 20s, which is brought into the improved ALINEA algorithm to calculate r(k+1)=20-70×0.14≈10s. That is, the total duration of the green light in the adjustment cycle is 10s, and the signal light controls the vehicle through the switching of the traffic light. The green light is fixed at 2s each time, and the remaining 40s in the adjustment cycle are allocated to the red light. The red light flashes 5 times in total, and each time lasts The time is 8s.

4.2 Information release

There are three main ways to release information. The first is to control the on-ramp through the signal lights at the ramp entrance and use the calculation results of the algorithm, which has the best control effect; the other two methods provide reference for drivers to make travel decisions through the release of congestion information.

Ramp lights. Use the ramp signal lights to control the vehicle to enter the ramp. During the green light period, vehicles are allowed to enter the main road from the ramp; during the red light period, vehicles must stop and wait on the ramp, and are not allowed to enter the main road.

The upstream LED display board of the road section. The networked LED display board installed upstream can inform drivers of road congestion ahead, and guide drivers to choose other road sections or change modes of transportation to avoid congestion.

Mobile devices. With the help of mobile devices such as broadcasting and navigation software, road conditions are released, and drivers are reminded of road congestion, providing reference for drivers to make travel decisions.

4.3 Error Analysis

The use of RMSE, MAE and MAPE can reflect the degree of dispersion of the forecast data and the size of the actual error, and can measure the forecasting effect of the forecasting model.

The formulas of the three evaluation indicators are as follows:

In the formula, x _i represents the actual traffic flow data at the i-th moment, x' _i represents the forecast data of the traffic flow output by the model at the i-th moment, and N represents the length of the traffic flow time series to be evaluated.

In this embodiment, the calculated sizes of the three indicators are shown in the following table:

Table 7 Forecast error analysis table

Through the calculation of each index, it is not difficult to see that the degree of dispersion of the predicted data in this embodiment is small, and the accuracy of data prediction is as high as 97.06%. The change trend and prediction error analysis of the traffic flow characteristic value are shown in Figure 3, which shows that the data prediction using the deep residual network model has low complexity and high accuracy, and the prediction value change is relatively stable.

4.4 Simulation and effect evaluation

Because the procedure of embedding the vehicle detector is relatively complicated, and the damage to the road surface is relatively large, before the present invention regulates and controls the road, its performance should be evaluated by simulation software. Six parameters are used to evaluate the road control effect, including the maximum and minimum values of the downstream mainline speed, the maximum and minimum flow rates, the lateral and longitudinal fluctuations of the downstream mainline speed, and the lateral and longitudinal fluctuations of the downstream mainline flow. The evaluation index uses the absolute mean of the sequence difference and the standard deviation of the sequence to represent the magnitude of the data sequence change and the speed (or frequency) of the data sequence change, that is, the horizontal volatility and the vertical volatility. The calculation formula of the evaluation index is as follows:

In the formula, x _i is the i- _{th value in the data, Δxi = x i -xi -1} ; is the data mean. The evaluation of the simulation results is shown in Table 8 and Table 9.

Table 8 Simulation evaluation results of downstream main line speed

Table 9 Simulation evaluation results of downstream mainline flow

It can be seen from Table 8 that under the ramp control state, the downstream speed of the main line has been significantly improved. In addition, compared with the general ALINEA algorithm, after using the improved ramp linkage control method of the present invention, the minimum speed of the downstream main line has been significantly improved.

After the ramp linkage control is adopted, the longitudinal fluctuations of speed and flow in the downstream of the main line are greatly improved compared with the longitudinal fluctuations of speed and flow in the uncontrolled and traditional control, which shows that the traffic conditions downstream of the main line can regulate the ramp control strategy quite sensitive. Through the regulation, the downstream of the main line has transitioned from the blocking state to the synchronous flow state; after the ramp linkage control is adopted, the lateral fluctuation of the flow downstream of the main line does not change much, which shows that the control strategy is more effective when it does not affect the normal travel of vehicles The traffic congestion has been greatly alleviated and the traffic conditions downstream of the main line have been improved.

Claims (6)

1. a kind of Entrance ramp inter-linked controlling method based on depth residual error network model, which comprises the steps of:

Step 1: collecting traffic flow character historical data, the laggard line number figure conversion of data prediction was converted data to the time For the two dimensional image of sequence；

Step 2: the prediction model based on depth residual error network traffic flow characteristic value, setting are established and trained to input image data Hyper parameter and network level carry out arameter optimization by propagated forward and backpropagation, complete model training, and deep learning is gone through History traffic flow character value variation tendency；

Step 3: collecting arithmetic for real-time traffic flow characteristic, number figure conversion will be carried out by pretreated data, is converted data to Using the time as the two dimensional image of sequence；After conversion by image data input trained depth residual error Network Prediction Model carry out it is short When predict, output trend figure in short-term, and converted using figure number and anticipation trend figure is switched into text data；

Step 4: the text data after conversion, using trained depth residual error Network Prediction Model to road traffic features value into The wagon flow state of prediction data and identification is input in On-ramp Control ALINEA algorithm by row short-term prediction, is calculated next Period vehicle occupancy rate and maximum queuing vehicle number are controlled, linkage control is carried out in advance to the vehicle flowrate for importing main line；

Step 5: carrying out simulation evaluation and the analysis of ALINEA algorithm using VB+VISSIM program, and issue traffic information.

2. the Entrance ramp inter-linked controlling method according to claim 1 based on depth residual error network model, feature exist In step 1 specifically comprises the following steps:

1.1 history data collection

Video is shot by unmanned plane and two methods of wagon detector collect history vehicle flowrate data, and unmanned plane shoots video can To collect section vehicle flowrate data and section wagon flow state change situation；Wagon detector collects section by detection and processing The traffic flow character value in each lane；

Definition traffic flow character value is vehicle flowrate Q, averagely flow speeds v, vehicle density k_mWith vehicle occupancy rate R_t, traffic flow spy Value indicative can be by directly collecting or being calculated indirectly, and vehicle occupancy rate uses time occupancy for index, and calculation formula is as follows:

T is the time span in each control period；t_iIt is i-th vehicle by the time shared by section, unit is the second；N is minute The interior vehicle number by section；

1.2 data prediction

Collected data must be pre-processed before number figure conversion, including disorder data recognition, abnormal data reparation and data Denoising and normalization；

1.2.1 disorder data recognition

The traffic flow abnormal data of acquisition mainly includes three kinds of situations:

(1) magnitude of traffic flow, speed and occupation rate data have exceeded reasonable threshold range；

(2) relationship between the magnitude of traffic flow, speed and occupation rate data does not meet traffic flow theory；

(3) there is missing in the magnitude of traffic flow, speed and occupation rate data；

For above type of traffic flow data, data reparation can be directly deleted or carried out；

1.2.2 abnormal data reparation

Traffic flow off-note data mainly include two kinds of situations of error in data and shortage of data, for the data exception identified Value, is directly deleted；For missing data, Data-parallel language is carried out using K minimum distance neighbour's method, first according to Euclidean distance or phase Analysis is closed to determine the distance K sample nearest with missing data sample, then this K value is weighted and averaged to estimate the sample Missing data；

1.2.3 data de-noising

Denoising uses single exponential smoothing algorithm, and calculation formula is as follows:

In formula,It is respectively the smoothed data and real data at m moment with X (m), k is Smoothness Index, takes 0.1；

1.2.4 data normalization

Using Logistic/Softmax transform method, all pending datas are transformed into [0,1] section；

1.3 number figure conversions

Two dimensional image conversion will be carried out by sequence of the time by pretreated traffic flow data, and obtain different road condition conditions Lower real-time time-traffic flow character data variation tendency chart, the data being disposed are saved with graphic form through subsequent place Manage the training and test as depth residual error network model.

3. the Entrance ramp inter-linked controlling method according to claim 1 based on depth residual error network model, feature exist In step 2 specifically comprises the following steps:

The input of 2.1 data and initialization

Image sample data input is carried out based on TensorFlow, carries out feature set selection, different types of sample is chosen, sample Polar plot and characteristic pattern conversion, TFRecords sample data set generate, data set inputs；

The training sample for reading TFRecords format uses an efficient coding One-Hot mode by data according to sample label It is encoded；

The setting of 2.2 hyper parameters

It needs to carry out hyper parameter setting before the training of depth residual error network model, the main parameter that is arranged includes batch training size, study Rate, weight attenuation rate, optimizer selection；

Learning rate setting is conducive to model gradient in suitable range and drops to optimal value, and biggish initial is arranged first The setting of habit rate, with the increase of model the number of iterations, gradually adjusts to minimum learning rate, to obtain faster training speed and mould Type precision；L is arranged using weight attenuation rate₂Regularization term parameter adjusts influence of the model complexity to loss function, prevents Model over-fitting；Momentum optimizer is selected, the movement index weighted average based on gradient improves loss function convergence speed Degree；

The setting of 2.3 network model levels

Network model first layer is convolutional layer, is responsible for extracting low-level feature；

The second layer is maximum pond layer, reduces estimation mean shift, retains picture texture information；

3-6 layers are convolutional layer, are responsible for extracting high-level feature；

Layer 7 is average pond layer, calculates the feature that convolutional layer extracts and is input to full articulamentum；

Full articulamentum is responsible for the output of class probability；

2.4 propagated forward

In propagated forward training process, desired learning objective function, function setup should be first set are as follows:

Wherein x is input feature vector value,For model prediction probability of outcome, w and b are the parameter that model training obtains；Pass through multiple volumes The sparse extraction of the feature of lamination calculates the sparse convolution feature of extraction using the operation of mean value pondization, the every batch of of input Sample image is converted into sparse features, into full articulamentum, calculates by logits, obtains the batch sample data for every [batch training size × classification number] class probability matrix of seed type；

Softmax operation ensure that all outputs are positive value, all line number values of matrix is stretched to [0,1] section, and any Row probability, which is added, is equal to the matrix that 1, softmax operation stretched, and the maximum value of every row is the maximum value of output probability, as originally The prediction result of secondary training；

2.5 backpropagations and arameter optimization

In depth Remanent Model training process, convolutional layer, which extracts each batch sample data, successively to be calculated sparse features and records Corresponding parametric values, the sparse features extracted from the bottom, are input to logits layers, calculate sample classification value；

Trained loss function is calculated as the cross entropy of sample actual types Yu model prediction result, the instruction of every batch of sample every time Practice loss function and calculates such as following formula

In formula, t_kiBelong to the probability of classification i, y for sample k_kiBelong to the model prediction probability of classification i for sample k；It is true by comparing Real and model prediction and identification classification results, are calculated model loss function, models fitting error back propagation；

Training set data is inputted, arameter optimization is carried out, selects the average weighted Momentum optimization of the movement index based on gradient Device is smoothed network parameter；

If current iterative step is t, as follows based on Momentum optimization algorithm calculation formula:

v_dw=β v_dw+(1-β)dW

v_db=β v_db+(1-β)db

W=W- α v_dw

B=b- α v_db

In above formula, v_dwAnd v_dbThe gradient momentum β that be loss function respectively accumulate in preceding t-1 wheel iterative process is that gradient is tired A long-pending index value, is set as 0.9；DW and db obtained gradient when being loss function backpropagation respectively, W, b are networks The more new formula of weight vectors and bias vector, α are the learning rates of network；After arameter optimization, model training is carried out most Latter step, input verifying collection data, test model performance, manual fine-tuning hyper parameter numerical value.

4. the Entrance ramp inter-linked controlling method according to claim 1 based on depth residual error network model, feature exist In step 3 specifically comprises the following steps:

Video is shot by unmanned plane and two methods of wagon detector collect arithmetic for real-time traffic flow characteristic, and in traffic flow Phase transition decision condition is arranged in mobile track of vehicle: being more than in given threshold value time interval, along the vehicle of rail running Speed get lower than or higher than the phase transitions point given threshold value speed；

Set data entry time window, the real-time traffic characteristic value that input wagon detector is collected, depth residual error net after input Network model is by data prediction and is converted to time series variation figure, and exports the variation of subsequent period traffic characteristic prediction data Tendency chart predicts that the two dimensional image of output is converted to text data by figure number；

Traffic flow is divided into free flow referring to the three-phase traffic flow theory that Kona proposes in conjunction with road traffic flow space-time characteristic F, synchronous stream S and Wide moving jam J three-phase；

Traffic flow phase change is considered as one and freely flows to the phase transition process step by step that synchronous stream arrives wide movement obstruction again: F → S → J； In conjunction with China's road status, each phase transitions point is set by threshold value of speed.

5. the Entrance ramp inter-linked controlling method according to claim 1 based on depth residual error network model, feature exist In: short-term prediction is carried out to road traffic features value described in step 4, short-term prediction includes flow speeds, vehicle density, vehicle The prediction data of occupation rate；

The Entrance ramp inter-linked controlling method, specifically comprises the following steps:

Before ALINEA algorithm, main line downstream vehicle occupation rate O in the k-1 control period is defined_out(k-1) by wagon detector Acquisition, k control downstream vehicle occupation rate O in the period_out(k) and Entrance ramp vehicle arriving rate d (k) is by depth residual error network mould Type prediction；

R (k) is by O in the k-1 period_out(k-1) data are calculated, then by r (k) and O_out(k) ring road in k+1 period can be obtained Regulation rate predicted value；

ALINEA algorithm long green light time is fixed, by adjusting the time interval of adjacent green light starting in per minute to remittance main line Vehicle carries out flow control；

Within a control period, regulation rate calculation formula is

In formula, r (k+1) is the metering rate that kth+1 controls period calculating；R (k) is the ramp metering in the kth control period Rate r (k) is calculated by vehicle occupancy rate measured data in the k-1 control period and is obtained, when regulation rate is green light in a control period It is long, unit s；K_RParameter adjusts external disturbance fixed in feedback control；It is the expectation occupation rate in main line downstream；O_out(k) It is the vehicle occupancy rate predicted value in main line downstream in the kth control period.

6. the Entrance ramp inter-linked controlling method according to claim 1 based on depth residual error network model, feature exist In, simulation evaluation described in step 5 and analysis, specifically

Program is developed using 4.3 COM of VB+VISSIM, ramp metering rate is realized based on ALINEA algorithm；Road network is loaded in a program Model is emulated, and the simulation run time takes 3600s, and corresponding practical peak one hour, in simulation process, each control period will Simulation status data, including green light end time, split and occupation rate can be returned；

If letter controls the period as t, unit s, program obtains the occupation rate of data detector return at the t-1 moment of each cycle, Detector returns to an occupation rate at interval of t-2, and the green letter of next cycle is calculated according to adjustment factor and best occupation rate Than；

When split is more than or equal to 0.8, ring road letter, which is controlled, continues green light；Less than 0.2, ring road letter, which is controlled, continues red light；Between 0.8 with Split between 0.2 carries out fixed cycle control, and the signal lamp duration of ramp metering rate is calculated according to the split after optimization.