CN110503833B - Entrance ramp linkage control method based on depth residual error network model - Google Patents
Entrance ramp linkage control method based on depth residual error network model Download PDFInfo
- Publication number
- CN110503833B CN110503833B CN201910809931.9A CN201910809931A CN110503833B CN 110503833 B CN110503833 B CN 110503833B CN 201910809931 A CN201910809931 A CN 201910809931A CN 110503833 B CN110503833 B CN 110503833B
- Authority
- CN
- China
- Prior art keywords
- data
- traffic flow
- model
- time
- prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0137—Measuring and analyzing of parameters relative to traffic conditions for specific applications
- G08G1/0145—Measuring and analyzing of parameters relative to traffic conditions for specific applications for active traffic flow control
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/07—Controlling traffic signals
- G08G1/075—Ramp control
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Traffic Control Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an entrance ramp linkage control method based on a deep residual error network model, which comprises the steps of firstly, collecting historical data of traffic flow characteristics, preprocessing and then carrying out digital-to-image conversion; secondly, inputting image data, and establishing and training a prediction model of a traffic flow characteristic value; thirdly, collecting real-time traffic flow characteristic data, converting the preprocessed data graph into an input model, outputting a short-time change trend prediction graph, and converting the prediction trend graph into text data by utilizing graph-number conversion; fourthly, the converted text data carries out short-time prediction on the road traffic characteristic value by utilizing a trained prediction model, and carries out linkage control on the traffic flow which is converged into the main line in advance; and finally, performing simulation evaluation and analysis on the ALINEA algorithm by using a VB + VISSIM program, and issuing road condition information. According to the control method, data are preprocessed and then are respectively subjected to digital-to-image conversion, more detailed characteristics of the two-dimensional image can be extracted through the digital-to-image conversion, model training and prediction time is reduced, and prediction precision and real-time information processing speed are improved.
Description
Technical Field
The invention belongs to the field of traffic data prediction, and relates to a short-time traffic flow characteristic data prediction and traffic flow state division method.
Background
The road traffic flow characteristic data mainly comprises traffic flow, traffic speed, traffic density and vehicle occupancy. The road traffic flow characteristic data prediction can predict the next period data and divide traffic flow states, and is a necessary premise for traffic management and control. Traffic management and control based on traffic flow prediction not only facilitates travelers to make better travel plans, but also facilitates traffic management departments to make better management decisions.
In the existing road traffic flow characteristic data prediction method, shallow layer models and time sequence models are most widely used at present. The shallow model cannot well mine information in traffic flow data, and the time sequence model only considers the characteristics of the traffic flow in time and ignores the influence in space. The deep residual error network can extract the characteristics in time and space based on the data change trend image, so the invention provides a traffic flow characteristic data prediction method based on the deep residual error network, which extracts the space-time characteristics in the traffic flow characteristic data change trend image through convolution and carries out nonlinear regression to finally realize the prediction of road traffic flow characteristic data and carry out traffic management and control based on the prediction data.
With the rapid development of deep learning and artificial intelligence technology, the accuracy of road traffic flow prediction is increasingly improved. The road traffic flow prediction can assist a traffic management department to make a more reasonable traffic control strategy, provide reference for a vehicle owner to respond to the traffic control strategy, relieve traffic jam and reduce waste of traffic resources. The road traffic flow prediction based on the depth residual error network model provides basic data for an intelligent traffic system, and promotes the development and application of the intelligent traffic system.
Object of the Invention
The invention provides a short-time traffic flow characteristic data prediction and entrance ramp linkage control method based on a deep residual error network model, aiming at improving the precision deficiency of the existing short-time traffic flow characteristic data prediction and traffic flow state division method and solving the problem that the traditional entrance ramp control method does not have a short plate with a feedforward mechanism and a prediction mechanism.
The method for realizing the entrance ramp linkage control based on the depth residual error network model mainly comprises the following steps:
step 1: collecting traffic flow characteristic historical data, carrying out data-map conversion after data preprocessing, and converting the data into two-dimensional images taking time as a sequence;
step 2: inputting image data, establishing and training a prediction model based on a deep residual error network traffic flow characteristic value, setting hyper-parameters and network levels, performing parameter tuning and optimization through forward propagation and backward propagation, completing model training, and deeply learning a history traffic flow characteristic value change trend;
and step 3: collecting real-time traffic flow characteristic data, performing figure-to-figure conversion on the preprocessed data, and converting the data into two-dimensional images taking time as a sequence; inputting the image data into a trained depth residual error network prediction model for short-time prediction after conversion, outputting a short-time change trend prediction graph, and converting the prediction trend graph into text data by utilizing graph number conversion;
and 4, step 4: the converted text data carries out short-time prediction on the road traffic characteristic value by utilizing a trained deep residual error network prediction model, the prediction data and the recognized traffic flow state are input into an entrance ramp control ALINEA algorithm, the vehicle occupancy and the maximum queuing vehicle number in the next control period are calculated, and the traffic flow which is imported into a main line is subjected to linkage control in advance;
and 5: and (4) carrying out simulation evaluation and analysis on the ALINEA algorithm by using a VB + VISSIM program, and issuing road condition information.
The step 1 specifically comprises the following steps:
1.1 historical data Collection
Historical traffic flow data are collected through two methods of unmanned aerial vehicle video shooting and vehicle detector detection, and road section traffic flow data and road section traffic flow state change conditions can be collected through the unmanned aerial vehicle video shooting; the vehicle detector can collect traffic flow characteristic values of all lanes of the road section by detecting and processing;
defining the characteristic values of traffic flow as traffic flow Q, average traffic flow speed v and traffic flow density kmAnd vehicle occupancy rate RtThe traffic flow characteristic value can be obtained by direct collection or indirect calculation, and the vehicleThe occupancy adopts the time occupancy as an index, and the calculation formula is as follows:
t is the time length of each control period; t is tiThe time occupied by the ith vehicle passing through the cross section is second; n is the number of vehicles passing through the cross section in the measurement time.
1.2 data preprocessing
The traffic flow characteristic values directly acquired by the unmanned aerial vehicle and the vehicle detector have the problems of data loss, error, invalidity, time drift and the like all the time inevitably, and if the problem data are directly fed back to the application of the deep residual error network model, the model prediction error is caused. Therefore, the acquired data needs to be preprocessed before the digital image conversion, namely data cleaning (data clean), which comprises abnormal data identification, abnormal data restoration and data denoising and normalization;
1.2.1 abnormal data identification
The collected traffic flow abnormal data mainly comprises three conditions:
(1) data such as traffic flow, speed, occupancy rate and the like exceed a reasonable threshold range;
(2) the relationship among data such as traffic flow, speed, occupancy and the like does not accord with the traffic flow theory;
(3) data such as traffic flow, speed, occupancy rate and the like are lacked;
for the traffic flow data of the types, the data can be directly deleted or repaired.
1.2.2 abnormal data repair
The traffic flow abnormal characteristic data mainly comprises two conditions of data error and data loss, and the identified data abnormal value is directly deleted; and aiming at missing data, performing data completion by adopting a K-means clustering method, determining K samples closest to the missing data sample according to Euclidean distance or correlation analysis, and then performing weighted average on the K values to estimate the missing data of the sample.
1.2.3 data De-noising
Under the condition that the acquisition period of the traffic flow data is short, the traffic flow data often contains more Gaussian noises, and the analysis and modeling of the traffic data are influenced to a certain extent, so that simple denoising processing needs to be performed on the sampled data. (ii) a The data denoising adopts a one-time exponential smoothing algorithm, and the short-term change trend of traffic flow data is reserved to the greatest extent on the basis of not increasing the complexity of the algorithm as much as possible; the first exponential smoothing method has the following calculation formula:
in the formula (I), the compound is shown in the specification,and x (m) are the smoothed data and the actual data at time m, respectively, and k is a smoothing index, which is taken to be 0.1.
1.2.4 data normalization
When a neural network model is constructed for traffic flow data, normalization processing is generally required to be carried out on the data so as to avoid the phenomenon of saturation of neurons, and a Logistic/Softmax conversion method is adopted to convert all data to be processed into a [0,1] interval.
1.3 digit map conversion
And carrying out two-dimensional image conversion on the preprocessed traffic flow data by taking time as a sequence to obtain a real-time-traffic flow characteristic data change trend graph under different road states, and storing the processed data in a picture form and using the processed data as a training and testing of a depth residual error network model after subsequent processing.
The step 2 specifically comprises the following steps:
2.1 data entry and initialization
Image sample data input is carried out based on TensorFlow, and 5 parts of characteristic set selection, different types of sample selection, sample vector diagram and characteristic diagram conversion, TFRecords sample data set generation and data set input are carried out; TFRecords is a binary file format, occupies small internal memory, is convenient to copy, move and store, and does not need a separate label file;
reading a training sample in a TFRecords format, and coding data by adopting a One-bit effective coding (One-Hot) mode according to a sample label;
in order to avoid overfitting and enhance the robustness of the model, original image data can be subjected to data enhancement methods such as image saturation and contrast conversion, characteristics are guaranteed to be unchanged, and meanwhile, information such as pixel positions is changed, so that a more sufficient sample data set is obtained.
2.2 Superparameter settings
The method comprises the following steps that hyper-parameter setting is required before training of a depth residual error network model, and main setting parameters comprise Batch training size (Batch-size), Learning rate (Learning rate), weight-decay rate (weight-decay-rate), optimizer selection (optimizer) and the like;
the batch training size value determines the descending direction and is inversely proportional to the size of the data set; when the data set is large enough, a suitable reduction may reduce the amount of computation; if the data volume is small and noise data exists, a large batch training size value is set to reduce the interference of the noise data;
the learning rate determines the updating amplitude of the weight value, and the learning rate is set in a proper range to be beneficial to the gradient of the model to be reduced to an optimal value; firstly, setting a larger initial learning rate, and gradually adjusting the initial learning rate to the minimum learning rate along with the increase of the iteration times of the model so as to obtain a faster training speed and a faster model precision;
overfitting can occur in the training process of the depth residual error network model, the larger the network weight is, the higher the corresponding overfitting degree is, and therefore L is set by adopting the weight attenuation rate2Regularization term parameters mainly play a role in adjusting the influence of model complexity on a loss function and preventing overfitting of the model;
and a Momentum optimizer is selected, wherein the optimizer is mainly based on gradient moving index weighted average, and the loss function has higher convergence speed and smaller swing amplitude during network optimization.
2.3 network model hierarchy setting
The depth residual error network model (ResNet) uses a 3x3 small convolution kernel mode, a plurality of small convolution kernels replace one large convolution kernel, model parameters are reduced, the number of nonlinear activation functions is increased, the model calculation amount is smaller, and the identification error is lower. For convolutional layers with the same size of the input and output feature maps, the number of filters is unchanged, when the size of the feature map is halved, the number of filters is doubled, and the pooling step length of the feature map is 2 so as to keep the time complexity between layers.
In the invention, only the image change trend needs to be identified for prediction, and excessive network layer number does not need to be set, so that unnecessary resource waste is avoided. The first layer of the network is a convolutional layer and is responsible for extracting low-level features, the second layer is a maximum pooling layer, the estimated mean shift is reduced, and the image texture information is reserved; 3-6 layers are convolution layers and are responsible for extracting high-level features; the seventh layer is an average pooling layer, inhibits the variance increase of the estimation value, calculates the characteristics extracted by the convolutional layer and inputs the characteristics into the full-connected layer; and the full connection layer is responsible for outputting the classification probability.
The depth residual error network is composed of a group of residual error blocks, each residual error block comprises a plurality of stacked convolutional layers, and a modified linear unit (Relu) and a batch normalization layer (BN) are used as the attachment of the convolutional layers, so that the situation of gradient disappearance or overflow is avoided.
The Relu activation function introduces non-linear characteristics into the model network, converts the input features of the model nodes into output features, and passes on to the next operation unit. By using the Relu function, part of output data is zeroed, the model network can introduce sparsity by itself, and the problem of gradient disappearance is effectively overcome by the performance of the segment line.
The depth residual network model rewrites the optimal mapping to 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 demapping.
Network residual errors are realized by adding a 'short Connections' in a feed-forward network, low-layer errors can be transmitted to an upper layer through shortcuts in a training process, phenomena such as gradient disappearance caused by layer numbers are reduced, and model training precision is improved under the condition that the calculated amount is slightly increased.
The shortcuts merge with the main path by skipping one or more layers with different step sizes, and the structure output can be expressed as
ml+1=Re lu(ml+F(ml,wl))
In the formula, mlAnd ml+1Respectively input and output of the 1 st residual block, Re lu () is a modified linear unit function, F represents a residual mapping function, wlAre parameters of the residual learning unit.
If the input and output dimensions are different, linear projection needs to be addedTo match the dimension size, and after increasing the linear projection, the formula is further converted into
After batch normalization is used, input data can still be distributed in a standard interval along with the depth acceleration of the model or the parameter change in the training process, the gradient disappearance is avoided, the convergence of the model is accelerated, the dependence of the model on the initial network weight is reduced, and a batch normalization operation formula is expressed as
In the formula, xkIs the activation degree of the neuron, Exk]X representing a collection of training data set acquisitionskIs determined by the average value of (a) of (b),for each batch of training data xkStandard deviation of (2).
Meanwhile, two learnable parameters (gamma, beta) are introduced into the batch normalization operation and are used for reconstructing the transformed activation and restoring the feature distribution learned by the original network. The operation does not destroy the characteristics learned by the previous layer operation of the data, and the expressive ability of the network is not influenced.
2.4 Forward propagation
The forward propagation can extract high-level features of the input images to obtain more abstract semantic features, and the features of the convolutional layer are extracted by considering the difference between a training set and actual traffic flow data and the expression capacity of deep features. In the forward propagation training process, an expected learning objective function is set firstly, and the function is set as follows:
wherein x is an input characteristic value,predicting the result probability for the model, and w and b are parameters obtained by model training;
performing feature sparse extraction on a plurality of convolution layers, calculating the extracted sparse convolution features by using a mean pooling operation, converting each input batch of sample images into sparse features, entering a full connection layer, and performing logits calculation to obtain a classification probability matrix of each type of the batch of sample data [ batch training size multiplied by classification number ];
and after softmax operation, ensuring that all outputs are positive values, extending all row values of the matrix to a [0,1] interval, and adding the probability of any row to be equal to 1. And (4) operating the stretched matrix by softmax, wherein the maximum value of each row is the value with the maximum output probability, and the value is the prediction result of the training.
2.5 Back propagation and parameter tuning
In the process of training the depth residual error model, the convolutional layer extracts the sparse features layer by layer from each batch of sample data and records corresponding parameter values, the sparse features extracted at the bottommost layer are input to the logits layer, and the sample classification value is calculated;
the loss function of each training is calculated as the cross entropy of the real type of the sample and the prediction result of the model, and the training loss function of each batch of samples is calculated as the following formula
In the formula, tkiIs the probability that a sample k belongs to class i, ykiPredicting a probability for a model for which sample k belongs to class i;
by comparing the actual and model prediction and identifying classification results, a model loss function is calculated, model fitting errors are propagated reversely, and all parameters are continuously adjusted in the process of depth residual model iteration, so that the robustness of the model is effectively increased, and the occurrence probability of overfitting is reduced;
inputting training set data, carrying out parameter tuning, and selecting a corresponding optimizer. The common mini-batch SGD training algorithm is easy to fall into local optimization and is greatly influenced by the learning rate. Therefore, a gradient-based mobility index weighted average Momentum optimizer is selected to smooth the network parameters, so that the problem of overlarge fluctuation of the update amplitude of a mini-batch SGD optimization algorithm can be solved, and the convergence speed of the network is accelerated;
assuming that the current iteration step is t, the calculation formula based on the Momentum optimization algorithm is as follows:
vdw=βvdw+(1-β)dW
vdb=βvdb+(1-β)db
W=W-αvdw
b=b-αvdb
in the above formula, vdwAnd vdbRespectively, the gradient momentum beta accumulated by the loss function in the previous t-1 iteration process is an index value of the gradient accumulation and is set to be 0.9; dW and db are gradients obtained when the loss function propagates reversely, W, b is an update formula of a network weight vector and an offset vector, and alpha is a learning rate of the network;
and after parameter tuning is finished, performing the last step of model training, inputting verification set data, testing the performance of the model, manually fine-tuning the learning rate and other hyper-parameter values.
The step 3 specifically comprises the following steps: after the training of the depth residual error network prediction model is completed, short-time prediction can be carried out on the road traffic characteristic value based on the depth residual error network model; real-time traffic flow characteristic data are collected through two methods of video shooting by an unmanned aerial vehicle and a vehicle detector, and phase conversion judgment conditions are set for vehicle tracks moving in the traffic flow: in exceeding a given threshold time interval, the vehicle speed traveling along the track becomes lower or higher than a given threshold speed of the phase transition point;
setting a data input time window, inputting real-time traffic characteristic values collected by a vehicle detector, preprocessing the data and converting the data into a time sequence change diagram by using an input depth residual error network model, outputting a traffic characteristic prediction data change trend diagram of the next period of time, and converting a predicted and output two-dimensional image into text data through a number diagram;
dividing the traffic flow into three phases of free flow (F), synchronous flow (S) and wide movement blockage (J) by combining the space-time characteristics of the traffic flow of the road; traffic flow phase change can be viewed as a gradual phase change process from free flow to synchronous flow to wide motion blockage (F → S → J); according to a three-phase traffic flow theory and the current road situation of China, the speed is taken as a threshold value to set each phase transition point.
The designed model can automatically divide each phase interval according to the set speed threshold value besides carrying out short-time prediction on the traffic characteristic value. The division into the predicted data phase sections is advantageous for performing the road control in step 4.
And 4, performing short-term prediction on the road traffic characteristic value, wherein the short-term prediction comprises prediction data of traffic flow speed, traffic flow density and vehicle occupancy.
The entrance ramp control is a traffic control mode which is widely applied and effective and is used for relieving road congestion, and the entrance ramp control strategy based on ALINEA is simple, efficient and easy to implement. By using the ALENIA algorithm, the regulation rate of the ramp can be controlled by controlling the duration of the red light, namely the number of vehicles passing per minute is regulated, so that the purpose of controlling the flow of the entrance ramp is achieved.
When the traditional ALENIA algorithm is used for queuing control, the turn-in mediation rate of the current period is calculated according to the vehicle occupancy of the downstream road section of the main line and the turn-in demodulation rate of the previous period, and a feed-forward mechanism and a prediction mechanism are not provided. The invention substitutes traffic characteristic prediction data into the algorithm, takes the short-time prediction time window as the control period of the ALENIA algorithm, can realize more accurate road advanced control, and makes up the defects of the traditional algorithm. The method not only improves the adjusting efficiency, but also effectively reduces the probability of congestion, and simultaneously can feed back the change after road control to the model to realize re-optimization control, thereby improving the adjusting precision.
And setting an entrance ramp control state by adopting three indexes of the upstream section vehicle speed, the downstream lane occupancy and the main line downstream flow. The upstream section vehicle speed is consistent with the division of the upper section traffic flow, namely, the three-phase traffic flow theory is linked with the entrance ramp control method, and the division of the traffic flow phase is used as one of the indexes of the entrance ramp control.
The entrance ramp control method in step 4 specifically comprises the following steps:
before adopting ALINEA algorithm, defining downstream vehicle occupancy rate O in k-1 control periodout(k-1) collected by the vehicle detector, k controlling the downstream vehicle occupancy O in the cycleout(k) And an ingress ramp vehicle arrival rate d (k) is predicted by the depth residual error network model;
r (k) is formed from O in k-1 periodout(k-1) data is calculated by r (k) and Oout(k) The predicted value of the ramp regulation rate in the k +1 period can be obtained;
the ALINEA algorithm has fixed green light duration, and the flow of vehicles merging into the main line is controlled by adjusting the time interval of starting of adjacent green lights in each minute;
in a control period, the regulation rate is calculated according to the formula
Wherein r (k +1) is the ramp regulation rate calculated in the k +1 control period; r (k) is in the kth control periodThe ramp regulation rate r (k) is obtained by calculating actual measurement data of vehicle occupancy in a k-1 control period, and the regulation rate is the green light time length in one control period and has the unit of s; kRExternal disturbance fixed in parameter adjustment feedback control;is the expected occupancy downstream of the main line; o isout(k) Is a predicted value of the vehicle occupancy rate downstream of the main line in the k-th control period.
The simulation evaluation and analysis in the step 5 are realized by using a VB + VISSIM 4.3COM development program and based on an ALINEA algorithm; loading a road network model in a program for simulation, wherein the simulation running time is 3600s, corresponding to one hour of an actual peak, and in the simulation process, each control cycle returns simulation state data comprising green light ending time, green signal ratio and occupancy rate;
setting a signal control period as t and the unit as s, acquiring the occupancy returned by the data detector at the t-1 th moment of each period by the program, returning the occupancy once every t-2 intervals by the detector, and calculating the green signal ratio of the next period according to the adjustment coefficient and the optimal occupancy;
when the green signal ratio is greater than or equal to 0.8, controlling the continuous green light by the ramp signal; less than 0.2, continuously red light is controlled by the ramp signal; and carrying out fixed period control on the green signal ratio between 0.8 and 0.2, and calculating the green light time of ramp control according to the optimized green signal ratio.
The invention aims to predict traffic characteristic values, identify road traffic states and adopt an entrance ramp control method to relieve traffic congestion, so that the entrance ramp control effect needs to be evaluated. Since the process of embedding the vehicle detector is complex and the damage to the road surface is large, the effect of the vehicle detector is evaluated by simulation software before the road is adjusted and controlled by the vehicle detector.
The deep residual error network has good robustness and effectively relieves the problem of gradient disappearance. According to the entrance ramp linkage control method based on the depth residual error network model, the historical data and the real-time data of traffic flow characteristics are collected, the data are preprocessed and then subjected to digital-to-image conversion, more detailed characteristics of a two-dimensional image can be extracted through the digital-to-image conversion, the model training and prediction time is reduced, and the prediction precision and the real-time information processing speed are improved. And the time-space characteristics in the traffic flow characteristic data change trend graph are extracted through convolution and nonlinear regression is carried out, the prediction of road traffic flow characteristic data is finally realized, traffic management and control are carried out based on the prediction data, basic data are provided for an intelligent traffic system, and the development and application of the intelligent traffic system are promoted.
Drawings
FIG. 1 is a general flow chart of a control method of the present invention;
FIG. 2 is a flowchart of the deep residual error network model training of the present invention;
FIG. 3 is an analysis diagram of the variation trend and prediction error of the traffic flow characteristic value in the embodiment.
Detailed Description
The present invention will be further described with reference to the following examples and drawings, but the present invention is not limited thereto.
Referring to fig. 1 to 3, the entrance ramp linkage control method based on the depth residual error network model includes the following steps:
1. data collection and processing
1.1 data Collection and screening
The traffic flow data acquisition method mainly comprises unmanned aerial vehicle video shooting and vehicle detector collection, wherein traffic flow characteristic data acquired by the vehicle detector is set as a training set for model training and parameter optimization, and the unmanned aerial vehicle video shooting data is used as a verification set for model hyper-parameter manual optimization.
And selecting a jamming and high-speed road section of Nanjing, wherein the road section has a ramp and is a working day on the same day, and the weather state is good. According to the data collected by the vehicle detector in real time, data with the duration of 350s is intercepted, the data reflects the formation process of traffic flow congestion, and the data has obvious change characteristics.
1.2 data preprocessing
Preprocessing the intercepted data, deleting and repairing traffic flow characteristic data which exceeds a reasonable threshold range and does not accord with a traffic flow theory and is missing, reducing noise of the data by adopting a one-time exponential smoothing algorithm, keeping a short-term variation trend of the traffic flow characteristic data, and normalizing the data by adopting a Logistic/Softmax conversion method. The pretreated traffic flow characteristic data are shown in the following table:
TABLE 1 real-time traffic flow characteristic data sheet
1.3 digit map conversion
After data preprocessing is finished, data text information is converted into a time-vehicle number change trend image, a time-vehicle flow speed change trend image, a time-vehicle flow density change trend image and a time-vehicle occupancy change trend image which take time as a sequence, four groups of image change trend time windows are set to be 50s, namely, the change trend of traffic flow feature data in the next 50s is predicted according to the traffic flow feature data in the current 50s period.
And after the conversion of the number map is finished, training a deep residual error network traffic flow characteristic data prediction model.
2. Deep residual error network traffic flow characteristic data prediction model training
2.1 Superparameter settings
The super-parameter setting is needed before the training of the depth residual error network model, and the main setting parameters comprise Batch training size (Batch-size), Learning rate (Learning rate), weight-decay rate (weight-decay-rate), optimizer (optimizer) and the like. Specific hyper-parameter settings are shown in the following table.
TABLE 2 deep residual network hyper-parameter settings
During model training, the effect of reducing the convergence rate of the Batch is far lower than the performance reduction amplitude caused by introducing a large amount of noise, and the GPU can exert better performance on the Batch of the power of 2, so that the adopted Batch value is 256.
And setting a larger initial learning rate, and gradually adjusting the initial learning rate to the minimum learning rate along with the increase of the iteration times of the model so as to obtain higher training speed and model precision. By adopting a Momentum optimizer which is based on the gradient moving index weighted average, the loss function convergence speed is higher and the swing amplitude is smaller during network optimization. And setting a weight attenuation ratio of 0.0001 to adjust the influence of the complexity of the model on the loss function and prevent the model from being over-fitted.
2.2 data entry and initialization
2.2.1 training set and test set partitioning
The division of the training set, validation set, and test set is based on the means and type of data acquisition.
The data volume collected by the vehicle detector is large, the data volume can be used as a training set for adjusting and optimizing the parameters of the depth residual error model, the parameters can be updated by gradient descent, and the minimization of the target function is realized.
The data volume that unmanned aerial vehicle gathered is less, and has the referential, so this data set supplies the model to carry out super parameter manual adjustment as verifying set, realizes the model and optimizes again.
The test set is used as a data set of the prediction accuracy of the test model, the data set is collected in real time through a vehicle detector, processed and uploaded, and the next period of data can be predicted through the deep residual error network model.
2.2.2 data conversion and enhancement
Image sample data input, feature set selection, different types of sample selection, sample vector graph and feature graph conversion, TFRecords sample data set generation and data set input are carried out on the basis of TensorFlow.
Reading a training sample in a TFRecords format, coding data by adopting a One-bit effective coding (One-Hot) mode according to a sample label, and performing data enhancement such as image saturation and contrast conversion on original image data.
2.3 deep residual network level configuration
Based on the training task and complexity of the invention, a depth residual error network model with the depth of 18 layers is set, so that the prediction precision is improved while a large amount of training resources are not occupied, and the following table shows each layer and the characteristics thereof.
TABLE 3 deep residual network hierarchy configuration
The model adopts a 3x3 small convolution kernel mode, and a plurality of small convolution kernels replace one large convolution kernel, so that the model parameters are reduced, and the number of nonlinear activation functions is increased. For convolutional layers with the same size of input and output feature map dimensions, the number of filters is unchanged, when the feature map dimension is halved, the number of filters is doubled, and the step length of pooling the feature map is 2 to maintain the time complexity between layers.
The depth residual error network is composed of a group of residual error blocks, each residual error block comprises a plurality of stacked convolutional layers, and a modified linear unit (Relu) and a batch normalization layer (BN) are used as the attachment of the convolutional layers, so that the situation of gradient disappearance or overflow is avoided.
The depth residual network model rewrites the optimal mapping to 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 demapping. The network residuals are implemented by adding a "short Connections" to the feed forward network. The shortcuts merge with the main path by skipping one or more layers with different step sizes, and the structure output can be expressed as
ml+1=Re lu(ml+F(ml,wl))
In the formula, mlAnd ml+1Respectively input and output of the 1 st residual block, Re lu () is a modified linear unit function, F represents a residual mapping function, wlAre parameters of the residual learning unit.
If the input and output dimensions are different, linear projection needs to be addedTo match the dimension size, and after increasing the linear projection, the formula is further converted into
2.4 Forward propagation
The forward propagation can extract high-level features of the input images to obtain more abstract semantic features, and the features of the convolutional layer are extracted by considering the difference between a training set and actual traffic flow data and the expression capacity of deep features. In the forward propagation training process, an expected learning objective function is set firstly, and the function is set as follows:
wherein x is the input characteristic data, and x is the input characteristic data,and w and b are parameters obtained by model training.
The method comprises the steps of extracting the characteristic sparsity of a plurality of convolution layers, calculating the extracted sparse convolution characteristic by using a mean pooling operation, converting each batch of input sample images into sparse characteristics, entering a full connection layer, calculating by logits to obtain a [ batch training size multiplied by classification number ] classification probability matrix of each type of the batch of sample data, ensuring all outputs to be positive values by softmax operation, extending all line values of the matrix to a [0,1] interval, and adding any line probability to be equal to 1. And operating the stretched matrix by softmax, wherein the maximum value of each row is the value with the maximum output probability, and the maximum value is the prediction result of the training.
2.5 Back propagation and parameter tuning
In the process of training the depth residual error model, the convolutional layer extracts the layer-by-layer calculation sparse features from each batch of sample data and records corresponding parameter values, the sparse features extracted from the bottommost layer are input to the logits layer, and the sample classification value is calculated.
The loss function of each training is calculated as the cross entropy of the real type of the sample and the prediction result of the model, and the training loss function of each batch of samples is calculated as follows:
in the formula, tkiIs the probability that a sample k belongs to class i, ykiThe probability is predicted for the model for which sample k belongs to class i.
By comparing the actual and model prediction and identifying classification results, a model loss function is calculated, model fitting errors are propagated reversely, and all parameters are continuously adjusted in the process of depth residual model iteration, so that the robustness of the model is effectively increased, and the occurrence probability of overfitting is reduced.
Inputting training set data, carrying out parameter tuning, and selecting a corresponding optimizer. The common mini-batch SGD training algorithm is easy to fall into local optimization and is greatly influenced by the learning rate. Therefore, a gradient-based Momentum optimizer of moving exponential weighted average is selected to smooth the network parameters, so that the problem of overlarge fluctuation of the update amplitude of the mini-batch SGD optimization algorithm can be solved, and the convergence speed of the network is accelerated.
Assuming that the current iteration step is t, the calculation formula based on the Momentum optimization algorithm is as follows:
vdw=βvdw+(1-β)dW
vdb=βvdb+(1-β)db
W=W-αvdw
b=b-αvdb
in the above formula, vdwAnd vdbRespectively, the gradient momentum beta accumulated by the loss function in the previous t-1 iteration process is an index value of the gradient accumulation and is set to be 0.9; dW and db are gradients obtained when the loss function propagates reversely, W, b is an update formula of a network weight vector and an offset vector, and alpha is a learning rate of the network;
after the parameters are adjusted and optimized, the last step of model training is carried out, verification set data is input, the performance of the model is tested, and the learning rate and other super-parameter values set in the table 2 are manually adjusted in a fine mode.
3. Traffic characteristic value prediction in next period
And after the parameters are adjusted and optimized, performing short-time prediction on the change trend of the road traffic flow characteristic data by using the trained deep residual error network model. The input data time points are from 1s to 351s, and the output data time points are from 51s to 401 s.
And converting the predicted two-dimensional change trend image into text data again through figure conversion. And the traffic flow speed and traffic flow density prediction data are used for three-phase traffic flow phase division, and the vehicle arrival rate and vehicle occupancy data are used for entrance ramp control. The prediction data are shown in the following table:
TABLE 4 predicted traffic flow characteristic data sheet
According to the three-phase traffic flow theory, the actual road conditions of the road sections are combined, and the judgment conditions, namely the threshold speed, are set for phase conversion. And identifying and calibrating corresponding traffic flow phase intervals according to the set speed threshold, wherein the division of the phase intervals also provides judgment conditions for the entrance ramp control in the step 4. Specific threshold speeds and interval divisions are given in table 5:
TABLE 5 threshold speed interval division Table
When the speed of the traffic flow is less than or equal to 50km/h, the free flow is converted into synchronous flow, and when the speed is less than 22km/h, the synchronous flow is converted into wide movement blockage; when the speed of the traffic flow is more than 25km/h, the wide movement blockage is converted into synchronous flow, and when the speed is more than 70km/h, the synchronous flow is converted into free flow.
4-ramp control and information release
4.1 ramp linkage control method
And taking the speed threshold value set in the previous step as one of the judgment conditions for judging the opening of the ramp, the adjustment of the ramp and the closing of the ramp. The traffic flow is switched to control the opening and closing of the ramp in the free flow, the synchronous flow and the wide motion blockage, and when the road is in the synchronous flow and wide motion blockage state, a corresponding entrance ramp control method is adopted.
And setting an entrance ramp control state by adopting three indexes of the upstream section vehicle speed, the downstream lane occupancy and the main line downstream flow. The road linkage control is realized by taking a speed threshold value for dividing a traffic flow phase as one of indexes for controlling an entrance ramp and combining a three-phase traffic flow theory and an entrance ramp control method, and a specific entrance ramp control state is shown in a table below.
TABLE 6 control conditions of entrance ramp
When the ramp adjustment is carried out, an ALINEA algorithm is used, the algorithm is improved, and the ramp adjustment rate is calculated by utilizing the predicted vehicle occupancy rate in the next control period, so that the ALINEA algorithm has a feedforward mechanism and a prediction mechanism. We take the same value for the adjustment period and the short prediction window, i.e. 50 s. The assignment mode can improve the engagement degree of the depth residual error model, the three-phase traffic flow and the ALINEA algorithm, so that the entrance ramp linkage control of the three algorithms becomes possible.
Let the current period be k-1 period, Oout(k-1) data collected by the vehicle detector in real time, Oout(k) The predicted value of k vehicle occupancy in the next period, r (k) is calculated from O in k-1 periodout(k-1) data is calculated by r (k) and Oout(k) And obtaining the predicted value of the ramp regulation rate in the k +1 period. The ALINEA algorithm has fixed green light duration, and the flow of vehicles merging into the main line is controlled by adjusting the time interval of starting of adjacent green lights in every minute.
The ALINEA algorithm is as follows:
before calculating the ramp regulation rate, the downstream expected occupancy, the regulation period and K are requiredRAnd calibrating the three parameters. Based on relevant research experience, it is recommended that the expected occupancy be set to 0.3, KRAnd is set to 70 in order to obtain the best control effect. Vehicle occupancy rate Oout(k-1)=0.37,Oout(k)=0.44。
And calculating the periodic ramp regulation rate r (k) value of 20s according to the measured data of the k-1 period vehicle collector, and substituting the improved ALINEA algorithm to calculate r (k +1) value of 20-70 multiplied by 0.14 and approximately 10 s. Namely, the total duration of the green light in the regulation period is 10s, the signal light controls the vehicle by switching the traffic light, the flashing time of the green light is fixed to 2s, the remaining 40s in the regulation period is distributed to the red light, the red light flashes for 5 times totally, and the duration of each time is 8 s.
4.2 information distribution
The first method is to use the calculation result of the algorithm to control the entrance ramp through the ramp signal lamp, and the control effect is the best; the other two types provide references for the decision of the driver to go through the congestion information release.
Ramp signal lamp. And controlling the vehicle to drive into the ramp by using the ramp signal lamp. In the green light period, allowing the vehicles to drive into the main road from the ramp; in the red light period, the vehicle must stop on the ramp for waiting and is not allowed to drive into the main road.
Road section upstream LED display panel. The networking LED display panel arranged at the upstream can inform the driver of the congestion in front of the road section, and guide the driver to select other road sections or change the traffic mode for traveling to avoid the congestion.
A mobile device. The road condition is published by mobile equipment such as broadcasting and navigation software, road congestion is prompted to the driver, and reference is provided for the trip decision of the driver.
4.3 error analysis
The three indexes of RMSE, MAE and MAPE can reflect the discrete degree and the actual error of the prediction data, and the prediction effect of the prediction model can be measured.
The formula for the three evaluation indices is as follows:
in the formula, xiRepresents actual traffic flow data, x 'at time i'iAnd N represents the length of the time sequence of the traffic flow to be evaluated.
The three index sizes calculated in this example are shown in the following table:
TABLE 7 prediction error analysis Table
Through the calculation of each index, it can be easily seen that the discrete degree of the predicted data in the embodiment is small, and the accuracy of data prediction is 97.06% as high. The traffic flow characteristic value change trend and prediction error analysis are shown in fig. 3, which shows that the data prediction by using the depth residual error network model has low complexity, high accuracy and stable prediction value change.
4.4 simulation and Effect evaluation
Because the process of embedding the vehicle detector is complex and the damage to the road surface is large, before the invention regulates and controls the road, the performance of the vehicle detector is evaluated by simulation software. And evaluating the control effect of the road by adopting six parameters of the maximum and minimum values of the downstream main line speed, the maximum and minimum values of the flow, the transverse and longitudinal wave properties of the downstream main line speed and the transverse and longitudinal wave properties of the downstream main line flow. The evaluation index respectively represents the amplitude of the change of the data sequence and the speed (or frequency) of the change of the data sequence by adopting the absolute mean value of the sequence difference and the standard deviation of the sequence, namely the transverse fluctuation and the longitudinal fluctuation, and the calculation formula of the evaluation index is as follows:
in the formula, xiFor the ith value in the data, Δ xi=xi-xi-1;Is the data mean. The simulation results are evaluated as shown in tables 8 and 9.
TABLE 8 downstream mainline velocity simulation evaluation results
Table 9 downstream mainline flow simulation evaluation results
As can be seen from table 8, there is a significant increase in the main line downstream speed under the ramp control condition. In addition, compared with the common ALINEA algorithm, the minimum value of the downstream main line speed is obviously improved after the improved ramp linkage control method is used.
After the ramp linkage control is adopted, the longitudinal fluctuation of the speed and the flow at the downstream of the main line is greatly improved compared with the longitudinal fluctuation of the speed and the flow when the control is not controlled and the traditional control is adopted, and the traffic condition at the downstream of the main line is quite sensitive to the regulation and control of the ramp control strategy. By regulation, the main line downstream is transited to a synchronous flow state through a blocking state; after the ramp linkage control is adopted, the transverse fluctuation of the downstream flow of the main line is small, which shows that the traffic jam is effectively relieved and the traffic condition of the downstream of the main line is improved by adopting the control strategy under the condition that the normal travel of the vehicle is not influenced.
Claims (4)
1. An entrance ramp linkage control method based on a depth residual error network model is characterized by comprising the following steps:
step 1: collecting traffic flow characteristic historical data, carrying out data-map conversion after data preprocessing, and converting the data into two-dimensional images taking time as a sequence;
the step 1 specifically comprises the following steps:
1.1 collecting historical data;
historical traffic flow data are collected by two methods, namely an unmanned aerial vehicle video shooting method and a vehicle detector method, and the road section traffic flow data and road section traffic flow state change conditions can be collected by the unmanned aerial vehicle video shooting method; the vehicle detector can collect traffic flow characteristic values of all lanes of the road section by detecting and processing;
defining characteristic value of traffic flow as vehicle flowAverage traffic speedTraffic densityAnd vehicle occupancyThe traffic flow characteristic value can be obtained by direct collection or indirect calculation, the vehicle occupancy adopts the time occupancy as an index, and the calculation formula is as follows:
for the length of time of each control cycle;first, theThe time of the vehicle passing through the section is second;the number of vehicles passing through the cross section in the time is measured;
1.2, preprocessing data;
preprocessing acquired data before digital image conversion, including abnormal data identification, abnormal data restoration and data denoising and normalization;
1.2.1 identifying abnormal data;
the collected traffic flow abnormal data mainly comprises three conditions:
traffic flow, speed and occupancy data are outside of reasonable threshold ranges;
the relationship between traffic flow, speed and occupancy data does not conform to the traffic flow theory;
the lack of traffic flow, speed and occupancy data;
aiming at the traffic flow data in the above situations, the data can be directly deleted or repaired;
1.2.2 abnormal data repair;
the traffic flow abnormal characteristic data mainly comprises two conditions of data error and data loss, and the identified data abnormal value is directly deleted; for missing data, performing data completion by adopting a K nearest distance neighbor method, determining K samples nearest to the sample with the missing data according to Euclidean distance or correlation analysis, and then performing weighted average on weighted values of the K samples to estimate the missing data of the sample;
1.2.3 denoising data;
denoising by adopting a first exponential smoothing algorithm, wherein the calculation formula is as follows:
in the formula (I), the compound is shown in the specification,andare respectively asThe smoothed data and the actual data of the time instants,taking 0.1 as a smoothing index;
1.2.4 data normalization;
converting all data to be processed into a [0,1] interval by adopting a Logistic/Softmax conversion method;
1.3 converting the number map;
carrying out two-dimensional image conversion on the preprocessed traffic flow data by taking time as a sequence to obtain a real-time-traffic flow characteristic data change trend graph under different road states, and storing the processed data in a picture form and using the processed data as a training and testing of a depth residual error network model after subsequent processing;
step 2: inputting image data, establishing and training a prediction model based on a deep residual error network model, setting hyper-parameters and network levels, carrying out parameter optimization through forward propagation and backward propagation, completing model training, and deeply learning the change trend of the historical traffic flow characteristic value;
the step 2 specifically comprises the following steps:
2.1 inputting and initializing data;
inputting image sample data based on TensorFlow, and performing feature set selection, different types of sample selection, sample vector diagram and feature diagram conversion, TFRecords sample data set generation and data set input;
reading a training sample in a TFRecords format, and coding data by adopting a One-bit effective coding One-Hot mode according to a sample label;
2.2 setting the super-parameter;
carrying out hyper-parameter setting before training a deep residual error network model, wherein main setting parameters comprise batch training size, learning rate, weight attenuation rate and optimizer selection;
the learning rate is set in a proper range, so that the gradient of the model is reduced to an optimal value, a larger initial learning rate is set at first, and the initial learning rate is gradually adjusted to the minimum learning rate along with the increase of the iteration times of the model so as to obtain a faster training speed and a faster model precision; setting L by weight attenuation ratio2Regularizing the parameter of the item, adjusting the influence of the complexity of the model on a loss function, and preventing the overfitting of the model; selecting a Momentum optimizer, and improving the convergence speed of the loss function based on the gradient moving index weighted average;
2.3 setting network model hierarchy;
the first layer of the network model is a convolution layer and is responsible for extracting low-level features;
the second layer is a maximum pooling layer, the estimated mean shift is reduced, and the image texture information is reserved;
3-6 layers are convolution layers and are responsible for extracting high-level features;
the seventh layer is an average pooling layer, and the features extracted by the convolutional layer are calculated and input into the full-connection layer;
the full connection layer is responsible for outputting classification probability;
2.4 forward propagation;
in the forward propagation training process, an expected learning objective function is set firstly, and the function is set as follows:
whereinIn order to input the characteristic value, the characteristic value is input,in order for the model to predict the probability of outcome,andparameters obtained for model training;
performing feature sparse extraction on a plurality of convolution layers, calculating the extracted sparse convolution features by using a mean pooling operation, converting each batch of input sample images into sparse features, entering a full connection layer, and performing logits calculation to obtain a [ batch training size multiplied by classification number ] classification probability matrix of each type of the batch of sample data;
the softmax operation ensures that all output is positive, all row values of the matrix are extended to a [0,1] interval, the probability sum of any row is equal to 1, the softmax operation is performed on the extended matrix, and the maximum value of each row is the value with the maximum output probability, namely the prediction result of the training;
2.5, back propagation and parameter tuning;
in the process of training the depth residual error model, the convolutional layer extracts the layer-by-layer calculation sparse features of each batch of sample data and records corresponding parameter values, the sparse features extracted from the bottommost layer are input to the logits layer, and the sample classification value is calculated;
the loss function of each training is calculated as the cross entropy of the real type of the sample and the prediction result of the model, and the training loss function of each batch of samples is calculated as follows:
in the formula (I), the compound is shown in the specification,is a sampleBelong to the categoryThe probability of (a) of (b) being,is a sampleBelong to the categoryThe model prediction probability of (3);
calculating to obtain a model loss function by comparing the actual and model prediction and identifying classification results, and performing back propagation on model fitting errors;
inputting training set data, performing parameter tuning, selecting a Momentum optimizer based on gradient moving index weighted average, and performing smoothing processing on network parameters;
let the current iteration step beThe calculation formula based on the Momentum optimization algorithm is as follows:
in the above formula, the first and second light sources are,W、bis an update formula of the network weight vector and the bias vector, alpha is the learning rate of the network,
dwthe gradient obtained when the loss function in the network weight vector updating formula is propagated reversely is obtained,dbthe gradient found when the loss function in the offset vector update formula propagates backwards,the gradient momentum accumulated by the loss function in the formula during the previous t-1 iteration is updated for the network weight vector,the gradient momentum accumulated by the loss function in the formula during the previous t-1 iterations is updated for the bias vector,is an index value of the gradient accumulation, set to 0.9;
after parameter tuning is finished, performing the last step of model training, inputting verification set data, testing the performance of the model, and manually fine-tuning the value of the hyper-parameter;
and step 3: collecting real-time traffic flow characteristic data, performing figure-to-figure conversion on the preprocessed data, and converting the data into two-dimensional images taking time as a sequence; inputting the image data into a trained depth residual error network model for short-time prediction after conversion, outputting a short-time change trend prediction graph, and converting the prediction trend graph into text data by utilizing graph number conversion;
and 4, step 4: the converted text data carries out short-time prediction on the road traffic characteristic value by utilizing a trained deep residual error network model, the prediction data and the recognized traffic flow state are input into an entrance ramp control ALINEA algorithm, the vehicle occupancy and the maximum queuing vehicle number in the next control period are calculated, and the traffic flow which is converged into a main line is subjected to linkage control in advance;
and 5: and (4) carrying out simulation evaluation and analysis on the ALINEA algorithm by using a VB + VISSIM program, and issuing road condition information.
2. The entrance ramp linkage control method based on the depth residual error network model according to claim 1, wherein step 3 specifically comprises the following steps:
real-time traffic flow characteristic data are collected by two methods of video shooting by an unmanned aerial vehicle and a vehicle detector, and phase conversion judgment conditions are set for vehicle tracks moving in the traffic flow: in the exceeding of the given threshold time interval, the speed of the vehicle traveling along the track becomes lower or higher than the given threshold speed of the phase transition point;
setting a data input time window, inputting real-time traffic characteristic values collected by a vehicle detector, preprocessing the data and converting the data into a time sequence change diagram by using an input depth residual error network model, outputting a traffic characteristic prediction data change trend diagram of the next period of time, and converting a predicted and output two-dimensional image into text data through the diagram number;
dividing the traffic flow into three phases of free flow F, synchronous flow S and wide movement blockage J by combining the space-time characteristics of the road traffic flow and referring to a three-phase traffic flow theory proposed by Konne;
traffic flow phase change is considered as a gradual phase change process from free flow to synchronous flow to wide motion blockage: f → S → J; and setting each phase transition point by taking the speed as a threshold value in combination with the current road situation of China.
3. The ingress ramp linkage control method based on the depth residual error network model according to claim 1, characterized in that: 4, short-time prediction is carried out on the road traffic characteristic value, wherein the short-time prediction comprises prediction data of traffic flow speed, traffic flow density and vehicle occupancy;
the linkage control method of the entrance ramp specifically comprises the following steps:
before using ALINEA algorithm, defineControl of occupancy of main line downstream vehicles within a cycleIs collected by a vehicle detector and is used for detecting the vehicle,downstream vehicle occupancy within a control cycleAnd rate of arrival of on-ramp vehiclesPredicting by a depth residual error network model;
byWithin a periodThe data is calculated byAndcan obtainPredicting the adjustment rate of the periodic ramp;
the ALINEA algorithm green light duration is fixed, and the flow of vehicles merging into the main line is controlled by adjusting the time interval of starting of adjacent green lights within every minute;
in one control cycle, the regulation rate is calculated by the formula:
in the formula (I), the compound is shown in the specification,is the firstControlling the ramp regulation rate of the periodic calculation;is the firstThe ramp regulation rate in the control period is,byCalculating actual measurement data of the vehicle occupancy in the control period, wherein the adjustment rate is the green light duration in one control period and the unit is s;external disturbance fixed in parameter adjustment feedback control;is the expected occupancy downstream of the main line;is the firstAnd (4) controlling the vehicle occupancy prediction value at the downstream of the main line in the period.
4. The entrance ramp linkage control method based on the depth residual error network model according to claim 1, wherein the simulation evaluation and analysis in step 5 specifically comprises:
the VB + VISSIM 4.3COM development program is used for realizing ramp control based on the ALINEA algorithm; loading a road network model in a program for simulation, wherein the simulation running time is 3600s, corresponding to one hour of an actual peak, and in the simulation process, each control cycle returns simulation state data comprising green light ending time, green signal ratio and occupancy rate;
setting the signal control period asIn units of s, program at the fourth of each cycle-1 acquisition of occupancy of data returned by the detector at time, detector per interval-2, returning the primary occupancy and calculating the green ratio of the next period according to the adjustment coefficient and the optimal occupancy;
when the green signal ratio is greater than or equal to 0.8, controlling the continuous green light by the ramp signal; less than 0.2, continuously red light is controlled by the ramp signal; and carrying out fixed period control on the green signal ratio between 0.8 and 0.2, and calculating the signal lamp time length of the ramp control according to the optimized green signal ratio.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910809931.9A CN110503833B (en) | 2019-08-29 | 2019-08-29 | Entrance ramp linkage control method based on depth residual error network model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910809931.9A CN110503833B (en) | 2019-08-29 | 2019-08-29 | Entrance ramp linkage control method based on depth residual error network model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110503833A CN110503833A (en) | 2019-11-26 |
CN110503833B true CN110503833B (en) | 2021-06-08 |
Family
ID=68590510
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910809931.9A Active CN110503833B (en) | 2019-08-29 | 2019-08-29 | Entrance ramp linkage control method based on depth residual error network model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110503833B (en) |
Families Citing this family (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111260922B (en) * | 2020-01-20 | 2021-01-29 | 浙江工业大学 | Ramp control method based on congestion situation classification |
CN112200077A (en) * | 2020-04-15 | 2021-01-08 | 陈建 | Artificial intelligent image processing method and system based on intelligent traffic |
CN111785088B (en) * | 2020-06-23 | 2021-08-20 | 大连理工大学 | Double-layer collaborative optimization method for merging network vehicle ramps |
CN111949703B (en) * | 2020-07-09 | 2023-06-06 | 广东工业大学 | Unmanned aerial vehicle deployment and flight trajectory optimization method and system for intelligent traffic |
CN114501010B (en) * | 2020-10-28 | 2023-06-06 | Oppo广东移动通信有限公司 | Image encoding method, image decoding method and related devices |
CN112669594B (en) * | 2020-12-11 | 2022-12-27 | 国汽(北京)智能网联汽车研究院有限公司 | Method, device, equipment and storage medium for predicting traffic road conditions |
CN113221442B (en) * | 2020-12-24 | 2022-08-30 | 山东鲁能软件技术有限公司 | Method and device for constructing health assessment model of power plant equipment |
CN112734703A (en) * | 2020-12-28 | 2021-04-30 | 佛山市南海区广工大数控装备协同创新研究院 | PCB defect optimization method by utilizing AI cloud collaborative detection |
CN113125992B (en) * | 2021-04-23 | 2022-07-19 | 合肥工业大学 | NPC three-level inverter fault diagnosis method and system based on DBN |
CN113256973B (en) * | 2021-05-11 | 2022-03-25 | 青岛海信网络科技股份有限公司 | Peak start time prediction method, device, equipment and medium |
CN113379156A (en) * | 2021-06-30 | 2021-09-10 | 南方科技大学 | Speed prediction method, device, equipment and storage medium |
CN114141029B (en) * | 2021-11-25 | 2022-11-18 | 东南大学 | Ramp control method based on offline reinforcement learning and macroscopic model |
CN115083038B (en) * | 2022-06-29 | 2024-02-27 | 南京理工大学泰州科技学院 | Device and method for detecting abnormal running of vehicle |
CN115662114A (en) * | 2022-10-08 | 2023-01-31 | 广州玩鑫信息科技有限公司 | Intelligent traffic system for relieving congestion based on big data and operation method thereof |
CN115544901B (en) * | 2022-11-28 | 2023-02-21 | 西南石油大学 | Intelligent thick oil working condition fault identification method for small sample |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109035779A (en) * | 2018-08-30 | 2018-12-18 | 南京邮电大学 | Freeway traffic flow prediction technique based on DenseNet |
CN109816983A (en) * | 2019-02-26 | 2019-05-28 | 昆明理工大学 | A kind of short-term traffic flow forecast method based on depth residual error network |
CN109903557A (en) * | 2019-03-04 | 2019-06-18 | 南京邮电大学 | Based on the freeway traffic flow prediction technique for improving independent loops neural network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10102744B2 (en) * | 2016-09-27 | 2018-10-16 | International Business Machines Corporation | Predictive traffic management using virtual lanes |
-
2019
- 2019-08-29 CN CN201910809931.9A patent/CN110503833B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109035779A (en) * | 2018-08-30 | 2018-12-18 | 南京邮电大学 | Freeway traffic flow prediction technique based on DenseNet |
CN109816983A (en) * | 2019-02-26 | 2019-05-28 | 昆明理工大学 | A kind of short-term traffic flow forecast method based on depth residual error network |
CN109903557A (en) * | 2019-03-04 | 2019-06-18 | 南京邮电大学 | Based on the freeway traffic flow prediction technique for improving independent loops neural network |
Non-Patent Citations (1)
Title |
---|
城市快速路入口匝道流量控制研究;方传武;《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》;20170115;第C034-377页 * |
Also Published As
Publication number | Publication date |
---|---|
CN110503833A (en) | 2019-11-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110503833B (en) | Entrance ramp linkage control method based on depth residual error network model | |
CN110097755B (en) | Highway traffic flow state identification method based on deep neural network | |
CN110164128B (en) | City-level intelligent traffic simulation system | |
CN112241814B (en) | Traffic prediction method based on reinforced space-time diagram neural network | |
CN109785618B (en) | Short-term traffic flow prediction method based on combinational logic | |
CN111612243A (en) | Traffic speed prediction method, system and storage medium | |
CN112949828B (en) | Graph convolution neural network traffic prediction method and system based on graph learning | |
CN112289034A (en) | Deep neural network robust traffic prediction method based on multi-mode space-time data | |
CN108986453A (en) | A kind of traffic movement prediction method based on contextual information, system and device | |
CN112052818B (en) | Method, system and storage medium for detecting pedestrians without supervision domain adaptation | |
CN111598325A (en) | Traffic speed prediction method based on hierarchical clustering and hierarchical attention mechanism | |
CN111046773B (en) | Method for judging accumulated water on road surface based on image technology | |
CN110956146B (en) | Road background modeling method and device, electronic equipment and storage medium | |
CN113591617B (en) | Deep learning-based water surface small target detection and classification method | |
CN107403222A (en) | A kind of motion tracking method based on auxiliary more new model and validity check | |
CN115034457A (en) | Urban rail transit short-time passenger flow prediction method based on CEEMDAN-LSTM | |
CN116596044B (en) | Power generation load prediction model training method and device based on multi-source data | |
CN114565187A (en) | Traffic network data prediction method based on graph space-time self-coding network | |
CN115206092A (en) | Traffic prediction method of BiLSTM and LightGBM model based on attention mechanism | |
CN112766603A (en) | Traffic flow prediction method, system, computer device and storage medium | |
CN114694382B (en) | Dynamic one-way traffic control system based on Internet of vehicles environment | |
CN114970815A (en) | Traffic flow prediction method and device based on improved PSO algorithm optimized LSTM | |
CN113627240B (en) | Unmanned aerial vehicle tree species identification method based on improved SSD learning model | |
CN114596726A (en) | Parking position prediction method based on interpretable space-time attention mechanism | |
CN113806857A (en) | High-speed train energy-saving braking method based on variational graph self-encoder |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |