CN110634294A - Time-space traffic flow prediction method driven by reinforced hierarchical learning - Google Patents
Time-space traffic flow prediction method driven by reinforced hierarchical learning Download PDFInfo
- Publication number
- CN110634294A CN110634294A CN201910920367.8A CN201910920367A CN110634294A CN 110634294 A CN110634294 A CN 110634294A CN 201910920367 A CN201910920367 A CN 201910920367A CN 110634294 A CN110634294 A CN 110634294A
- Authority
- CN
- China
- Prior art keywords
- traffic flow
- road
- correlation
- predicted
- time
- 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.)
- Granted
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
- G08G1/0125—Traffic data processing
- G08G1/0133—Traffic data processing for classifying traffic situation
Landscapes
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention provides a time-space traffic flow prediction method driven by reinforced hierarchical learning, which fully utilizes the correlation of relevant road sections in time and space, dynamically simulates a nonlinear, high-dimensional and random road traffic flow evolution mode through a reinforced hierarchical learning network, designs and realizes the extraction of road traffic flow characteristics based on a restricted Boltzmann machine model, further reduces the dimension of the recorded road traffic flow data, classifies the road traffic flow characteristics after dimension reduction by an SVM method, obtains the final traffic flow prediction result, and tests and field tests show that when the sample reliability is 75%, the prediction result accuracy is over 85.6%, when the sample reliability is over 90%, the prediction result accuracy is over 96.3%, the traffic flow prediction accuracy and reliability are greatly improved, and the theoretical basis of the prediction method is very strong, the traffic flow prediction timeliness is good, the prediction result real-time performance is strong, and the application space is wide.
Description
Technical Field
The invention relates to a space-time traffic flow prediction method, in particular to a space-time traffic flow prediction method driven by reinforced hierarchical learning, and belongs to the technical field of traffic flow prediction.
Background
With the high-speed development of urban road traffic, traffic pressure is higher and higher, and road traffic congestion is more and more serious, so that the intelligent road traffic system plays an important role in relieving the urban road traffic pressure, improving urban life and work quality, reducing environmental pollution and the like, and road traffic conditions, particularly traffic flow prediction, are a very important part in the road traffic condition prediction. Because the road traffic flow has the characteristics of nonlinearity, high dimensionality, randomness and the like, a large amount of road traffic flow information contains rich information and plays a very key role in predicting the traffic flow, the comprehensive excavation of traffic flow characteristics has important significance in improving the predicting speed and precision of the road traffic flow, and the accurate traffic flow condition prediction can help people to plan a trip route better, improve the working and living quality and efficiency and reduce the environmental pollution and the energy consumption.
Because the road traffic flow has the characteristics of nonlinearity, high dimensionality, randomness and the like, the uncertain factors are many, and the road traffic flow prediction difficulty is high. The urban road traffic condition prediction in the prior art mainly depends on a large amount of historical data and real-time road traffic flow information, belongs to the problem of pattern recognition, and does not really realize intelligent prediction, at present, the research on the road traffic condition prediction at home and abroad is mostly based on a probability statistical method of a time sequence and the prediction analysis of space correlation among roads, the methods do not comprehensively consider a space-time change pattern of the road traffic flow, are based on the direct application of traffic data or shallow learning, and do not fully utilize abundant dynamic evolution patterns and laws contained in a large amount of traffic flow data.
In summary, the traffic flow prediction method in the prior art mainly has the following defects: firstly, a great deal of work in the prior art is to use the evolution law of the road traffic flow on a time sequence to predict the traffic flow, an autoregressive moving average model is generally adopted, the autoregressive moving average model is a random sequence formed by the road traffic flow changing along with time, the time change law of the random sequence reflects the continuity of the traffic flow data on time, so that the traffic flow state of a time point to be predicted is obtained by mining the time sequence law in historical data, but the traffic flow prediction methods based on the time sequence characteristics are all statistical methods only depending on the law of the traffic flow data on time, and the accuracy and the reliability of the traffic flow prediction are very low. Secondly, there are some methods in the prior art that the evolution trend of the road traffic flow is analyzed by using the spatial correlation in the road traffic network, but the method is based on an assumption that the traffic flow density on the road is only related to the traffic flow density on the road section adjacent to the road, so that the method only uses the correlation between the adjacent road and the predicted road to predict the flow of the road to be predicted, and the influence of the distant road on the predicted road and the correlation of the distant road and the road to be predicted on the time space are not fully considered. Thirdly, in the prior art, some reinforcement learning methods are used for extracting typical characteristics in complex data such as traffic flow by using a multilayer network structure, but the reinforcement learning methods used in the prior art are static networks, cannot well express dynamic modes in a road traffic flow network, and road traffic flow data need to be completely recorded in, and no indexes are set for measuring the space-time correlation among road sections in an area, so that a lot of redundancy is caused, the training effect and the speed of the network are seriously influenced, the traffic flow prediction timeliness is low, and the prediction result lags seriously.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a time-space traffic flow prediction method driven by reinforced hierarchical learning, which fully utilizes the correlation of relevant road sections in time and space, dynamically simulates a nonlinear, high-dimensional and random road traffic flow evolution mode through a reinforced hierarchical learning network, designs and realizes the extraction of the road traffic flow characteristics based on a restricted Boltzmann machine model, further reduces the dimension of the recorded road traffic flow data, classifies the road traffic flow characteristics after the dimension reduction by an SVM method to obtain the final prediction result, tests and field detection show that when the sample reliability is 75%, the prediction result accuracy is over 85.6%, when the sample reliability is 90%, the prediction result accuracy is over 96.3%, the traffic flow prediction accuracy and reliability are greatly improved, and the dynamic intelligence is obviously improved, the traffic flow prediction method has the advantages of strong theoretical basis, good timeliness of traffic flow prediction, strong real-time prediction result and wide application space.
In order to achieve the technical effects, the technical scheme adopted by the invention is as follows:
a time-space traffic flow prediction method driven by reinforced hierarchical learning comprises the steps of extracting input data, driving and extracting traffic flow characteristics by a limited Boltzmann machine model, predicting road traffic flow based on an SVM model, extracting the input data, including high-correlation road selection extraction and original data compression based on principal component analysis, and specifically comprises the following steps:
firstly, selecting and extracting high-correlation roads;
secondly, compressing the original data based on principal component analysis;
thirdly, driving and extracting traffic flow characteristics by the limited boltzmann machine model;
and fourthly, predicting the road traffic flow based on the SVM model.
A method for predicting the time-space traffic flow driven by reinforced hierarchical learning includes such steps as evaluating the influence of different roads on predicted road at different time by a correlation measuring mode, choosing the Pearson product-moment correlation coefficient from the high-correlation roads to measure the correlation between the traffic flow of each road at different time and the traffic flow of road to be predicted at the time point, comparing the calculation results to order the correlation between them, choosing the first h high-correlation roads and their relative time points, obtaining corresponding parameters including the distance between the road and the road to be predicted, the traffic flow and the state of the road at the time point, the length and the width of the road, further compressing the information by a principal component analysis method, extracting the most main components from a series of mutually contained and redundant information by the principal component analysis method to represent all information, extracting only the first r principal components by the principal component analysis method, then representing all relevant information of the whole h relevant roads and corresponding time points by the information, training the compressed data more effectively to strengthen a hierarchical learning network, inputting the data to a limit Boltzmann machine to obtain and record hidden layer characteristics of which the data contain the same principal characteristics, extracting the essential characteristics of the traffic flow of the road through network iteration of the limited Boltzmann machine, and finally classifying the extracted essential characteristics of the traffic flow by an SVM, and the SVM maps the low-dimensional data to a high-level feature space, finds a hyperplane with an optimal interval, and classifies the data to obtain a final traffic flow prediction result.
3. The method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: extracting the logging data sequentially comprises two steps of high-correlation road selection extraction and original data compression based on principal component analysis, on the aspect of logging layer data selection, firstly, according to the calculated correlation coefficient of Pearson's product moment of the correlation road and the predicted road, the front h roads with the highest correlation degree are selected and extracted, then, from the relevant information of the h roads, including average flow, time interval, geographical distance of traffic flow and traffic flow condition, road width and length of the h road at the f time point, the principal component capable of reflecting all information is extracted by using a principal component analysis method, and the extraction of the original logging data is completed.
A method for predicting the time-space traffic flow driven by reinforced hierarchical learning further comprises the following steps that in the step of selecting and extracting high-correlation roads, a Pearson product moment correlation coefficient is used as a measurement index to select the roads highly correlated with the road to be predicted, and time intervals are used as input variables:
wherein C represents a correlation coefficient between road flow rates, wherein AiAnd BiTraffic flow observations of an adjacent road a and a road B to be predicted respectively,and
respectively is the mean value of two variables, C describes the degree of the linear correlation between the adjacent road A and the road B to be predicted, the value of C is between-1 and +1, if C is more than 0, the adjacent road A and the road B to be predicted are in positive correlation, namely the flow value of the adjacent road is larger, the flow value of the road to be predicted is correspondingly larger, if C is less than 0, the adjacent road and the road to be predicted are in negative correlation, namely the flow value of the adjacent road is larger, the flow value of the road to be predicted is largerConversely, the smaller the value, the larger the absolute value of C is, the higher the correlation between the adjacent road and the road to be predicted is;
setting the current time point as F, firstly finding out a road set A which can reach a road B to be predicted in a time range F delta F, wherein delta F is a time interval, and the g-th traffic flow sampling data set of each road in a record is calculated according to e days:
{Ag(f),Ag(f-Δf),Ag(f-2Δf),…,Ag(f-nΔf)}(g＝1,2,…,e)，
wherein, the road AiThe g-th day traffic flow sample dataset in the record is:
{Ai g(f),Ai g(f-Δf),Ai g(f-2Δf),…,Ai g(f-nΔf)},
the g-th day traffic flow sampling data set of the road B to be predicted in the record is Bg(F + F. DELTA.f), road AiThe degree of correlation between the traffic flow at k (j ═ 1,2, …, n) time intervals after the time point F and the traffic flow of the road B to be predicted at the time point F + F Δ F to be predicted is calculated by the following formula:
wherein
When C is presentiThe closer the absolute value of (f + k Δ f) is to 1, the higher the correlation.
A method for predicting the time-space traffic flow driven by reinforced hierarchical learning further includes the second step of selecting and extracting h pieces of information of high-correlation roads from original data compression based on principal component analysis to form an original matrix of 6 x h:
K＝[k(1)k(2) … k(6)]F∈R6×h
wherein k (i) ═ k1(i)k2(i)，...，kh(i)]∈RhI-th (i-1, 2.., 6) information representing h related roads is obtainedThe data regularization yields:
A＝[a(1)a(2) … a(6)]F∈R6×h
wherein
Is the variance of the ith information of the jth relevant road segment,
is the average value of the ith information of the jth related road, and the covariance matrix of the road traffic flow is as follows:
wherein
All characteristic values of the covariance matrix can be obtained through matrix calculation, and the sequence is that lambda 1 is more than or equal to lambda 2.
Because the principal component contains the most important information, only part of the principal component needs to be extracted, the information contained in all the components can be reflected, and the contribution rate of each characteristic value is as follows:
the cumulative contribution of the first r principal components is ∑ αrWhen the contribution ratio exceeds a constant ψ, only the first r principal components need to be extracted to represent all the component information:
a method for predicting a space-time traffic flow driven by reinforced hierarchical learning further comprises a third step of utilizing a deep model of a limited Boltzmann machine to model a traffic network in driving and extracting traffic flow characteristics, extracting abstract characteristics from the traffic flow and facilitating subsequent classification;
h traffic flow characteristics of r main components extracted by a main component analysis method are used as traffic flow characteristic parameter input vectors v-v (v is v1,v2,...,vrh)FThe output is q hidden feature vectors s ═ s(s)1,s2,...,sq)F. Each traffic flow characteristic recording node is only related to q hidden characteristic nodes;
the limited Boltzmann machine network comprises several parameters, namely weight M between a traffic flow characteristic recording layer and a hidden characteristic layerijAnd the second is offset vector a of traffic flow characteristic input node (a ═ a)1,a2,...,ars)FAnd thirdly, the offset vector b of the hidden feature node is equal to (b)1,b2,...,bq)F，MijAnd a and b determine that the limited Boltzmann machine network encodes an rh-dimension traffic flow characteristic input sample into a q-dimension hidden characteristic sample, and for a given parameter theta (M)ij,ai,bj) The limited boltzmann machine model has the following energy function value between a traffic flow characteristic input vector v and a hidden characteristic layer output vector s:
in the road traffic flow prediction based on an SVM model, a limited Boltzmann machine extracts traffic flow characteristics and then uses the extracted traffic flow characteristics as input of an SVM, and the input is mapped into a training data set on a characteristic space:
F＝{(A1,B1),(A2,B2),...,(AN,BN)}
wherein A isi＝(A11,A12,...,A1q)，i＝1,2,...,N，AiIs the ith characteristic vector formed by inputting and mapping traffic flow data, BiIs AiThe class mark of (1), namely a traffic flow state mark;
the SVM learning aims at finding an optimal separation hyperplane in a feature space formed by inputting and mapping traffic flow data and classifying feature vectors into five different traffic flow states, wherein the equation corresponding to the separation hyperplane is f (A) wFA + B, determined by the normal vector w and the intercept B, when (a, B) this example is on the hyperplane, w a + B is 0, | w a + B | can represent the distance from the feature vector a to the hyperplane, and whether the classification is correct can be judged by observing whether the symbol of w a + B is consistent with the symbol of the traffic flow state mark B;
wherein the spacing of the maximum spaced classification hyperplane is represented by the geometric spacing:
the goal of the maximum interval classifier is to find maxJ, which is at the corresponding constraint bar Bi(wFAi+ b) is not less than 1, i is 1,2,.. times, n, further converted to:
because of the need to
Is equivalent to obtaining
The above objective function is equivalent to:
due to the special structure of the problem, the problem can be optimized by changing Lagrange duality into dual variables, namely, the optimal solution of the original problem, namely the dual algorithm of the SVM under the linear separable condition, is obtained by solving the dual problem equivalent to the original problem;
through linear SVM classification, the probability vector of 5 states of the traffic flow at the time stamp f can be finally obtained:
p (f) ("p (state ═ congestion"), p (state ═ slowness "), p (state ═ normality"), p (state ═ smoothness ") ]
And after SVM classification, obtaining a final road traffic flow prediction result according to the probability vector of the traffic flow at the time of the timestamp f.
Compared with the prior art, the invention has the advantages that:
1. the invention provides a time-space traffic flow prediction method driven by reinforced hierarchical learning, which fully utilizes the correlation of relevant road sections in time and space, dynamically simulates a nonlinear, high-dimensional and random road traffic flow evolution mode through a reinforced hierarchical learning network, designs and realizes the extraction of road traffic flow characteristics based on a restricted Boltzmann machine model, further reduces the dimension of the recorded road traffic flow data, classifies the road traffic flow characteristics after dimension reduction by an SVM method to obtain the final prediction result, and tests and field tests show that when the reliability of a sample is 75%, the accuracy of the prediction result is more than 85.6%, and when the reliability of the sample is 90%, the accuracy of the prediction result is more than 96.3%, and the accuracy of the traffic flow prediction method is high.
2. The invention provides a time-space traffic flow prediction method driven by reinforced hierarchical learning, which solves the problems that the prior art utilizes the evolution rule of road traffic flow on a time sequence to predict the traffic flow, the time-sequence characteristic-based traffic flow prediction method is a statistical method only depending on the rule of traffic flow data on time, the accuracy and the reliability of the traffic flow prediction are very low, the time-space traffic flow prediction method driven by the reinforced hierarchical learning is not a simple statistical method based on samples, low-dimensional data are mapped to a high-level characteristic space through network iteration, a hyperplane with an optimal interval is found, the data are classified, the accuracy and the reliability of the traffic flow prediction are greatly improved, and the real-time performance and the dynamic intelligence are obviously improved.
3. The invention provides a method for predicting a space-time traffic flow driven by reinforced hierarchical learning, which solves the problems that the evolution trend of a road traffic flow is analyzed only by utilizing spatial correlation in a road traffic network and only the assumption is made that the traffic flow density on a road is only related to the traffic flow density on a road section adjacent to the road, the correlation between the adjacent road and a predicted road is utilized to predict the flow of the road to be predicted, the influence of a farther road on the predicted road and the correlation of the road to be predicted on the time space are fully considered, so that the theoretical basis of the traffic flow prediction method is very strong, the prediction is more scientific, stable and reliable, and the prediction result is more comprehensive and clear.
4. The invention provides a time-space traffic flow prediction method driven by reinforcement hierarchical learning, which solves the problems that in the prior art, all methods using reinforcement learning are static networks, dynamic modes in a road traffic flow network cannot be well expressed, road traffic flow data needs to be completely recorded in, and no indexes are set to measure the time-space correlation between road sections in an area, so that a lot of redundancy is caused.
Drawings
FIG. 1 is a step diagram of a method for predicting spatiotemporal traffic flow driven by reinforced hierarchical learning provided by the present invention.
Fig. 2 is a schematic view of a traffic network in the high-correlation road selection extraction according to the present invention.
FIG. 3 is a schematic diagram of a feature extraction model of a restricted Boltzmann machine provided by the present invention.
Detailed Description
The technical scheme of the method for predicting the spatio-temporal traffic flow driven by the reinforced hierarchical learning provided by the invention is further described below with reference to the accompanying drawings, so that the technical scheme can be better understood and implemented by those skilled in the art.
Referring to fig. 1, the invention provides a reinforced hierarchical learning-driven space-time traffic flow prediction method, which comprises the steps of extracting recorded data, extracting traffic flow characteristics under the drive of a limited boltzmann model, and predicting road traffic flow based on an SVM model, wherein the step of extracting the recorded data comprises high-correlation road selection extraction and raw data compression based on principal component analysis, and the specific steps are as follows:
firstly, selecting and extracting high-correlation roads;
secondly, compressing the original data based on principal component analysis;
thirdly, driving and extracting traffic flow characteristics by the limited boltzmann machine model;
and fourthly, predicting the road traffic flow based on the SVM model.
The traffic flow of the road at a certain future time is formed by converging the traffic flows of adjacent or far roads, but the influence time and the influence degree between the adjacent or far roads are different, particularly, the traffic flow of the road close to the predicted road is converged to the predicted road in a shorter time; the influence of different roads on the predicted road at different times is evaluated in a correlation measurement mode, the high-correlation road selects and selects a Pearson product moment correlation coefficient to measure the correlation between the traffic flow of each road at different moments and the traffic flow of the road to be predicted at the prediction time point, the absolute value of the Pearson product moment correlation coefficient is closer to 1, which indicates that the correlation between the traffic flows of the two roads is higher, the correlation is ranked according to the comparison calculation result, the calculation result is finally selected from high to low, the first h high-correlation roads and the corresponding time point are selected, namely, the correlation between the traffic flow of the road at the time point and the traffic flow of the road to be predicted at the prediction time point is high, and the part of the traffic flow of the road to be predicted at the prediction time point is determined by the correlation of the road at the time point Traffic flow at intermediate points converges. After the first h related roads with high relevance and the time point are picked, corresponding parameters are obtained, wherein the corresponding parameters comprise the distance between the road and the road to be predicted, the traffic flow and the state of the road at the time point, and the length and the width of the road. The information has mutual inclusion relationship, the data has certain redundancy, so the information is further compressed by a principal component analysis method, the principal component analysis extracts the most main components from a series of mutually contained and redundant information to represent all the information, only the first r principal components are extracted by the principal component analysis method, the information can represent all the relevant information of the whole h relevant roads and corresponding time points, the compressed data can train a reinforced hierarchical learning network more effectively and input the reinforced hierarchical learning network into a limiting Boltzmann machine, the limiting Boltzmann machine is a very common reinforced hierarchical learning method, the core idea of the reinforced hierarchical learning is to obtain and record hidden layer characteristics of the data containing the same main characteristics through network iteration, the network iteration of the limited Boltzmann machine is performed, the method comprises the steps of extracting essential features of road traffic flow, better using the extracted essential features for classification prediction, and finally classifying the extracted essential features of the traffic flow through an SVM (support vector machine), wherein the SVM is used for mapping low-dimensional data to a high-level feature space, so that the hyperplane with the optimal interval can be found more conveniently, and the data are classified to obtain the final traffic flow prediction result.
Firstly, extracting the recorded data
The quality of the input data is very important for training the reinforced hierarchical learning network, the input data is good in selection, the network can be trained better and faster by the reinforced hierarchical learning, and the input data is high in reliability, large in data density and wide in coverage range as far as possible.
In the data selection of the input layer, firstly, according to the calculated correlation coefficient of the Pearson product moment of the correlated road and the predicted road, the front h roads with the highest correlation degree are selected and extracted, then, the related information of the h roads, including the average flow, the time interval, the geographic distance of the traffic flow and the traffic flow condition, the road width and the length of the h road at the f time point, is extracted by using a principal component analysis method, and the principal component capable of reflecting all the information is extracted, thereby completing the extraction of the original input data.
(one) high-connected road selection extraction
The traffic flow of a road in a future time period is predicted in a traffic network, the traffic flow of the road at the time point is formed by merging the traffic flows of the adjacent roads at different previous time points, the closer the road is to the road to be predicted, the traffic flow can be merged onto the road to be predicted in a shorter time, the longer the road is away from the road to be predicted, the longer the time for merging the traffic flow onto the road to be predicted is, and the road and the time period which are highly related to the road to be predicted are selected as the input variables by using the Pearson's product moment correlation coefficient as the measurement index:
wherein C represents a correlation coefficient between road flow rates, wherein AiAnd BiTraffic flow observations of an adjacent road a and a road B to be predicted respectively,
and
the average values of the two variables are respectively described by C, the degree of the strength of the linear correlation between the adjacent road A and the road B to be predicted is described by C, the value of C is between-1 and +1, if C is greater than 0, the adjacent road A and the road B to be predicted are in positive correlation, namely the flow value of the adjacent road is larger, the flow value of the road to be predicted is correspondingly larger, if C is less than 0, the adjacent road and the road to be predicted are in negative correlation, namely the flow value of the adjacent road is larger, the flow value of the road to be predicted is smaller on the contrary, and the larger the absolute value of C is, the higher the correlation degree between the adjacent road and the road to be predicted is.
Setting the current time point as F, firstly finding out a road set A which can reach a road B to be predicted in a time range F delta F, wherein delta F is a time interval, and the g-th traffic flow sampling data set of each road in a record is calculated according to e days:
{Ag(f),Ag(f-Δf),Ag(f-2Δf),…,Ag(f-nΔf)}(g＝1,2,…,e)，
wherein, the road AiThe g-th day traffic flow sample dataset in the record is:
{Ai g(f),Ai g(f-Δf),Ai g(f-2Δf),…,Ai g(f-nΔf)},
the g-th day traffic flow sampling data set of the road B to be predicted in the record is Bg(F + F. DELTA.f), road AiThe degree of correlation between the traffic flow at k (j ═ 1,2, …, n) time intervals after the time point F and the traffic flow of the road B to be predicted at the time point F + F Δ F to be predicted is calculated by the following formula:
wherein
When C is presentiThe closer the absolute value of (f + k Δ f) is to 1, the higher the correlation.
Specific example As shown in FIG. 2, let the current time be fa，faFor 10 am, F is 5, the set of roads which can reach the road B to be predicted in 10 minutes is { a1, a2, A3, a4, a5, a6}, sampling is specified to be taken every Δ F is 2 minutes, traffic flow data of each related road every 2 minutes from 10 am to 9 am in the past 10 days is selected, wherein the traffic flow sampling data set of the a2 road on the first day is { a22 1(fa),A2 1(fa-Δf),A2 1(fa-2Δf),A2 1(fa-3Δf),A2 1(fa-4Δf),...,A2 1(fa15 deltaf), and sampling data of 10 am at 10 points in the past 10 days of the road to be predicted, wherein the sampling data of the road section B to be predicted in the first day is B1(fa+5 Δ f), a sampled data set { a2 road segment of 58 points a2 at 9 am over 10 days2 1(fa-Δf),A2 2(fa-Δf),A2 3(fa-Δf),...,A2 10(faΔ f) and a sampled data set { B } of 10 am 10 points in 10 days for a road segment B to be predicted1(fa+5Δf),B2(fa+5Δf),...,B10(fa+5 Δ f), calculating the correlation between the traffic flow of the A2 road at 58 minutes (a certain sampling time point) in the morning and the traffic flow of the road to be predicted at 10 minutes (a prediction time point) in the morning:
similarly, the correlation degree of the traffic flow of each road at each sampling time point and the traffic flow of the road to be predicted at the prediction time point is obtained by calculating a Pearson product moment correlation coefficient, the calculation results are compared, and the first h high-correlation roads, the correlation time points and information are selected as original input data of the traffic flow, wherein Ff hIndicates the traffic flow, Δ f, of the selected h-th road at the time fhTime interval, Δ e, representing a distance prediction timehARepresenting the geographical distance, K, from the road A to be predictedf hIndicating the traffic flow condition of the h-th road at the time point f, LhIndicates the length of the h-th road of interest, MhIndicating the width of the road, MARepresenting the position of the road to be predicted, the geographic distance between the h-th associated road and the road to be predicted is calculated by the following formula:
ΔehA＝Geo_Distance(MA,Mh)
(II) raw data compression based on principal component analysis
After the roads highly associated with the predicted road are selected and extracted, redundant information still exists, the training rate of the reinforced hierarchical learning network can be reduced, and the training effect is influenced.
The information of h high-correlation roads selected from the high-correlation road selection extraction constitutes an original matrix of 6 × h:
K＝[k(1)k(2) … k(6)]F∈R6×h
wherein k (i) ═ k1(i)k2(i)，...，kh(i)]∈RhThe ith (i ═ 1,2.., 6) information indicating h related roads, such as traffic flow at relevant time, is obtained by normalizing the original data in order to overcome the non-uniformity of the original data:
A＝[a(1)a(2) … a(6)]F∈R6×h
whereinIs the variance of the ith information of the jth relevant road segment,
is the average value of the ith information of the jth related road, and the covariance matrix of the road traffic flow is as follows:
wherein
All eigenvalues of the covariance matrix can be obtained by matrix calculation, and the sequence is that lambda 1 is more than or equal to lambda 2.
Because the principal component contains the most important information, only part of the principal component needs to be extracted, the information contained in all the components can be reflected, and the contribution rate of each characteristic value is as follows:
the cumulative contribution of the first r principal components is ∑ αrWhen the contribution ratio exceeds a constant ψ, it is only necessary to extract that the first r principal components represent allComponent information of (2):
in the embodiment, if there are 2 eigenvalues, each is λ1＝7.31，λ2When equal to 0.67, then λ1The contribution ratio is 7.31/(7.31+0.67) ═ 0.916 or 91.6%, the constant ψ is set to 90%, and at this time, only 1 principal component is needed to satisfy the requirement, and this principal component can be used to represent all information of all data, and finally the sampled data can be projected onto the selected r eigenvectors, thereby obtaining an r × h eigenvector for feature learning.
Second, restricted boltzmann model driving extraction traffic flow characteristic
A deep model of the limited Boltzmann machine is used for modeling a traffic network, and a limited Boltzmann machine feature extraction model is shown in figure 3, so that abstract features can be extracted from a traffic flow by using the model, and subsequent classification is facilitated, and irregular and random behaviors in the traffic flow can be better presumed and predicted.
H traffic flow characteristics of r main components extracted by a main component analysis method are used as traffic flow characteristic parameter input vectors v-v (v is v1,v2,...,vrh)FThe output is q hidden feature vectors s ═ s(s)1,s2,...,sq)F. Each traffic flow characteristic input node is only related to q hidden characteristic nodes and is independent of other characteristic input nodes, the state of each traffic flow characteristic input node is only influenced by the q hidden characteristic nodes, and each hidden characteristic node is only influenced by rh traffic flow characteristic input nodes, so that the training of the limited Boltzmann machine is easier.
The limited Boltzmann machine network comprises several parameters, namely weight M between a traffic flow characteristic recording layer and a hidden characteristic layerijAnd the second is offset vector a of traffic flow characteristic input node (a ═ a)1,a2,...,ars)FAnd thirdly a hidden featureThe offset vector b of the symbolic node is (b)1,b2,...,bq)F，MijAnd a and b determine that the limited Boltzmann machine network encodes an rh-dimension traffic flow characteristic input sample into a q-dimension hidden characteristic sample, and for a given parameter theta (M)ij,ai,bj) The limited boltzmann machine model has the following energy function value between a traffic flow characteristic input vector v and a hidden characteristic layer output vector s:
determining joint probability P (v, s) of occurrence of a group of traffic flow characteristic values of a traffic flow characteristic input node and a group of traffic flow state characteristic values of a hidden characteristic node by an energy function, wherein the joint probability between the two is as follows:
wherein
Is a normalization factor used to normalize the probability function.
In the restricted boltzmann model, the following probability value formula is given:
the probability distribution function for the traffic flow characteristic entry v is:
and (3) recording the traffic flow characteristics into an independent variable x of the function changed from v, wherein the probability distribution function about x is as follows:
an intermediate variable f (x) is defined as:
the probability distribution for x is written as:
the partial derivative function for the probability distribution of x takes negative:
from the above formula, it can be seen that when the energy value of the system is minimum, the network of the restricted boltzmann machine is stable, and to minimize the energy U, f (v) is minimized, that is, p (v) is maximized:
parameter θ ═ Mij,ai,bj) Learning by a contrast divergence algorithm to obtain:
third, road traffic flow prediction based on SVM model
The SVM is a support vector machine, has been successfully applied in the fields of text and handwriting recognition, bioinformatics and the like, has the characteristics of randomness, high dimension, nonlinearity and the like in the traffic flow prediction problem, and has a plurality of specific advantages in solving the problems of nonlinearity and high dimension mode recognition, so that the advantages can be well used for classification.
In the SVM, the entry is converted from an entry space to a feature space, then learning is carried out in the feature space, the linear SVM assumes that elements of the two spaces are in one-to-one correspondence, the entry in the entry space is mapped to a feature vector in the feature space, and the nonlinear SVM maps the entry to the feature vector by utilizing a nonlinear mapping from the entry space to the feature space.
In the invention, because the reinforced hierarchical learning method utilizes a deep network structure to simulate the nonlinear distribution in traffic flow data, after the data is eliminated redundancy through a limited Boltzmann machine, the most representative and relatively more distinguishable characteristics are obtained for further classification, and the method is more efficient than other classification methods when the linear separable Support Vector Machine (SVM) is utilized to process the distinguishable characteristics.
After the traffic flow characteristics are extracted by a limited Boltzmann machine, the traffic flow characteristics are used as the input of an SVM, and the input is mapped into a training data set on a characteristic space:
F＝{(A1,B1),(A2,B2),...,(AN,BN)}
wherein A isi＝(A11,A12,...,A1q)，i＝1,2,...,N，AiIs the ith characteristic vector formed by inputting and mapping traffic flow data, BiIs AiThe category mark of (2), namely the traffic flow state mark.
The SVM learning aims at finding an optimal separation hyperplane in a feature space formed by inputting and mapping traffic flow data and classifying feature vectors into five different traffic flow states, wherein the equation corresponding to the separation hyperplane is f (A) wFAnd a + B is determined by a normal vector w and an intercept B, when the example (a, B) is on the hyperplane, w a + B is 0, | w a + B | can represent the distance from the feature vector a to the hyperplane, and whether the classification is correct can be judged by observing whether the symbol of w a + B is consistent with the symbol of the traffic flow state mark B.
Wherein the spacing of the maximum spaced classification hyperplane is represented by the geometric spacing:
the goal of the maximum interval classifier is to find max J, which is at the corresponding constraint bar Bi(wFAi+ b) is not less than 1, i is 1,2,.. times, n, further converted to:
because of the need toIs equivalent to obtainingThe above objective function is equivalent to:
due to the special structure of the problem, the problem can be transformed to an optimization problem of dual variables through Lagrange duality, namely, the optimal solution of the original problem, namely the dual algorithm of the SVM under the linear separable condition, is obtained by solving the dual problem equivalent to the original problem.
Through linear SVM classification, the probability vector of 5 states of the traffic flow at the time stamp f can be finally obtained:
p (f) ("p (state ═ congestion"), p (state ═ slowness "), p (state ═ normality"), p (state ═ smoothness ") ]
And after SVM classification, obtaining a final road traffic flow prediction result according to the probability vector of the traffic flow at the time of the timestamp f.
Claims (7)
1. A method for predicting the time-space traffic flow driven by reinforced hierarchical learning is characterized in that: the method comprises the steps of extracting input data, driving and extracting traffic flow characteristics by a limited Boltzmann machine model, predicting road traffic flow based on an SVM model, extracting the input data, and performing high-correlation road selection extraction and raw data compression based on principal component analysis, wherein the specific steps are as follows:
firstly, selecting and extracting high-correlation roads;
secondly, compressing the original data based on principal component analysis;
thirdly, driving and extracting traffic flow characteristics by the limited boltzmann machine model;
and fourthly, predicting the road traffic flow based on the SVM model.
2. The method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: evaluating the influence of different roads on the predicted road at different time by a correlation measurement mode, selecting and selecting a Pearson product moment correlation coefficient by a highly-correlated road selection to measure the correlation between the traffic flow of each road at different time and the traffic flow of the road to be predicted at the predicted time point, wherein the more the absolute value of the Pearson product moment correlation coefficient is close to 1, the higher the correlation between the traffic flows of the two roads is, comparing the calculation results, ranking the calculation results according to the correlation degree from high to low, finally selecting the previous h highly-correlated roads and the corresponding time points, namely the correlation between the traffic flow of the relevant road at the time point and the traffic flow of the road to be predicted at the predicted time point is very high, and obtaining corresponding parameters including the distance between the road and the road to be predicted after selecting the previous h highly-correlated roads and the time points, the traffic flow and state of the road at the time point, the length and width of the road are further compressed by a principal component analysis method, the principal component analysis extracts the most main components from a series of mutually contained and redundant information to represent all the information, only the first r principal components are extracted by the principal component analysis method, the information can represent all the relevant information of the whole h relevant roads and corresponding time points, the compressed data is more effectively trained and strengthened by a hierarchical learning network and input into a limiting Boltzmann machine to obtain and record hidden layer characteristics of which the data contains the same main characteristics, the essential characteristics of the traffic flow of the road are extracted by network iteration of the limiting Boltzmann machine, finally the extracted essential traffic flow characteristics are classified by an SVM, the SVM maps the low-dimensional data to a high-layer characteristic space, and finding out the hyperplane with the optimal interval, and classifying the data to obtain the final traffic flow prediction result.
3. The method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: extracting the logging data sequentially comprises two steps of high-correlation road selection extraction and original data compression based on principal component analysis, on the aspect of logging layer data selection, firstly, according to the calculated correlation coefficient of Pearson's product moment of the correlation road and the predicted road, the front h roads with the highest correlation degree are selected and extracted, then, from the relevant information of the h roads, including average flow, time interval, geographical distance of traffic flow and traffic flow condition, road width and length of the h road at the f time point, the principal component capable of reflecting all information is extracted by using a principal component analysis method, and the extraction of the original logging data is completed.
4. The method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: firstly, in the selection and extraction of the high-correlation road, a road highly correlated with a road to be predicted and a time period are selected as input variables by using a Pearson product moment correlation coefficient as a measurement index:
wherein C represents a correlation coefficient between road flow rates, wherein AiAnd BiTraffic flow observations of an adjacent road a and a road B to be predicted respectively,and
respectively, the mean value of two variables, C describes the adjacent road A and the road to be predictedThe degree of the linear correlation between the roads B is strong and weak, the value of C is between-1 and +1, if C is greater than 0, the adjacent road A and the road B to be predicted are in positive correlation, namely the flow value of the adjacent road is larger, the flow value of the road to be predicted is correspondingly larger, if C is less than 0, the adjacent road and the road to be predicted are in negative correlation, namely the flow value of the adjacent road is larger, the flow value of the road to be predicted is smaller on the contrary, and the larger the absolute value of C is, the higher the correlation between the adjacent road and the road to be predicted is shown;
setting the current time point as F, firstly finding out a road set A which can reach a road B to be predicted in a time range F delta F, wherein delta F is a time interval, and the g-th traffic flow sampling data set of each road in a record is calculated according to e days:
{Ag(f),Ag(f-Δf),Ag(f-2Δf),…,Ag(f-nΔf)}(g＝1,2,…,e)，
wherein, the road AiThe g-th day traffic flow sample dataset in the record is:
{Ai g(f),Ai g(f-Δf),Ai g(f-2Δf),…,Ai g(f-nΔf)},
the g-th day traffic flow sampling data set of the road B to be predicted in the record is Bg(F + F. DELTA.f), road AiThe degree of correlation between the traffic flow at k (j ═ 1,2, …, n) time intervals after the time point F and the traffic flow of the road B to be predicted at the time point F + F Δ F to be predicted is calculated by the following formula:
wherein
When C is presentiThe closer the absolute value of (f + k Δ f) is to 1, the higher the correlation.
5. The method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: secondly, in the original data compression based on principal component analysis, the information of h high-correlation roads selected and extracted by the high-correlation road selection extraction forms an original matrix of 6 multiplied by h:
K＝[k(1)k(2)…k(6)]F∈R6×h
wherein k (i) ═ k1(i)k2(i)，...，kh(i)]∈RhThe ith (i ═ 1,2.., 6) information indicating the h associated roads is obtained by normalizing the raw data:
A＝[a(1)a(2)…a(6)]F∈R6×h
wherein
Is the variance of the ith information of the jth relevant road segment,
is the average value of the ith information of the jth related road, and the covariance matrix of the road traffic flow is as follows:
wherein
All characteristic values of the covariance matrix can be obtained through matrix calculation, and the sequence is that lambda 1 is more than or equal to lambda 2.
Because the principal component contains the most important information, only part of the principal component needs to be extracted, the information contained in all the components can be reflected, and the contribution rate of each characteristic value is as follows:
the cumulative contribution of the first r principal components is ∑ αrWhen the contribution ratio exceeds a constant ψ, only the first r principal components need to be extracted to represent all the component information:
6. the method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: thirdly, in the driving and extracting of the traffic flow characteristics by the limited Boltzmann machine model, a deep layer model of the limited Boltzmann machine is utilized to model a traffic network, abstract characteristics are extracted from the traffic flow, and the subsequent classification is facilitated;
h traffic flow characteristics of r main components extracted by a main component analysis method are used as traffic flow characteristic parameter input vectors v-v (v is v1,v2,...,vrh)FThe output is q hidden feature vectors s ═ s(s)1,s2,...,sq)F. Each traffic flow characteristic recording node is only related to q hidden characteristic nodes;
the limited Boltzmann machine network comprises several parameters, namely weight M between a traffic flow characteristic recording layer and a hidden characteristic layerijAnd the second is offset vector a of traffic flow characteristic input node (a ═ a)1,a2,...,ars)FAnd thirdly, the offset vector b of the hidden feature node is equal to (b)1,b2,...,bq)F，MijAnd a and b determine that the limited Boltzmann machine network encodes an rh-dimension traffic flow characteristic input sample into a q-dimension hidden characteristic sample, and for a given parameter theta (M)ij,ai,bj) Restricted boltzmann model, traffic flow profile
The energy function value between the input vector v and the hidden feature layer output vector s is:
7. the method for predicting the spatiotemporal traffic flow driven by the reinforced hierarchical learning according to claim 1, which is characterized in that: fourthly, in the road traffic flow prediction based on the SVM model, the traffic flow characteristics are extracted by a limited Boltzmann machine and then used as the input of the SVM, and the input is mapped into a training data set on a characteristic space:
F＝{(A1,B1),(A2,B2),...,(AN,BN)}
wherein A isi＝(A11,A12,...,A1q)，i＝1,2,...,N，AiIs the ith characteristic vector formed by inputting and mapping traffic flow data, BiIs AiThe class mark of (1), namely a traffic flow state mark;
the SVM learning aims at finding an optimal separation hyperplane in a feature space formed by inputting and mapping traffic flow data and classifying feature vectors into five different traffic flow states, wherein the equation corresponding to the separation hyperplane is f (A) wFA + B, determined by the normal vector w and the intercept B, when (a, B) this example is on the hyperplane, w a + B is 0, | w a + B | can represent the distance from the feature vector a to the hyperplane, and whether the classification is correct can be judged by observing whether the symbol of w a + B is consistent with the symbol of the traffic flow state mark B;
wherein the spacing of the maximum spaced classification hyperplane is represented by the geometric spacing:
the goal of the maximum interval classifier is to find max J, which is at the corresponding constraint bar Bi(wFAi+ b) is not less than 1, i is 1,2,.. times, n, further converted to:
because of the need to
Is equivalent to obtaining
The above objective function is equivalent to:
due to the special structure of the problem, the problem can be optimized by changing Lagrange duality into dual variables, namely, the optimal solution of the original problem, namely the dual algorithm of the SVM under the linear separable condition, is obtained by solving the dual problem equivalent to the original problem;
through linear SVM classification, the probability vector of 5 states of the traffic flow at the time stamp f can be finally obtained:
p (f) ("p (state ═ congestion"), p (state ═ slowness "), p (state ═ normality"), p (state ═ smoothness ") ]
And after SVM classification, obtaining a final road traffic flow prediction result according to the probability vector of the traffic flow at the time of the timestamp f.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910920367.8A CN110634294B (en) | 2019-09-26 | 2019-09-26 | Time-space traffic flow prediction method driven by reinforced hierarchical learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910920367.8A CN110634294B (en) | 2019-09-26 | 2019-09-26 | Time-space traffic flow prediction method driven by reinforced hierarchical learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110634294A true CN110634294A (en) | 2019-12-31 |
| CN110634294B CN110634294B (en) | 2021-12-24 |
Family
ID=68972951
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910920367.8A Active CN110634294B (en) | 2019-09-26 | 2019-09-26 | Time-space traffic flow prediction method driven by reinforced hierarchical learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110634294B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116665456A (en) * | 2023-07-27 | 2023-08-29 | 交通运输部科学研究院 | Method for evaluating traffic state by combining high-dimensional index dimension reduction processing |
| CN117533356A (en) * | 2024-01-09 | 2024-02-09 | 北京市北太机电设备工贸有限公司 | Intelligent driving assistance system and method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107465643A (en) * | 2016-06-02 | 2017-12-12 | 国家计算机网络与信息安全管理中心 | A kind of net flow assorted method of deep learning |
| CN108960496A (en) * | 2018-06-26 | 2018-12-07 | 浙江工业大学 | A kind of deep learning traffic flow forecasting method based on improvement learning rate |
-
2019
- 2019-09-26 CN CN201910920367.8A patent/CN110634294B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107465643A (en) * | 2016-06-02 | 2017-12-12 | 国家计算机网络与信息安全管理中心 | A kind of net flow assorted method of deep learning |
| CN108960496A (en) * | 2018-06-26 | 2018-12-07 | 浙江工业大学 | A kind of deep learning traffic flow forecasting method based on improvement learning rate |
Non-Patent Citations (3)
| Title |
|---|
| XIAOGUANG NIU 等: "DeepSense: A Novel Learning Mechanism for Traffic Prediction with Taxi GPS Traces", 《2014 IEEE GLOBAL COMMUNICATIONS CONFERENCE》 * |
| 刘艳忠: "基于Lagrange支持向量回归机交通流量预测模型的研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
| 孙占全 等: "基于主成分分析与支持向量机结合的交通流预测", 《公路交通科技》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116665456A (en) * | 2023-07-27 | 2023-08-29 | 交通运输部科学研究院 | Method for evaluating traffic state by combining high-dimensional index dimension reduction processing |
| CN116665456B (en) * | 2023-07-27 | 2023-12-19 | 交通运输部科学研究院 | Method for evaluating traffic state by combining high-dimensional index dimension reduction processing |
| CN117533356A (en) * | 2024-01-09 | 2024-02-09 | 北京市北太机电设备工贸有限公司 | Intelligent driving assistance system and method |
| CN117533356B (en) * | 2024-01-09 | 2024-03-29 | 北京市北太机电设备工贸有限公司 | Intelligent driving assistance system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110634294B (en) | 2021-12-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108629978B (en) | Traffic track prediction method based on high-dimensional road network and recurrent neural network | |
| CN111832814A (en) | Air pollutant concentration prediction method based on graph attention machine mechanism | |
| CN112784881A (en) | Network abnormal flow detection method, model and system | |
| CN112085947A (en) | Traffic jam prediction method based on deep learning and fuzzy clustering | |
| Wu et al. | A hybrid support vector regression approach for rainfall forecasting using particle swarm optimization and projection pursuit technology | |
| CN113159364A (en) | Passenger flow prediction method and system for large-scale traffic station | |
| CN104850865A (en) | Real-time compression tracking method of multi-characteristic transfer learning | |
| CN112966871B (en) | Traffic jam prediction method and system based on convolution long-short-term memory neural network | |
| CN110111578B (en) | Tensor recovery-based sporadic traffic congestion detection method | |
| CN110634294B (en) | Time-space traffic flow prediction method driven by reinforced hierarchical learning | |
| CN111861756B (en) | Group partner detection method based on financial transaction network and realization device thereof | |
| CN110674858B (en) | Traffic public opinion detection method based on space-time correlation and big data mining | |
| CN112270355A (en) | Active safety prediction method based on big data technology and SAE-GRU | |
| CN114944053A (en) | Traffic flow prediction method based on spatio-temporal hypergraph neural network | |
| CN114529038A (en) | Intelligent matching business recruitment strategy system and method based on enterprise demands | |
| CN115392554A (en) | Track passenger flow prediction method based on depth map neural network and environment fusion | |
| CN117611015B (en) | Real-time monitoring system for quality of building engineering | |
| CN115099296A (en) | Sea wave height prediction method based on deep learning algorithm | |
| CN111667135A (en) | Load structure analysis method based on typical feature extraction | |
| CN112884014A (en) | Traffic speed short-time prediction method based on road section topological structure classification | |
| CN114882069A (en) | Taxi track abnormity detection method based on LSTM network and attention mechanism | |
| CN116307103A (en) | Traffic accident prediction method based on hard parameter sharing multitask learning | |
| CN117494034A (en) | Air quality prediction method based on traffic congestion index and multi-source data fusion | |
| CN115359437A (en) | Accompanying vehicle identification method based on semantic track | |
| CN116110219A (en) | Traffic accident prediction method |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20211206 Address after: 518000 floor 11, Yantian international administrative office building, Shenzhen, Guangdong Applicant after: Yantian Port International Information Co.,Ltd. Address before: 78 No. 324000 Zhejiang province Kecheng District of Quzhou City Jiuhua North Avenue Applicant before: Liu Xiuping |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |