CN113627089B - Urban traffic flow simulation method based on mixed density neural network - Google Patents

Urban traffic flow simulation method based on mixed density neural network Download PDF

Info

Publication number
CN113627089B
CN113627089B CN202110994275.1A CN202110994275A CN113627089B CN 113627089 B CN113627089 B CN 113627089B CN 202110994275 A CN202110994275 A CN 202110994275A CN 113627089 B CN113627089 B CN 113627089B
Authority
CN
China
Prior art keywords
neural network
mixed density
density neural
mixed
matrix
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
Application number
CN202110994275.1A
Other languages
Chinese (zh)
Other versions
CN113627089A (en
Inventor
王之畅
朱越
朱浩嘉
金嘉晖
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Southeast University
Original Assignee
Southeast University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Southeast University filed Critical Southeast University
Priority to CN202110994275.1A priority Critical patent/CN113627089B/en
Publication of CN113627089A publication Critical patent/CN113627089A/en
Application granted granted Critical
Publication of CN113627089B publication Critical patent/CN113627089B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an urban traffic flow simulation method based on a mixed density neural network. The method comprises the steps of firstly collecting an urban traffic data set, and carrying out regional division on the obtained cities. And respectively setting the divided areas as a starting point (O) and an end point (D) to create a data structure corresponding to the starting point and the end point in pairs, which is called an OD matrix. The time slice length of the flow prediction is determined and corresponding slice flow data is created. And the slice flow data is used for training the mixed density neural network, if the mixed density neural network is not converged, the mixed density neural network is continuously trained, if the mixed density neural network is converged, whether the prediction effect is good is judged, and if the prediction effect is good, the background flow is generated. If the effect is poor, the time slice length is determined again, and the process is repeated until the final prediction effect is good.

Description

Urban traffic flow simulation method based on mixed density neural network
Technical Field
The invention relates to the field of intelligent city simulation systems and background traffic generation, in particular to a time sequence background traffic generation method based on a mixed density neural network.
Background
The construction of the smart city is an important issue in the process of urbanization in China, and in order to assist the construction of the smart city, researchers need to build an analog simulator of the smart city for researching the smart city. In order to reflect the real city appearance, simulation and simulation of the details of the city are needed. Urban road network traffic is used as an important simulation object of a simulator, and has a plurality of derivative purposes: a strategy for estimating congestion conditions in a city, thereby adjusting traffic lights in real time; the method is used for determining the real-time demand of taxis in a region, so that the taxis can be better dispatched; the method is used for determining the regional heat of the city, so as to assist subsequent city planning. Therefore, the method is important for the simulation and simulation of urban road network traffic.
Because traffic track data are rare and difficult to obtain, the method for simulating urban road network flow mainly depends on flow generation, and the existing scheme adopts a method for manually injecting flow, namely, the flow is manually injected at urban intersections by considering the possible flow condition of each intersection according to the information of the urban road network. And secondly, a random flow injection method is adopted, namely background flow is generated in a random mode according to a pre-designed mixed density neural network.
However, for the smart city simulator, it is required to generate not only the background traffic, but also the background traffic with accurate and large-scale features. According to the random flow injection method, the generated background flow has too strong randomness, cannot reflect the real urban traffic flow distribution, and is difficult to be used in subsequent application scenes. Due to the energy efficiency limitation of people, the large-scale simulation required by the simulator cannot be achieved by adopting a mode of manually generating background flow. Meanwhile, most of track data labels given in the data set are too thin, the data magnitude is large, and computing resources needed for directly predicting background flow are too large. And along with the prediction, the error is easy to accumulate in the prediction process and continuously deviates from the real flow distribution state. Therefore, a more effective method is to convert the prediction of the background flow into the prediction of the vehicle demand between each region, model the mixed density network with better robustness, convert the prediction of the demand value between two places into the simulation of the demand distribution, and finally sample the demand in a probability density mode. And generating background flow in the simulator according to the acquired demand between the two places.
Therefore, how to utilize a limited track data set and limited computing resources to generate real urban area vehicle demands as much as possible is very important for simulating real flow state and creating a smart city simulator.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the problems in the prior art, the invention provides an urban traffic flow simulation method based on a mixed density neural network, which emphasizes on solving the problems of generating real urban area vehicle demands as much as possible by using a limited track data set and limited computing resources and simulating a real flow state. On the premise of having historical vehicle track information, a function mixed density neural network with vehicle requirements changing along with time sequence is correctly established.
The technical scheme is as follows: in order to realize the purpose of the invention, the technical scheme adopted by the invention is as follows: a city traffic flow simulation method based on a mixed density neural network comprises the following steps:
(1) an urban traffic data set is collected. The method needs to obtain a certain amount of initial flow data set, which is a track data set S in a certain area D within a certain range of time T. Determining the latitude and longitude range of the data S, and extending D into a minimum square region D containing D and S k . According to the size of the region D and the sparsity degree of the data set S, the D is divided into k Divided into N by longitude and latitude l *N h Equal size regions, count total number of regions N, number [0 … (N-1)]；
(2) The time slice length is determined. Determining a time slice length t, t [1,5,10 … ]]The length of t can be determined according to the effect of the final experiment. Slice flow data is created. According to the starting time t of the track s Dividing the trajectory data set into N according to t as the mark time of the trajectory t T | T +1 time slices and create N t Empty OD matrices of size N x N. For each time slice, we find the corresponding track slice, obtain the initial and stopping area numbers, and add to the OD matrix of the current time number. After traversing all tracks, obtaining an OD matrix data set in a time sequence;
(3) and performing time sequence prediction according to the OD mixed density neural network. The training time slice length of each sample is determined and set to sl. The predicted slice length for each sample is determined as pl. And (3) enabling the OD matrix data sets on the time sequence to correspond to each other according to sl and pl, and generating corresponding training sets and test sets. And inputting the training set into an OD mixed density neural network according to a preset batch size to obtain an output value. Comparing the output result with the test sample to obtain a difference value between a predicted value and a true value, and performing gradient updating on the mixed density neural network;
(4) and (4) repeating the process (3) until the mixed density neural network converges, judging whether the effect of the mixed density neural network is good or not, and judging that the effect of the mixed density neural network is good if the effect of the mixed density neural network is expected on a plurality of indexes. If the index prediction is not met, repeatedly executing the steps (2) and (3) until the mixed density neural network reaches the final output expectation;
(5) and using the finally generated mixed density neural network for subsequent flow generation and generating corresponding background flow.
Further, in step (3), the mixed density neural network prediction method is as follows:
(3.1) determining the length of the training time slice of each sample, and setting the length to be sl. The predicted slice length for each sample is determined as pl. And (3) enabling the OD matrix data sets on the time sequence to correspond to each other according to sl and pl, and generating corresponding training sets and test sets. For each training set sample X t Are all shaped as (sl, N) l ，N h ) Of the three-dimensional matrix of (a). For each prediction sample Y simultaneously t All are in the shape of (pl, N) l ，N h ) A three-dimensional matrix of (a);
(3.2) the training set firstly passes through a time sequence module when passing through the module, and the time sequence module adopts the traditional GRU mixed density neural network structure to output so as to obtain an output parameter H under the time sequence module;
(3.3) representing the distribution of each OD point pair in time sequence by using a mixed density network, constructing a parameter matrix mu, sigma, alpha in the mixed density network, and taking the output H under the time sequence module as the input of the mixed density network module. Distribution of each OD point pair
Figure GDA0003693134150000031
Each Gaussian mixture density distribution phi k (t) is represented by the following formula:
Figure GDA0003693134150000032
Figure GDA0003693134150000033
Figure GDA0003693134150000034
refers to the distribution of numbers (i, j) on the OD matrix at time slice t. Wherein alpha is k (t) is the weight of the kth distribution at time slice t, phi k (t) is the distribution expression in the case of time slicing at t.
μ k (t),σ k (t) is the mean and standard deviation of the kth distribution in time slice t.
(3.4) outputting the mixed density network
Figure GDA0003693134150000035
And performing gradient updating by using a class maximum likelihood method, and checking whether convergence occurs or not.
Further, in step (3.2), the time sequence module adopts a traditional GRU mixed density neural network to construct an output, construct an update gate, reset a gate, candidate hidden states, and obtain an output parameter H under the time sequence module, wherein the mixed density neural network construction method is as follows:
(a) update gate z for mixed density neural network input t The formula of (1) is as follows:
z t ＝sigmoid(x t W z +H t-1 U z )
wherein sigmoid is an activation function, x t For slicing input data, W z As a weight matrix, H t-1 For the last hidden state, U z Is a weight matrix.
(b) Reset gate r of mixed density neural network input t Is as follows
r t ＝sigmoid(x t W r +H t-1 U r )
Wherein sigmoid is an activation function, x t For slicing input data, W r As a weight matrix, U r Is a weight matrix.
(c) Candidate hidden states for mixed density neural networks
Figure GDA0003693134150000036
The formula is as follows:
Figure GDA0003693134150000037
wherein
Figure GDA0003693134150000038
In order to be a weight matrix, the weight matrix,
Figure GDA0003693134150000039
is a weight matrix.
(d) Hidden state H of mixed density neural network t The formula is as follows:
Figure GDA0003693134150000041
(e) the final output formula of the mixed density neural network is as follows:
H d ＝sigmoid(H t W d )
further, in step (2.3), we use a mixed density network to represent the distribution of each OD point pair in time sequence, construct a parameter matrix μ, σ, α in the mixed density network, and use the output H under the time sequence module as the input of the mixed density network module, wherein the mixed density network construction method is as follows:
(A) the parameter μ represents the mean of the constructed gaussian mixture density network, which is constructed as follows:
μ＝sigmoid(H d W μ +b μ )
(B) the parameter σ represents the standard deviation of a constructed gaussian mixture density network, constructed as follows:
σ＝sigmoid(H d W σ +b σ )
(C) the parameter α represents the weight of the constructed gaussian mixture density network in the final overall distribution expression, and the construction mode is shown as the following formula:
α＝sigmoid(H d W α +b α )
further, a distribution expression of the mixed density neural network is generated after step (C):
Figure GDA0003693134150000042
Figure GDA0003693134150000043
since in a real test set sample we can know the real value of the sample, and the result we output is a distribution that we should let as much as possible fit the real value situation. We therefore use maximum likelihood to represent the matching state of the distribution and set our penalty function accordingly. The loss function we set is shown as follows:
Figure GDA0003693134150000044
from this loss function we can further perform a gradient update parameter, the magnitude of the gradient update being given by the set learning rate.
Further, in step (3), evaluation of the mixed density neural network index is required. And integrating the finally calculated distribution, thereby assuming a predicted value, judging whether the effect of the mixed density neural network is good or not, and judging that the mixed density neural network is good if the effect of the mixed density neural network is expected on a plurality of indexes. And (3) judging indexes of the mixed density neural network are MSE and RMSE, judging whether the mixed density neural network belongs to a good mixed density neural network according to experience, and determining whether the processes of the steps (2) and (3) are repeated.
The invention has the beneficial effects that: the method emphasizes the problem that the real urban area vehicle demand is generated as much as possible by using a limited track data set and limited computing resources, and the real flow state is simulated; on the premise of having historical vehicle track information, a function mixed density neural network with vehicle requirements changing along with time sequence is correctly established.
Drawings
FIG. 1 is a flow chart of an algorithm of an urban traffic flow simulation method based on a mixed density neural network.
Detailed Description
The following describes a specific embodiment of the present invention with reference to the drawings, and fig. 1 is a flowchart of an algorithm of a mixed density neural network-based urban traffic flow simulation method implemented by the present invention. In light of the drawings, it will be understood that the invention may be practiced as follows:
(1) an urban traffic data set is collected. The method needs to obtain a certain amount of initial flow data set, which is a track data set S in a certain area D within a certain range of time T. Determining the latitude and longitude range of the data S, and extending D into a minimum square region D containing D and S k . According to the size of the region D and the sparsity degree of the data set S, the D is divided into k Divided into N by longitude and latitude l *N h Equal size regions, count total number of regions N, number [0 … (N-1)]；
(2) The time slice length is determined. Determining a time slice length t, t [1,5,10 … ]]The length of t can be determined according to the effect of the final experiment. Slice flow data is created. According to the starting time t of the track s Dividing the trajectory data set into N according to t as the mark time of the trajectory t T | T +1 time slices and create N t Empty OD matrices of size N x N. For each time slice, we find the corresponding track slice, obtain the initial and stopping area numbers, and add to the OD matrix of the current time number. After traversing all the tracks, obtaining an OD matrix data set on a time sequence;
(3) and performing time sequence prediction according to the OD mixed density neural network. The training time slice length of each sample is determined and set to sl. The predicted slice length for each sample is determined, set to pl. And correspondingly generating a corresponding training set and a corresponding test set by the OD matrix data set on the time sequence according to sl and pl. And inputting the training set into an OD mixed density neural network according to a preset batch size to obtain an output value. Comparing the output result with the test sample to obtain a difference value between a predicted value and a true value, and performing gradient updating on the mixed density neural network;
(4) and (4) repeating the process (3) until the mixed density neural network is converged, judging whether the mixed density neural network is good or bad, and judging that the mixed density neural network is good if the mixed density neural network is expected on a plurality of indexes. If the index prediction is not met, repeatedly executing the steps (2) and (3) until the mixed density neural network reaches the final output expectation;
(5) and using the finally generated mixed density neural network for subsequent flow generation and generating corresponding background flow.
In this embodiment, in step (3), the mixed density neural network prediction method is as follows:
(3.1) determining the training time slice length of each sample, and setting the training time slice length as sl. The predicted slice length for each sample is determined, set to pl. And correspondingly generating a corresponding training set and a corresponding test set by the OD matrix data set on the time sequence according to sl and pl. For each training set sample X t Are all shaped as (sl, N) l ，N h ) Of the three-dimensional matrix of (a). For each prediction sample Y simultaneously t Are all in the shape of (pl, N) l ，N h ) A three-dimensional matrix of (a);
(3.2) the training set firstly passes through a time sequence module when passing through the module, and the time sequence module adopts the traditional GRU mixed density neural network structure to output so as to obtain an output parameter H under the time sequence module;
(3.3) we represent the distribution of each pair of OD points in time sequence by means of a mixed density network, in which a parameter matrix mu, sigma, alpha is constructed, and the output H under the time sequence module is used as the input of the mixed density network module. Distribution of each OD point pair
Figure GDA0003693134150000061
Each Gaussian mixture density distribution phi k (t) is represented by the following formula:
Figure GDA0003693134150000062
Figure GDA0003693134150000063
Figure GDA0003693134150000064
refers to the distribution of numbers (i, j) on the OD matrix at time slice t. Wherein alpha is k (t) is the weight of the kth distribution at time slice t, phi k (t) is the distribution expression in the case of time slicing at t. Mu.s k (t),σ k (t) is the mean and standard deviation of the kth distribution in time slice t.
(3.4) outputting the mixed density network
Figure GDA0003693134150000065
And performing gradient updating by using a class maximum likelihood method, and checking whether convergence occurs or not.
In this embodiment, in step (3.2), the timing module adopts a conventional GRU mixed density neural network to construct an output, construct an update gate, reset a gate, candidate hidden states, and obtain an output parameter H under the timing module, where the mixed density neural network construction method is as follows:
(a) update gate z for mixed density neural network input t The formula of (1) is as follows:
z t ＝sigmoid(x t W z +H t-1 U z )
wherein sigmoid is an activation function, x t For slicing input data, W z As a weight matrix, H t-1
For the last hidden state, U z Is a weight matrix.
(b) Reset gate r of mixed density neural network input t Is as follows
r t ＝sigmoid(x t W r +H t-1 U r )
Wherein sigmoid is an activation function, x t For slicing input data, W r As a weight matrix, U r Is a weight matrix.
(c) MixingCandidate hidden states for a dense neural network
Figure GDA0003693134150000071
The formula is as follows:
Figure GDA0003693134150000072
wherein
Figure GDA0003693134150000073
In order to be a weight matrix, the weight matrix,
Figure GDA0003693134150000074
is a weight matrix.
(d) Hidden state H of mixed density neural network t The formula is as follows:
Figure GDA0003693134150000075
(e) the final output formula of the mixed density neural network is as follows:
H d ＝sigmoid(H t W d )
in this embodiment, in step (2.3), we use a mixed density network to represent the distribution of each OD point pair in time sequence, construct a parameter matrix μ, σ, α in the mixed density network, and use the output H under the time sequence module as the input of the mixed density network module, where the mixed density network construction method is as follows:
(A) the parameter μ represents the mean of a constructed gaussian mixture density network, constructed in the following way:
μ＝sigmoid(H d W μ +b μ )
(B) the parameter σ represents the standard deviation of a constructed gaussian mixture density network, constructed as follows:
σ＝sigmoid(H d W σ +b σ )
(C) the parameter α represents the weight of the constructed gaussian mixture density network in the final overall distribution expression, and the construction mode is shown as the following formula:
α＝sigmoid(H d W α +b α )
in this embodiment, a distribution expression of the mixed density neural network is generated after step (C):
Figure GDA0003693134150000076
Figure GDA0003693134150000077
since in a real test set sample we can know the real value of the sample, and the result we output is a distribution that we should let as much as possible fit the real value situation. We therefore use maximum likelihood to represent the matching state of the distribution and set our penalty function accordingly. The loss function we set is shown as follows:
Figure GDA0003693134150000081
from this loss function we can further perform a gradient update parameter, the magnitude of the gradient update being given by the set learning rate.
In this embodiment, step (3) requires evaluation of the mixed density neural network index. And integrating the finally calculated distribution, thereby assuming a predicted value, judging whether the effect of the mixed density neural network is good or not, and judging that the mixed density neural network is good if the effect of the mixed density neural network is expected on a plurality of indexes. And (4) judging indexes of the mixed density neural network are MSE and RMSE, judging whether the mixed density neural network belongs to a good mixed density neural network according to experience, and determining whether the processes of the steps (2) and (3) are repeated.

Claims (6)

1. A city traffic flow simulation method based on a mixed density neural network is characterized by comprising the following steps:
(1) collecting an urban traffic data set; obtaining a certain amount of initial flow data set which is a track data set S in a certain area D within a certain range of time T; determining the latitude and longitude range of the data set S, and extending the region D into a minimum square region D containing the region D and the data set S k (ii) a According to the size of the region D and the sparsity degree of the data set S, the region D is divided into k Divided into N by longitude and latitude l *N h The area of the size is recorded as the total number of the areas is N, and the number is 0 (N-1);
(2) determining a time slice length; determining the time slice length t, wherein t is 1,5,10, determining the length of the time slice length t according to the effect of the final experiment, creating slice flow data, and obtaining the start time t of the track s Dividing the trajectory data set into N according to the time slice length t as the mark time of the trajectory t T | T +1 time slices and create N t Empty OD matrices of size N × N; for each time slice, finding the corresponding track slice, obtaining the initial and stopping area numbers, adding the initial and stopping area numbers to the OD matrix of the current time number, and obtaining the OD matrix data set in the time sequence after traversing all tracks;
(3) performing time sequence prediction according to the OD mixed density neural network; determining the length of a training time slice of each sample, and setting the length as sl; determining the prediction time slice length of each sample, and setting the prediction time slice length as pl; enabling an OD matrix data set on a time sequence to correspond to an sl matrix data set and a pl matrix data set, and generating a corresponding training set and a corresponding testing set; inputting the training set into an OD mixed density neural network according to a preset batch size to obtain an output value; comparing the output result with the test sample through sampling to obtain a difference value between a predicted value and a true value, and performing gradient updating on the mixed density neural network;
(4) repeating the step (3) until the mixed density neural network is converged, judging whether the effect of the mixed density neural network is good or not, and judging that the effect of the mixed density neural network is good if the effect of the mixed density neural network is expected on a plurality of indexes; if the index prediction is not met, repeatedly executing the steps (2) and (3) until the OD mixed density neural network reaches the final output expectation;
(5) and using the finally generated OD mixed density neural network for subsequent traffic generation and generating corresponding background traffic.
2. The urban traffic flow simulation method based on the mixed density neural network as claimed in claim 1, wherein in step (3), the mixed density neural network prediction method is as follows:
(3.1) determining the length of a training time slice of each sample, and setting the length as sl; determining the prediction time slice length of each sample, and setting the prediction time slice length as pl; enabling an OD matrix data set on a time sequence to correspond to an sl matrix data set and a pl matrix data set, and generating a corresponding training set and a corresponding testing set; for each training set sample X t Are all shaped as (sl, N) l ，N h ) Of each prediction sample Y, while for each prediction sample Y t Are all in the shape of (pl, N) l ，N h ) A three-dimensional matrix of (a);
(3.2) the training set firstly passes through a time sequence module when passing through the module, and the time sequence module adopts the traditional GRU mixed density neural network structure to output so as to obtain an output parameter H under the time sequence module;
(3.3) representing the distribution of each OD point pair in time sequence by using a mixed density network, constructing a parameter matrix mu, sigma, alpha in the mixed density network, and taking the output H under a time sequence module as the input of a mixed density network module, wherein the distribution of each OD point pair
Figure FDA0003693134140000021
Each Gaussian mixture density distribution phi k (t) is represented by the following formula:
Figure FDA0003693134140000022
Figure FDA0003693134140000023
Figure FDA0003693134140000024
refers to the distribution of numbers (i, j) on the OD matrix at time slice t, where α k (t) is the weight of the kth distribution at time slice t, phi k (t) is the distribution expression in the case of time slicing at t, μ k (t),σ k (t) is the mean and standard deviation of the kth distribution in time slice t;
(3.4) outputting the mixed density network
Figure FDA0003693134140000025
And performing gradient updating by using a class maximum likelihood method, and checking whether convergence occurs or not.
3. The urban traffic flow simulation method based on the mixed density neural network as claimed in claim 2, wherein in step (3.2), the timing module adopts a traditional GRU mixed density neural network to construct an output, construct an update gate, reset a gate, candidate hidden states, and obtain an output parameter H under the timing module, and the mixed density neural network construction method is as follows:
(a) update gate z for mixed density neural network input t The formula of (1) is as follows:
z t ＝sigmoid(x t W z +H t-1 U z )
wherein sigmoid is an activation function, x t For slicing input data, W z As a weight matrix, H t-1 For the last hidden state, U z Is a weight matrix;
(b) reset gate r of mixed density neural network input t Is as follows
r t ＝sigmoid(x t W r +H t-1 U r )
Wherein sigmoid is an activation function, x t For slicing input data, W r As a weight matrix, U r Is a weightA matrix;
(c) candidate hidden states for mixed density neural networks
Figure FDA0003693134140000026
The formula is as follows:
Figure FDA0003693134140000027
wherein
Figure FDA0003693134140000028
In order to be a weight matrix, the weight matrix,
Figure FDA0003693134140000029
is a weight matrix;
(d) hidden state H of mixed density neural network t The formula is as follows:
Figure FDA0003693134140000031
(e) the final output formula of the mixed density neural network is as follows:
H d ＝sigmoid(H t W d )。
4. the urban traffic simulation method based on the mixed density neural network according to the claim 2 or 3, wherein in the step (2.3), the distribution of each OD point pair in time sequence is represented by using a mixed density network, a parameter matrix μ, σ, α is constructed in the mixed density network, and the output H under a time sequence module is used as the input of a mixed density network module, wherein the mixed density network construction method is as follows:
(A) the parameter μ represents the mean value of the constructed gaussian mixture density network, which is constructed in the following way:
μ＝sigmoid(H d W μ +b μ )
(B) the parameter σ represents the standard deviation of the constructed gaussian mixture density network, which is constructed in the following way:
σ＝sigmoid(H d W σ +b σ )
(C) the parameter α represents the weight of the constructed gaussian mixture density network in the final overall distribution expression, and the construction mode is shown as the following formula:
α＝sigmoid(H d W α +b α )。
5. the method according to claim 4, wherein after step (C) a distribution expression of the mixed density neural network is generated:
Figure FDA0003693134140000032
Figure FDA0003693134140000033
since in a real test set sample, the real value of the sample can be known, and the output result is a distribution, the distribution should be fitted to the real value situation, so that the maximum likelihood manner is adopted to represent the matching state of the distribution, and thus a loss function is set, wherein the set loss function is shown in the following formula:
Figure FDA0003693134140000034
according to the loss function, gradient update parameters are performed, and the magnitude of the gradient update is given by the set learning rate.
6. The urban traffic flow simulation method based on the mixed density neural network according to claim 1, wherein in the step (3), the mixed density neural network index needs to be evaluated, the finally calculated distribution is integrated, so that a predicted value is assumed, the mixed density neural network effect is judged to be good or bad, and if the mixed density neural network effect is expected to be good on a plurality of indexes, the mixed density neural network is judged to be good; and (3) judging whether the mixed density neural network belongs to a good mixed density neural network according to experience, and determining whether the processes of the steps (2) and (3) are repeated.
CN202110994275.1A 2021-08-27 2021-08-27 Urban traffic flow simulation method based on mixed density neural network Active CN113627089B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110994275.1A CN113627089B (en) 2021-08-27 2021-08-27 Urban traffic flow simulation method based on mixed density neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110994275.1A CN113627089B (en) 2021-08-27 2021-08-27 Urban traffic flow simulation method based on mixed density neural network

Publications (2)

Publication Number Publication Date
CN113627089A CN113627089A (en) 2021-11-09
CN113627089B true CN113627089B (en) 2022-08-02

Family

ID=78388115

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110994275.1A Active CN113627089B (en) 2021-08-27 2021-08-27 Urban traffic flow simulation method based on mixed density neural network

Country Status (1)

Country Link
CN (1) CN113627089B (en)

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109492814B (en) * 2018-11-15 2021-04-20 中国科学院深圳先进技术研究院 Urban traffic flow prediction method, system and electronic equipment
CN109859469A (en) * 2019-02-15 2019-06-07 重庆邮电大学 A kind of vehicle flowrate prediction technique based on integrated LSTM neural network
CN110111575B (en) * 2019-05-16 2020-10-27 北京航空航天大学 Urban traffic flow network analysis method based on complex network theory
CN111292525B (en) * 2019-12-30 2023-01-13 南京艾特斯科技有限公司 Traffic flow prediction method based on neural network
CN112508173A (en) * 2020-12-02 2021-03-16 中南大学 Traffic space-time sequence multi-step prediction method, system and storage medium
CN112906948B (en) * 2021-02-02 2023-12-22 湖南大学 Urban area attraction prediction method, device and medium based on private car track big data
CN112966853B (en) * 2021-02-04 2023-12-22 南通大学 Urban road network short-time traffic flow prediction method based on space-time residual mixed model

Also Published As

Publication number Publication date
CN113627089A (en) 2021-11-09

Similar Documents

Publication Publication Date Title
CN109142171B (en) Urban PM10 concentration prediction method based on feature expansion and fusing with neural network
CN110097755A (en) Freeway traffic flow amount state identification method based on deep neural network
CN106781489B (en) A kind of road network trend prediction method based on recurrent neural network
CN110021165A (en) A kind of traffic flow forecasting method based on Autoencoder-LSTM Fusion Model
CN111292525B (en) Traffic flow prediction method based on neural network
CN105096614B (en) Newly-built crossing traffic flow Forecasting Methodology based on generation moldeed depth belief network
CN112489426B (en) Urban traffic flow space-time prediction scheme based on graph convolution neural network
CN109754113A (en) Load forecasting method based on dynamic time warping Yu length time memory
CN110517482B (en) Short-term traffic flow prediction method based on 3D convolutional neural network
CN109840660A (en) A kind of vehicular characteristics data processing method and vehicle risk prediction model training method
CN110164129B (en) Single-intersection multi-lane traffic flow prediction method based on GERNN
CN110148296A (en) A kind of trans-city magnitude of traffic flow unified prediction based on depth migration study
CN110909909A (en) Short-term traffic flow prediction method based on deep learning and multi-layer spatiotemporal feature map
CN112966853B (en) Urban road network short-time traffic flow prediction method based on space-time residual mixed model
CN112907970B (en) Variable lane steering control method based on vehicle queuing length change rate
CN110826244B (en) Conjugated gradient cellular automaton method for simulating influence of rail transit on urban growth
CN110517510A (en) Based on the Intersections timing designing method for improving Webster function and genetic algorithm
CN115862324A (en) Space-time synchronization graph convolution neural network for intelligent traffic and traffic prediction method
CN109214565A (en) A kind of subregion system loading prediction technique suitable for the scheduling of bulk power grid subregion
CN114780739A (en) Time sequence knowledge graph completion method and system based on time graph convolution network
CN112633602A (en) Traffic congestion index prediction method and device based on GIS map information
Zhao et al. Traffic signal control with deep reinforcement learning
CN106297296A (en) A kind of fine granularity distribution method hourage based on sparse tracing point data
CN112098869B (en) Self-adaptive electric vehicle SOC estimation method based on big data
CN113627089B (en) Urban traffic flow simulation method based on mixed density neural network

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