CN114266406A - Method for predicting traffic flow state of large-scale road network based on federal learning - Google Patents

Method for predicting traffic flow state of large-scale road network based on federal learning Download PDF

Info

Publication number
CN114266406A
CN114266406A CN202111601256.4A CN202111601256A CN114266406A CN 114266406 A CN114266406 A CN 114266406A CN 202111601256 A CN202111601256 A CN 202111601256A CN 114266406 A CN114266406 A CN 114266406A
Authority
CN
China
Prior art keywords
network
road network
traffic flow
training
federal learning
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.)
Pending
Application number
CN202111601256.4A
Other languages
Chinese (zh)
Inventor
于海洋
梁育豪
任毅龙
赵亚楠
兰征兴
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.)
Beihang University
Original Assignee
Beihang 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 Beihang University filed Critical Beihang University
Priority to CN202111601256.4A priority Critical patent/CN114266406A/en
Publication of CN114266406A publication Critical patent/CN114266406A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Traffic Control Systems (AREA)

Abstract

The present disclosure relates to a large-scale road network traffic flow state prediction scheme based on federal learning, which is characterized by comprising: step one, constructing a directed graph; step two, establishing an initial model; step three, updating the training parameters by using a back propagation algorithm; and step four, obtaining a prediction result by using a federal average algorithm. Based on the method, a large-scale road network is decomposed into a plurality of sub-networks, a plurality of base stations in each sub-network collect traffic flow characteristics of vehicles in a certain range within a period of time, each base station serves as a participant in federal learning, receives global models respectively, trains current sub-network traffic flow prediction models by locally using own data sets, uploads the models to a server for global aggregation, and predicts future states of the road network by the server. The operation cost of the server can be effectively reduced, the training efficiency of the model can be higher, and the prediction effect is better.

Description

Method for predicting traffic flow state of large-scale road network based on federal learning
Technical Field
The invention belongs to the field of road network traffic flow prediction, federal learning and intelligent traffic systems, and particularly relates to a technology for predicting traffic flow states of a large-range road network segmentation sub-road network by using a distributed machine learning framework of federal learning.
Background
The quantity of motor vehicles in China rapidly increases year by year, a series of road network resources such as traffic jam, frequent traffic accidents and the like and the problem of contradiction between supply and demand among motor vehicles are generated, and great inconvenience and trouble are brought to travelers and traffic management departments. The intelligent traffic system is a comprehensive transportation system which effectively and comprehensively applies new-generation scientific technologies such as computer technology, data communication technology and the like to the traffic fields such as traffic transportation, service control and the like, road network traffic flow prediction is an important research direction in the intelligent traffic system, the traffic flow prediction can predict the situation of traffic network evolution in a future period of time of road conditions, accurate travel information is provided for travelers, a basis is provided for traffic managers to actively control traffic, the road network efficiency can be effectively improved, the traffic safety is guaranteed, and meanwhile, the environment is improved and energy is saved.
Federal learning is a distributed machine learning technique proposed by google in 2016. In particular, people train algorithms on multiple decentralized edge devices or servers that own local data samples. The method is obviously different from the traditional centralized machine learning technology, the traditional centralized machine learning technology uploads all local data sets to one server, and the federal learning is that the results are transmitted to the server after the data are trained locally, so that the direct disclosure of personal data is avoided from the source, and the privacy safety of users is protected. Meanwhile, the operation pressure of the central server can be relieved without concentrating the data on the server for operation, so that the model training efficiency is higher.
Conventional traffic flow prediction methods may be classified into parametric models and non-parametric models. The parametric model mainly includes an autoregressive sum moving average model (ARIMA) and a kalman filtering method, and the non-parametric model mainly includes a KNN model and a support vector machine method. With the prosperity of the machine learning method in the new era, a plurality of scholars apply the machine learning method to traffic flow prediction, and the method mainly comprises a convolutional neural network, a generation countermeasure network, a tensor neural network and the like. However, the prediction of the road network traffic flow by the current method is often directed to the road network in a small range, and the load of the server is too large in the case of a large amount of data, so that the prediction time is prolonged, and the prediction of the road network state in a large range is not facilitated.
Disclosure of Invention
The invention aims to solve the technical problem of designing a large-scale road network traffic flow state prediction scheme based on federal learning.
The invention adopts the technical scheme for solving the requirements as follows:
the method comprises the following steps: a large-scale road network is divided into several subnets.
Step two: the server constructs a global prediction model by using a gating cycle unit and a fully connected neural network, generates initial parameters, and distributes the initial global model to the base stations participating in the federal learning in each subnet.
Step three: and the base station in each subnet uses the traffic flow characteristic time sequence data acquired by the base station in a certain time period to train the global model for a plurality of rounds, and uploads the trained local model parameters to the server after the training is finished.
Step four: and the server aggregates the models by receiving the uploaded local model parameters by using a federal learning average aggregation algorithm to generate a new global model, and predicts the traffic flow state at a plurality of moments in the future by using the new global model.
Specifically, the method includes:
a large-scale road network traffic flow state prediction scheme based on federal learning comprises the following steps:
step one, constructing a directed graph
And (C) simulating the large-range road network into a directed graph G (V, E), wherein V is a point set, intersections are simulated into vertexes in the directed graph, E is an edge set, and road sections between the two intersections are simulated into directed edges. Dividing a road network into n disjoint directed sub-graphs according to actual physical characteristics of the traffic road network
Figure BDA0003431884950000031
Step two, establishing an initial model
For each sub-network
Figure BDA0003431884950000032
The number of segments obtained by dividing the time segments according to the time interval delta T is marked as T, each base station collects GPS information transmitted by vehicles, and data are gathered and expanded into a matrix
Figure BDA0003431884950000033
N represents the number of time series; the building gate is formed by connecting T GRUs in series to control a circulation unit htOutput for the t GRU unit; h is to beTAs input of fully connected neural network, input into the network for training to obtain predicted result
Figure BDA0003431884950000036
Step three, updating the training parameters by using a back propagation algorithm
Updating and training parameters of the fully-connected neural network and parameters of the gating cycle unit by using a back propagation algorithm; wherein the loss function of the back propagation algorithm
Figure BDA0003431884950000034
Wherein v is the true value in the training set;
obtaining a trained parameter set through a plurality of times of forward propagation and backward propagation
Figure BDA0003431884950000035
Parameter set
Figure BDA0003431884950000041
Sending the data to a server;
step four, obtaining a prediction result by using a federal average algorithm
For each subnet GqThe global model is updated using the federal averaging algorithm:
Figure BDA0003431884950000042
wherein, | BSqI is the sum of the number of all base stations participating in federal learning of the current subnet, and i is a base station label; each subnetwork GqAfter the global model of (2) is updated, W is usedqConstructing a global prediction model by the determined gating cycle units and the fully-connected neural network; inputting the stored historical traffic flow data omega into the model to obtain a prediction result
Figure BDA0003431884950000043
Preferably, in the second step, in the gating cycle unit, Wr，Wz，W，Ur，UzU is a weight parameter matrix to be trained, htFor the t GRU output, xtIs a column vector of matrix X; the forward propagation formula is: r ist＝σ(Wrxt+Urht-1)；zt＝σ(Wzxt+Uzht-1)；
Figure BDA0003431884950000044
Figure BDA0003431884950000045
Wherein the function is:
Figure BDA0003431884950000046
Figure BDA0003431884950000047
output of gated cyclic unit
Figure BDA0003431884950000048
H is to beTAs the input of the fully connected neural network, inputting the input into the network for training; w(i)Is the weight matrix of the i-th layer of the fully-connected neural network, b(i)Is the bias of the i-th layer, z(i)Is the output of the i-th layer, a(i-1)Is an input to the ith layer; the objective function is defined as v ═ WTy + b, then the formula for the i-th layer forward propagation is: z is a radical of(i)＝W(i)a(i-1)+b(i)；a(i)＝σ(z(i)) And a is a(0)＝hT(ii) a When the number of hidden layers l is 1, the predicted result
Figure BDA0003431884950000049
Preferably, in said third step, the parameter W for the gated-cycle cellr，Wz，W，Ur，UzU is trained using back propagation, the formula is as follows:
Figure BDA00034318849500000410
Figure BDA00034318849500000411
Figure BDA0003431884950000051
wherein:
Figure BDA0003431884950000052
Figure BDA0003431884950000053
Figure BDA0003431884950000054
after a plurality of times of forward propagation and backward propagation, a group of trained parameter sets can be obtained
Figure BDA0003431884950000055
Figure BDA0003431884950000056
According to the technical scheme, the method for predicting the traffic flow state of the large-scale road network based on the federal learning comprises the steps that the large-scale road network is decomposed into a plurality of sub-networks, a plurality of base stations in each sub-network collect traffic flow characteristics of vehicles in a certain range within a period of time, each base station serves as a participant in the federal learning and is enabled to receive a global model respectively, the current sub-network traffic flow prediction model is trained by locally using a data set of the base station, then the current sub-network traffic flow prediction model is uploaded to a server to be subjected to global aggregation, and the future state of the road network is predicted by the server. The method can effectively reduce the operation cost of the server, and can also improve the training efficiency and the prediction effect of the model.
Drawings
FIG. 1 is a schematic flow chart of the present invention.
Detailed Description
The following describes in detail specific embodiments of the present invention.
The method comprises the following steps: and (C) simulating the large-range road network into a directed graph G (V, E), wherein V is a point set, intersections are simulated into vertexes in the directed graph, E is an edge set, and road sections between the two intersections are simulated into directed edges. Dividing a road network into a plurality of disjoint directed sub-graphs according to actual physical characteristics of the traffic road network
Figure BDA0003431884950000061
Namely:
Figure BDA0003431884950000062
step two: establishing an initial model:
for each sub-network
Figure BDA0003431884950000063
Determining a time interval delta T, determining a time period (supposing 1 hour), wherein the number of the segments obtained by dividing the time period according to the time interval delta T is marked as T, and each base station has N time sequence sequences by collecting GPS information transmitted by vehicles. The data set for each base station is spanned into a matrix
Figure BDA0003431884950000064
A gated cycle unit (GRU) is established. Is formed by connecting T GRUs in series, wherein Wr，Wz，W，Ur，UzU is a weight parameter matrix to be trained, htFor the t GRU output, xtIs the column vector of matrix X. The forward propagation formula is:
rt＝σ(Wrxt+Urht-1)
zt＝σ(Wzxt+Uzht-1)
Figure BDA0003431884950000065
Figure BDA0003431884950000066
wherein the function is:
Figure BDA0003431884950000067
Figure BDA0003431884950000068
output of gated cyclic unit
Figure BDA0003431884950000069
H is to beTThe input of the fully connected neural network is input into the network for training. W(i)Is the weight matrix of the i-th layer of the fully-connected neural network, b(i)Is the bias of the i-th layer, z(i)Is the output of the i-th layer, a(i-1)Is an input to the ith layer. Let the objective function be v ═ WTy + b, then the formula for the i-th layer forward propagation is:
z(i)＝W(i)a(i-1)+b(i)
a(i)＝σ(z(i)) And a is a(0)＝hT
Assuming that the number of hidden layers is 1, the final prediction result is obtained
Figure BDA0003431884950000071
Step three: each subnet GqMiddle base station
Figure BDA0003431884950000072
Extracting the average speed of the road sections divided by time in a period of time from the database, and opening the data set into a matrix
Figure BDA0003431884950000073
Each base station uses the data set
Figure BDA0003431884950000074
Freely segmenting the training set and the test set, for the training set
Figure BDA0003431884950000075
Using the algorithm as step 2 to forward propagate and obtain the result
Figure BDA0003431884950000076
v is the true value in the training set. Defining a loss function
Figure BDA0003431884950000077
The training parameters are then updated using a back propagation algorithm. For parameters of the fully-connected neural network, the formula for the i-th layer back propagation is as follows:
Figure BDA0003431884950000078
Figure BDA0003431884950000079
where α represents the learning rate.
Parameter W for gated cycle cellr，Wz，W，Ur，UzU is trained using back propagation, the formula is as follows:
Figure BDA00034318849500000710
Figure BDA00034318849500000711
Figure BDA0003431884950000081
Figure BDA0003431884950000082
Figure BDA0003431884950000083
Figure BDA0003431884950000084
wherein:
Figure BDA0003431884950000085
Figure BDA0003431884950000086
Figure BDA0003431884950000087
Figure BDA0003431884950000088
Figure BDA0003431884950000089
Figure BDA00034318849500000810
Figure BDA00034318849500000811
Figure BDA00034318849500000812
through a plurality of forward propagation and backward propagation, a group of trained parameter sets can be obtained:
Figure BDA00034318849500000813
and sending the parameter list to the server.
Step four: server for each subnet Gq，|BSqAnd | is the sum of all base stations participating in federal learning of the current subnet, and a global model is updated by using a federal mean (FedAVG) algorithm:
Figure BDA0003431884950000091
each subnetwork GqAfter the global model of (2) is updated, W is usedqAnd constructing a global prediction model by the determined gating cycle units and the fully-connected neural network. The server uses the stored historical traffic flow data omega to input the historical traffic flow data omega into the model to obtain a prediction result
Figure BDA0003431884950000092

Claims (3)

1. A method for predicting traffic flow states of a large-scale road network based on federal learning is characterized by comprising the following steps:
step one, constructing a directed graph
And (C) simulating the large-range road network into a directed graph G (V, E), wherein V is a point set, intersections are simulated into vertexes in the directed graph, E is an edge set, and road sections between the two intersections are simulated into directed edges. Dividing a road network into n disjoint directed sub-graphs according to actual physical characteristics of the traffic road network
Figure FDA0003431884940000011
Step two, establishing an initial model
For each sub-network
Figure FDA0003431884940000012
The number of segments obtained by dividing the time segments according to the time interval delta T is marked as T, each base station collects GPS information transmitted by vehicles, and data are gathered and expanded into a matrix
Figure FDA0003431884940000013
N represents the number of time series; the building gate is formed by connecting T GRUs in series to control a circulation unit htOutput for the t GRU unit; h is to beTAs input of fully connected neural network, input into the network for training to obtain predicted result
Figure FDA0003431884940000014
Step three, updating the training parameters by using a back propagation algorithm
Updating and training parameters of the fully-connected neural network and parameters of the gating cycle unit by using a back propagation algorithm; wherein the loss function of the back propagation algorithm
Figure FDA0003431884940000015
Wherein v is the true value in the training set;
obtaining a trained parameter set through a plurality of times of forward propagation and backward propagation
Figure FDA0003431884940000016
Parameter set
Figure FDA0003431884940000017
Sending the data to a server;
step four, obtaining a prediction result by using a federal average algorithm
For each subnet GqThe global model is updated using the federal averaging algorithm:
Figure FDA0003431884940000021
wherein, | BSqI is the sum of the number of all base stations participating in federal learning of the current subnet, and i is a base station label; each subnetwork GqIs updated and then used
Figure FDA0003431884940000029
Constructing a global prediction model by the determined gating cycle units and the fully-connected neural network; inputting the stored historical traffic flow data omega into the model to obtain a prediction result
Figure FDA0003431884940000022
2. The method according to claim 1, wherein in step two,
in the gated cyclic unit, Wr，Wz，W，Ur，UzU is a weight parameter matrix to be trained, htFor the t GRU output, xtIs a column vector of matrix X; the forward propagation formula is: r ist＝σ(Wrxt+Urht-1)；zt＝σ(Wzxt+Uzht-1)；
Figure FDA0003431884940000023
Figure FDA0003431884940000024
Wherein the function is:
Figure FDA0003431884940000026
output of gated cyclic unit
Figure FDA0003431884940000027
H is to beTAs the input of the fully connected neural network, inputting the input into the network for training; w(i)Is the weight matrix of the i-th layer of the fully-connected neural network, b(i)Is the bias of the i-th layer, z(i)Is the output of the i-th layer, a(i -1)Is an input to the ith layer; the objective function is defined as v ═ WTy + b, then the formula for the i-th layer forward propagation is:z(i)＝W(i)a(i-1)+b(i)；a(i)＝σ(z(i)) And a is a(0)＝hT(ii) a When the number of hidden layers l is 1, the predicted result
Figure FDA0003431884940000028
3. The method for predicting traffic flow status of a road network in a wide range based on federal learning according to claim 3, wherein in the third step,
parameter W for gated cycle cellr，Wz，W，Ur，UzU is trained using back propagation, the formula is as follows:
Figure FDA0003431884940000031
Figure FDA0003431884940000032
Figure FDA0003431884940000033
Figure FDA0003431884940000034
Figure FDA0003431884940000035
Figure FDA0003431884940000036
wherein:
Figure FDA0003431884940000037
Figure FDA0003431884940000038
Figure FDA0003431884940000039
Figure FDA00034318849400000310
Figure FDA00034318849400000311
Figure FDA00034318849400000312
Figure FDA00034318849400000313
Figure FDA00034318849400000314
after a plurality of times of forward propagation and backward propagation, a group of trained parameter sets can be obtained
Figure FDA0003431884940000041
CN202111601256.4A 2021-12-24 2021-12-24 Method for predicting traffic flow state of large-scale road network based on federal learning Pending CN114266406A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111601256.4A CN114266406A (en) 2021-12-24 2021-12-24 Method for predicting traffic flow state of large-scale road network based on federal learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111601256.4A CN114266406A (en) 2021-12-24 2021-12-24 Method for predicting traffic flow state of large-scale road network based on federal learning

Publications (1)

Publication Number Publication Date
CN114266406A true CN114266406A (en) 2022-04-01

Family

ID=80829890

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111601256.4A Pending CN114266406A (en) 2021-12-24 2021-12-24 Method for predicting traffic flow state of large-scale road network based on federal learning

Country Status (1)

Country Link
CN (1) CN114266406A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115311860A (en) * 2022-08-09 2022-11-08 中国科学院计算技术研究所 Online federal learning method of traffic flow prediction model
CN115909746A (en) * 2023-01-04 2023-04-04 中南大学 Traffic flow prediction method, system and medium based on federal learning
CN116346863A (en) * 2023-05-29 2023-06-27 湘江实验室 Vehicle-mounted network data processing method, device, equipment and medium based on federal learning
CN117540215A (en) * 2024-01-05 2024-02-09 中国民航大学 Flight delay prediction model training method, electronic equipment and storage medium

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115311860A (en) * 2022-08-09 2022-11-08 中国科学院计算技术研究所 Online federal learning method of traffic flow prediction model
CN115909746A (en) * 2023-01-04 2023-04-04 中南大学 Traffic flow prediction method, system and medium based on federal learning
CN116346863A (en) * 2023-05-29 2023-06-27 湘江实验室 Vehicle-mounted network data processing method, device, equipment and medium based on federal learning
CN116346863B (en) * 2023-05-29 2023-08-01 湘江实验室 Vehicle-mounted network data processing method, device, equipment and medium based on federal learning
CN117540215A (en) * 2024-01-05 2024-02-09 中国民航大学 Flight delay prediction model training method, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN114266406A (en) Method for predicting traffic flow state of large-scale road network based on federal learning
Li et al. A deep learning method based on an attention mechanism for wireless network traffic prediction
Hu et al. Federated region-learning: An edge computing based framework for urban environment sensing
Xu et al. Deep belief network-based support vector regression method for traffic flow forecasting
Yang et al. Bus arrival time prediction using support vector machine with genetic algorithm
CN113313947B (en) Road condition evaluation method of short-term traffic prediction graph convolution network
CN113268916A (en) Traffic accident prediction method based on space-time graph convolutional network
CN112949999A (en) High-speed traffic accident risk early warning method based on Bayesian deep learning
CN113283581B (en) Multi-fusion graph network collaborative multi-channel attention model and application method thereof
CN113139341B (en) Electric quantity demand prediction method and system based on federal integrated learning
CN109255726A (en) A kind of ultra-short term wind power prediction method of Hybrid Intelligent Technology
Zhou Attention based stack resnet for citywide traffic accident prediction
Wang et al. Ultra-short-term wind power forecasting based on deep belief network
Wang et al. Electric vehicle charging load clustering and load forecasting based on long short term memory neural network
Srinivasan et al. Computational intelligence-based congestion prediction for a dynamic urban street network
Zhang et al. Method of predicting bus arrival time based on MapReduce combining clustering with neural network
Tian et al. Spatio-temporal position prediction model for mobile users based on LSTM
Yang et al. Dynamic origin-destination matrix estimation based on urban Rail transit AFC data: deep optimization framework with forward passing and backpropagation techniques
Li et al. Optimal combinations and variable departure intervals for micro bus system
CN116612633A (en) Self-adaptive dynamic path planning method based on vehicle-road cooperative sensing
Peng et al. Spatial-Temporal Dynamic Forecasting of EVs Charging Load Based on DCC-2D
CN112200372A (en) Method for calculating and guiding passenger sharing rate of land-side comprehensive traffic optimization of large-scale airport hub
CN111144611A (en) Spatial load prediction method based on clustering and nonlinear autoregression
Kosolsombat et al. A review of the prediction method for intelligent transport system
CN115841229A (en) Intelligent scheduling method of internet vehicle connecting system

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