CN115909746A - Traffic flow prediction method, system and medium based on federal learning - Google Patents

Traffic flow prediction method, system and medium based on federal learning Download PDF

Info

Publication number
CN115909746A
CN115909746A CN202310006261.3A CN202310006261A CN115909746A CN 115909746 A CN115909746 A CN 115909746A CN 202310006261 A CN202310006261 A CN 202310006261A CN 115909746 A CN115909746 A CN 115909746A
Authority
CN
China
Prior art keywords
model
client
local
server
parameters
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
CN202310006261.3A
Other languages
Chinese (zh)
Inventor
鲁鸣鸣
何文勇
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.)
Central South University
Original Assignee
Central South 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 Central South University filed Critical Central South University
Priority to CN202310006261.3A priority Critical patent/CN115909746A/en
Publication of CN115909746A publication Critical patent/CN115909746A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a traffic flow prediction method, a system and a medium based on federal learning, wherein the method comprises the following steps: setting hyper-parameters of a server model and initializing to obtain a global model; the server distributes the global model to the client to obtain each local model; each client updates the local model by using the local traffic flow data set; calculating the correlation between each local model update and the global model update, and screening the client according to the correlation; the screened client side sends the local model parameters to a server; the server converges the received local model parameters to complete the global model updating; repeating the steps until the models converge; and finally, predicting the traffic flow by using the converged local model by each client. The invention can avoid invalid parameter uploading and reduce the communication overhead in the process of federal learning training.

Description

Traffic flow prediction method, system and medium based on federal learning
Technical Field
The invention belongs to the technical field of intelligent traffic, and particularly relates to a traffic flow prediction method, a traffic flow prediction system and a traffic flow prediction medium based on federal learning.
Background
With the acceleration of urbanization, population living density becomes more dense, and with the continuous increase of the number of private cars in cities, the demand of residents for public transportation services is rapidly increased. On one hand, the urban environment is rapidly deteriorated due to a large amount of exhaust gas emitted by automobiles; on the other hand, the traffic congestion degree is more serious due to the intensive traveling of the vehicles. The problems not only cause the consumption of a large amount of manpower and financial resources, but also seriously affect the travel experience of people. Therefore, how to relieve traffic pressure and improve urban trip efficiency is an urgent research problem. In recent years, an Intelligent Transportation System (ITS) [1] can prevent some traffic congestion and accidents in time by providing a rational traffic management decision. Therefore, research and development of intelligent transportation systems are receiving more and more attention from researchers. Traffic flow prediction is an important research field in intelligent traffic systems [2], and can effectively depict real-time traffic flow conditions of roads and capture change rules of traffic flow on the roads along with time, so the traffic flow prediction is widely applied to various applications of the traffic systems.
At present, deep learning has performed well in a series of traffic prediction problems. The existing traffic flow prediction method mainly focuses on how to train and predict traffic flow data collected by a sensor in a cloud by using a machine learning model, so that on one hand, the computing power of edge end sensor equipment is not fully utilized, and on the other hand, the risk of data leakage in the transmission process is caused. In addition, as the amount of traffic data increases explosively, the increase of the GPU/CPU computing power is relatively slow, which makes the cloud server far from meeting the computing requirements of real scenes [3]. In recent years, as computing, storage, and the like of terminal equipment have been greatly improved, researchers have begun to put part of services and computing down on terminal equipment, and have proposed a completely new solution in combination with Federal Learning (FL) [4 ]. The scheme converts a centralized training mode into a distributed terminal equipment cooperative training mode, thereby effectively solving the problems. The current traffic flow prediction work based on federal learning is either too simple to have insufficient representation capacity [5], so that the model has the problem of poor prediction performance; or the algorithm design is too complex [6], resulting in huge communication overhead at the edge end and the computation end.
Disclosure of Invention
In order to solve the problems of insufficient performance and huge communication overhead in the conventional federal learning-based traffic flow prediction method, the invention provides a federal learning-based traffic flow prediction method, a system and a medium with lighter communication weight, which can avoid invalid parameter uploading and reduce the communication overhead in the federal learning training process.
In order to achieve the technical purpose, the invention adopts the following technical scheme:
a traffic flow prediction method based on federal learning comprises the following steps:
step 1, setting hyper-parameters of a server model, and initializing a basic structure of the model to obtain global model parameters
Figure 541088DEST_PATH_IMAGE001
(ii) a The parameter subscript G designates the global model, superscript @>
Figure 756431DEST_PATH_IMAGE002
Represents iteration turns, and makes initialization->
Figure 67327DEST_PATH_IMAGE003
；
Step 2, the server sends the global model parameters
Figure 196957DEST_PATH_IMAGE001
Distributing the parameters to each client to obtain the local model parameters of each client>
Figure 542488DEST_PATH_IMAGE004
(ii) a In the parameter subscript>
Figure 161688DEST_PATH_IMAGE005
Respectively denote>
Figure 959879DEST_PATH_IMAGE006
A client-side local model;
step 3, training local models respectively by using local traffic flow data sets of all clients to finish the first local model
Figure 893200DEST_PATH_IMAGE007
Updating the wheel to obtain the parameters of each local model as->
Figure 358817DEST_PATH_IMAGE008
；
Step 4, calculating the correlation between local model update of each client and global model update of the server, and screening the clients to be uploaded with local model parameters to the server in the current turn according to the correlation;
step 5, the selected client sides respectively send the local model parameters to a server;
step 6, the server converges the received local model parameters, namely the global model is the first
Figure 883339DEST_PATH_IMAGE007
Updating the wheel to obtain a global model parameter of ^ 4>
Figure 372089DEST_PATH_IMAGE009
；
Step 7, updating
Figure 640260DEST_PATH_IMAGE010
Repeating the steps 2 to 6 until the global model and each local model are converged;
and 8, predicting the local traffic flow by each client by using the local model.
Further, the hyper-parameters set in step 1 include a learning rate
Figure 461847DEST_PATH_IMAGE011
Number of communication rounds->
Figure 422850DEST_PATH_IMAGE012
Local training round>
Figure 398896DEST_PATH_IMAGE013
。
Further, the basic structure of the model adopts an Encoder-Decoder architecture: encoder module using gating basedThe recurrent neural network captures the context information in the input traffic flow time series, namely, converts the hidden time dynamic characteristics in the input time series into an intermediate hidden vector
Figure 205178DEST_PATH_IMAGE014
(ii) a The Decoder module uses a gate-based recurrent neural network and a fully-connected network for result prediction.
Further, step 4 comprises:
(1) Calculating the local model of each client
Figure 645387DEST_PATH_IMAGE015
Updating the wheel:
Figure 511712DEST_PATH_IMAGE016
in the formula (I), the compound is shown in the specification,
Figure 975054DEST_PATH_IMAGE017
indicates the fifth->
Figure 585027DEST_PATH_IMAGE018
Individual client in the fifth or fifth place>
Figure 614163DEST_PATH_IMAGE002
Partial model update on wheel, based on a comparison of the number of partial models in the wheel and the number of partial models in the wheel>
Figure 182547DEST_PATH_IMAGE019
And &>
Figure 165809DEST_PATH_IMAGE020
Respectively denote a fifth->
Figure 782735DEST_PATH_IMAGE018
Individual client local model is at ^ h>
Figure 931957DEST_PATH_IMAGE007
Wheel and a fifth->
Figure 405664DEST_PATH_IMAGE002
A parameter of the wheel;
(2) Computing server Global model number
Figure 374757DEST_PATH_IMAGE021
Updating of the wheel to evaluate it approximately as ^ based>
Figure 529794DEST_PATH_IMAGE002
Updating the wheel:
Figure 799102DEST_PATH_IMAGE022
in the formula (I), the compound is shown in the specification,
Figure 443710DEST_PATH_IMAGE023
presentation server in the fifth or fifth place>
Figure 103361DEST_PATH_IMAGE002
Global model update of the wheel;
(3) Calculating the correlation between the local model update and the global model update of each client, specifically by comparing
Figure 858828DEST_PATH_IMAGE017
And &>
Figure 717062DEST_PATH_IMAGE023
The number of parameters with the same symbol in the Chinese character is measured and expressed as:
Figure 34036DEST_PATH_IMAGE024
in the formula (I), the compound is shown in the specification,
Figure 180984DEST_PATH_IMAGE025
represents the correlation of local and global model updates for client k @>
Figure 740141DEST_PATH_IMAGE026
An index representing the parameters of the model is determined,Mrepresents the total number of model parameters, <' > is selected>
Figure 718461DEST_PATH_IMAGE027
Represents->
Figure 439293DEST_PATH_IMAGE017
Corresponds to the partial model->
Figure 73536DEST_PATH_IMAGE028
Is/are>
Figure 170805DEST_PATH_IMAGE026
An updated value of a parameter->
Figure 269211DEST_PATH_IMAGE029
Represents->
Figure 160944DEST_PATH_IMAGE023
Corresponds to the global pattern G £ th>
Figure 282484DEST_PATH_IMAGE026
Update values of individual parameters; sgn denotes a sign calculation function;
(4) Screening all clients: and if the correlation between the local model update and the global model update of a certain client is greater than a set correlation threshold, the local model of the client is a model which meets the requirement of uploading to the server.
Further, the server uses a FedAVG algorithm to converge the received local model parameters.
A traffic flow prediction system based on federal learning comprises 1 server and
Figure 183444DEST_PATH_IMAGE030
a client, the server and each client comprising a memory and a processor, the memory having stored therein a computer program which, when executed by the processor, causes the processors of the server and the client to cooperate to implement any of the aboveThe technical scheme is that the traffic flow prediction method based on the federal learning is adopted.
A computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the federal learning based traffic flow prediction method in any of the above technical solutions.
Advantageous effects
Compared with the prior art, the invention has the beneficial effects that:
the method of the invention filters the client model parameters which are effective for updating the global model according to the correlation by calculating the correlation between the local model update of each client and the global model update of the server, avoids the uploading of ineffective parameters, and greatly reduces the communication overhead of the whole system under the condition of ensuring equivalent precision. In addition, compared with the traditional centralized method, the method not only makes full use of the computing power of the terminal equipment, but also ensures the data privacy of each equipment.
Drawings
FIG. 1 is a flow chart of an embodiment of the method of the present invention;
FIG. 2 is a schematic structural diagram of a global model in the method of the present invention;
FIG. 3 is a schematic diagram of the overall structure of the process of the present invention;
FIG. 4 is a graph comparing the results of the method of the present invention and other algorithms in terms of prediction accuracy;
FIG. 5 is a comparison graph of the results of the method of the present invention and other algorithms on the communication overhead.
Detailed Description
In order to make the technical solutions in the embodiments of the present invention better understood and make the above objects, features and advantages of the present invention more obvious and understandable, the technical solutions of the present invention are further described in detail below with reference to the accompanying drawings:
the flow chart of the specific embodiment of the method of the invention is shown in fig. 1, and the process is as follows:
step 1, setting hyper-parameters of server model, such as learning rate
Figure 372242DEST_PATH_IMAGE011
Number of communication rounds->
Figure 700455DEST_PATH_IMAGE012
Local training round>
Figure 43712DEST_PATH_IMAGE013
Etc., and initializing the basic structure of the model to obtain global model parameters>
Figure 13942DEST_PATH_IMAGE001
(ii) a The parameter subscript G designates the global model, superscript @>
Figure 555782DEST_PATH_IMAGE002
Represents iteration turns, and makes initialization->
Figure 54896DEST_PATH_IMAGE003
So that the initialized global model is denoted as +>
Figure 885449DEST_PATH_IMAGE031
。
In order to better capture the time sequence characteristics of dynamic change in traffic flow data, the method adopts an Encoder-Decoder architecture as the infrastructure of a global model, and the structure of the Encoder-Decoder architecture is shown in figure 2. The Encoder module captures context information on a time sequence by using a gate-controlled recurrent neural network (GRU), and converts hidden time dynamic characteristics in an input time sequence into an intermediate hidden vector
Figure 128211DEST_PATH_IMAGE014
. And the Decoder module uses GRU and full connection network (FNN) to predict the result. By initializing a global model->
Figure 790137DEST_PATH_IMAGE001
(r = 0), r represents the iteration count, and is thus ready for subsequent model distribution.
Step 2, the server sends the global model parameters
Figure 460153DEST_PATH_IMAGE001
Distributing the parameters to each client to obtain the local model parameters of each client>
Figure 778001DEST_PATH_IMAGE004
(ii) a In parameter subscript>
Figure 48622DEST_PATH_IMAGE005
Respectively denote->
Figure 830633DEST_PATH_IMAGE006
And (4) local models of the clients.
Flow (1) in fig. 3 represents a process in which the server distributes parameters to the respective clients.
Step 3, training local models respectively by using local traffic flow data sets of all clients to finish the first local model
Figure 671550DEST_PATH_IMAGE007
Updating the wheel to obtain the parameters of each local model as->
Figure 476695DEST_PATH_IMAGE008
。
Flow (2) in fig. 3 represents the client-side local model training process.
First, the method of the invention uses
Figure 61260DEST_PATH_IMAGE032
Represents an input time series of the traffic flow, wherein n represents the length of the series, and ` er `>
Figure 697778DEST_PATH_IMAGE033
For the time series->
Figure 912858DEST_PATH_IMAGE034
D represents the characteristic dimension of traffic flow data at each time point. The Encoder will output one after reading the whole sequence XAn intermediate hidden state vector
Figure 736458DEST_PATH_IMAGE014
Representing a potential dynamic temporal pattern in the time series. In the client k, the calculation process is as follows:
Figure 124714DEST_PATH_IMAGE035
（1）
wherein
Figure 615738DEST_PATH_IMAGE036
Representing the initial hidden state vector.
Next, the method of the present embodiment uses
Figure 1720DEST_PATH_IMAGE037
Representing the time series predicted by the model, wherein n represents the length of the predicted sequence. The module is characterized in that the input of each time step is the output of the previous time step. It is noted that the input at the first time step is the intermediate concealment vector ≧ which is the Encoder output>
Figure 79660DEST_PATH_IMAGE014
And the last value of the input sequence->
Figure 6027DEST_PATH_IMAGE038
. In the client k, the calculation process is as follows:
Figure 351558DEST_PATH_IMAGE039
（2）。
finally, the method of this embodiment uses a gradient descent method to update the parameters, and the procedure is as follows:
Figure 970758DEST_PATH_IMAGE040
（3）
wherein
Figure 972212DEST_PATH_IMAGE041
Indicates a learning rate, is selected>
Figure 702271DEST_PATH_IMAGE042
Indicating that the gradient is calculated.
And 4, calculating the correlation between local model update of each client and global model update of the server, and screening the clients to be uploaded with local model parameters to the server in the current turn according to the correlation.
Flow (3) in fig. 3 represents calculating the correlation between the local model update and the global model update of the client, and the method of the present embodiment divides the correlation into the following key 4 processes:
(1) And (3) calculating the update of the r-th round of the local model of each client:
Figure 902308DEST_PATH_IMAGE043
（4）/>
in the formula (I), the compound is shown in the specification,
Figure 692410DEST_PATH_IMAGE017
indicates the fifth->
Figure 977897DEST_PATH_IMAGE018
Individual client is on the ^ h>
Figure 449330DEST_PATH_IMAGE002
Partial model update of the wheel, ->
Figure 5339DEST_PATH_IMAGE019
And &>
Figure 231921DEST_PATH_IMAGE020
Respectively denote a fifth->
Figure 739125DEST_PATH_IMAGE018
Individual client local model is at ^ h>
Figure 810987DEST_PATH_IMAGE007
Wheel and a fifth->
Figure 188878DEST_PATH_IMAGE002
A parameter of the wheel;
(2) Computing server Global model number
Figure 55203DEST_PATH_IMAGE021
Updating of the wheel, evaluating it approximately as ^ based>
Figure 580862DEST_PATH_IMAGE002
And (3) updating the wheel:
Figure 190835DEST_PATH_IMAGE022
（4）
in the formula (I), the compound is shown in the specification,
Figure 219971DEST_PATH_IMAGE023
indicating that the server is in the ^ th->
Figure 460460DEST_PATH_IMAGE002
Global model update of the wheel;
(3) Calculating the correlation between the local model update and the global model update of each client, specifically by comparing
Figure 709301DEST_PATH_IMAGE017
And &>
Figure 388544DEST_PATH_IMAGE023
The number of parameters with the same symbol in the Chinese character is measured and expressed as:
Figure 272186DEST_PATH_IMAGE044
（6）
Figure 949155DEST_PATH_IMAGE045
（7）
in the formula (I), the compound is shown in the specification,
Figure 918248DEST_PATH_IMAGE025
represents the correlation of local and global model updates for client k @>
Figure 135603DEST_PATH_IMAGE026
An index representing the parameters of the model,Mrepresents the total number of model parameters, <' > is selected>
Figure 139331DEST_PATH_IMAGE027
Represents->
Figure 721622DEST_PATH_IMAGE017
Corresponds to the partial model->
Figure 178011DEST_PATH_IMAGE028
Is/are>
Figure 199057DEST_PATH_IMAGE026
An updated value of a parameter->
Figure 293177DEST_PATH_IMAGE029
Represents->
Figure 374265DEST_PATH_IMAGE023
Corresponds to the global pattern G £ th>
Figure 521213DEST_PATH_IMAGE026
Update values of individual parameters; sgn denotes a sign calculation function;
(4) Screening all clients: and if the correlation between the local model update and the global model update of a certain client is greater than a set correlation threshold, the local model of the client is a model which meets the requirement of uploading to the server.
And 5, respectively sending the local model parameters to a server by each screened client.
Flow (4) in fig. 3 represents the process of each client sending the local model parameters to the server.
Step 6, the server uses a FedAVG algorithm to converge the received local model parameters, namely the global model is the first
Figure 814791DEST_PATH_IMAGE007
Updating the wheel to obtain a global model parameter of>
Figure 793111DEST_PATH_IMAGE019
。
The flow (5) in fig. 3 represents the process of aggregating the parameters by the server.
Step 7, updating
Figure 982784DEST_PATH_IMAGE010
And repeating the steps 2 to 6 until the global model and each local model converge.
And 8, predicting the local traffic flow by each client by using the local model.
In the traffic flow prediction experiment performed by the embodiment of the invention, the main evaluation indexes comprise a mean square error, a root mean square error and an average absolute error. The data sets used in this example are the METR-LA and PEMS-BAY traffic stream data sets. The METR-LA contained 207 monitored values of highway traffic speed in los Angeles county in 4 months, with each sensor monitored every five minutes for a total of 34249 data samples; PEMS-BAY contains 325 sensors monitoring BAY expressway vehicle flow speed within 6 months, and comprises 52093 data samples in total.
Fig. 4 a) compares the predicted performance of this example method (CM-FedSeq 2 Seq) with other methods on the METR-LA dataset. Firstly, compared with some classical centralized methods, the CM-FedSeq2Seq of the method has better results on multiple evaluation indexes. In particular, CM-FedSeq2Seq reduced the MAE and MAPE errors by 38.4% and 44.0%, respectively, compared to ARIMA. In addition, compared with the existing federal learning method, the method CM-FedSeq2Seq of the embodiment also far exceeds the two latest methods FedGRU and CNFGNN. Particularly, compared with FedGRU, the CM-FedSeq2Seq reduces the error indexes of each index by 24.6 percent, 19.2 percent and 25.1 percent respectively; CM-FedSeq2Seq showed a 4.5% reduction in RMSE index compared to CNFGNN (only RMSE was compared here since CNFGNN only provided an RMSE error index).
Fig. 4 b) compares the predicted performance of this example method (CM-FedSeq 2 Seq) with other methods on PEMS-BAY data sets. First, compared with the compared centralized method, the CM-FedSeq2Seq of the present embodiment method has better results in each evaluation index. In particular, CM-FedSeq2Seq reduced the MAE, RMSE, and MAPE errors by 27.4%, 23.4%, and 29.8%, respectively, compared to FC-LSTM (the better accurate model). Meanwhile, compared with the existing federal learning method, the CM-FedSeq2Seq of the method is also superior to the two latest methods FedGRU and CNFGNN. Particularly, compared with FedGRU, the CM-FedSeq2Seq reduces the error indexes of each index by 22.9%, 21.6% and 27.5% respectively; the RMSE index of CM-FedSeq2Seq was reduced by 0.52% compared to CNFGNN.
In order to highlight the advantages of the method in the communication overhead, the communication overhead conditions of the methods in the training phase are compared. As shown in fig. 5, the CM-FedSeq2Seq of the present embodiment also ensures that the communication overhead of the CM-FedSeq2Seq is far smaller than that of CNFGNN on the premise that the RMSE index is better than that of the conventional method CNFGNN.
Experiments show that the method has higher prediction accuracy and lower communication overhead, so that the method is a more effective federal learning traffic flow prediction method.
The above embodiments are preferred embodiments of the present application, and those skilled in the art can make various changes or modifications without departing from the general concept of the present application, and such changes or modifications should fall within the scope of the claims of the present application.
Reference is made herein to:
[1] Lin Y, Wang P, Ma M. Intelligent transportation system (ITS): Concept, challenge and opportunity[C]//2017 ieee 3rd international conference on big data security on cloud (bigdatasecurity), ieee international conference on high performance and smart computing (hpsc), and ieee international conference on intelligent data and security (ids). IEEE, 2017: 167-172.
[2] Luo Q, Zhou Y. Spatial-temporal Structures of Deep Learning Models for Traffic Flow Forecasting: A Survey[C]//2021 4th International Conference on Intelligent Autonomous Systems (ICoIAS). IEEE, 2021: 187-193s.
[3] Guo Y, Zhao R, Lai S, et al. Distributed machine learning for multiuser mobile edge computing systems[J]. IEEE Journal of Selected Topics in Signal Processing, 2022.
[4] Khan L U, Saad W, Han Z, et al. Federated learning for internet of things: Recent advances, taxonomy, and open challenges[J]. IEEE Communications Surveys & Tutorials, 2021.
[5] Liu Y, James J Q, Kang J, et al. Privacy-preserving traffic flow prediction: A federated learning approach[J]. IEEE Internet of Things Journal, 2020, 7(8): 7751-7763.
[6] Meng C, Rambhatla S, Liu Y. Cross-node federated graph neural network for spatio-temporal data modeling[C]//Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2021: 1202-1211。

Claims (7)

1. a traffic flow prediction method based on federal learning is characterized by comprising the following steps:
step 1, setting hyper-parameters of a server model, and initializing a basic structure of the model to obtain global model parameters
Figure 939413DEST_PATH_IMAGE001
(ii) a The parameter subscript G designates the global model, superscript @>
Figure 453089DEST_PATH_IMAGE002
Represents iteration turns, and makes initialization->
Figure 140423DEST_PATH_IMAGE003
；
Step 2, the server sends the global model parameters
Figure 203057DEST_PATH_IMAGE001
Distributing the parameters to each client to obtain the local model parameters of each client>
Figure 952838DEST_PATH_IMAGE004
(ii) a In the parameter subscript>
Figure 391910DEST_PATH_IMAGE005
Respectively denote->
Figure 668170DEST_PATH_IMAGE006
A client-side local model;
step 3, training local models respectively by using local traffic flow data sets of all clients to finish the first local model
Figure 777072DEST_PATH_IMAGE007
Updating the wheel to obtain the parameters of each local model as->
Figure 404362DEST_PATH_IMAGE008
；
Step 4, calculating the correlation between local model update of each client and global model update of the server, and screening the clients to be uploaded with local model parameters to the server in the current turn according to the correlation;
step 5, the screened client sides respectively send local model parameters to a server;
step 6, the server converges the received local model parameters, namely the global model is the first
Figure 522491DEST_PATH_IMAGE007
Updating the wheel to obtain a global model parameter of ^ 4>
Figure 918837DEST_PATH_IMAGE009
；
Step 7, updating
Figure 588853DEST_PATH_IMAGE010
Repeating the steps 2 to 6 until the global model and each local model are converged;
and 8, predicting the local traffic flow by each client by using the local model.
2. The federally-learned traffic flow prediction method according to claim 1, wherein the hyper-parameter set in step 1 includes a learning rate
Figure 47647DEST_PATH_IMAGE011
Number of communication rounds->
Figure 828521DEST_PATH_IMAGE012
And local training round>
Figure 344953DEST_PATH_IMAGE013
。
3. The federally-learned traffic flow prediction method according to claim 1, wherein the basic structure of the model employs an Encoder-Decoder architecture: the Encoder module uses a gate-based recurrent neural network to capture the context information in the input traffic flow time sequence, namely, the hidden time dynamic characteristics in the input time sequence are converted into an intermediate hidden vector
Figure 795657DEST_PATH_IMAGE014
(ii) a The Decoder module uses a gate-based recurrent neural network and a fully-connected network for result prediction.
4. The federal learning based traffic flow prediction method as claimed in claim 1, wherein step 4 includes:
(1) Calculating the local model number of each client
Figure 131961DEST_PATH_IMAGE002
Updating the wheel:
Figure 591892DEST_PATH_IMAGE015
in the formula (I), the compound is shown in the specification,
Figure 697251DEST_PATH_IMAGE016
indicates the fifth->
Figure 709070DEST_PATH_IMAGE017
Individual client is on the ^ h>
Figure DEST_PATH_IMAGE018
Partial model update of the wheel, ->
Figure 939194DEST_PATH_IMAGE019
And &>
Figure 796291DEST_PATH_IMAGE020
Respectively denote a fifth->
Figure 159752DEST_PATH_IMAGE017
Individual client side local model in ^ th>
Figure 76892DEST_PATH_IMAGE007
Wheel and a fifth->
Figure 122209DEST_PATH_IMAGE002
A parameter of the wheel;
(2) Computing server Global model number
Figure 923943DEST_PATH_IMAGE021
Updating of the wheel, evaluating it approximately as ^ based>
Figure 3894DEST_PATH_IMAGE002
Updating the wheel:
Figure 232881DEST_PATH_IMAGE022
in the formula (I), the compound is shown in the specification,
Figure 765494DEST_PATH_IMAGE023
indicating that the server is in the ^ th->
Figure 229973DEST_PATH_IMAGE002
Global model update of the wheel;
(3) Calculating the correlation between the local model update and the global model update of each client, specifically by comparing
Figure 305377DEST_PATH_IMAGE016
And &>
Figure 564320DEST_PATH_IMAGE023
The number of parameters with the same symbol in the Chinese character is measured and expressed as: />
Figure 584228DEST_PATH_IMAGE024
In the formula (I), the compound is shown in the specification,
Figure DEST_PATH_IMAGE025
represents the correlation of local and global model updates for client k @>
Figure 993344DEST_PATH_IMAGE026
An index representing the parameters of the model,Mrepresents the total number of model parameters, <' > is selected>
Figure 47888DEST_PATH_IMAGE027
Represents->
Figure 477732DEST_PATH_IMAGE016
In correspondence with a partial model>
Figure 860303DEST_PATH_IMAGE028
Is/are>
Figure 932164DEST_PATH_IMAGE026
An updated value of a parameter->
Figure 575635DEST_PATH_IMAGE029
Represents->
Figure 317326DEST_PATH_IMAGE023
Corresponds to the global pattern G £ th>
Figure 46248DEST_PATH_IMAGE026
Update values of individual parameters; sgn represents a sign calculation function;
(4) Screening all clients: and if the correlation between the local model update and the global model update of a certain client is greater than a set correlation threshold, the local model of the client is a model which meets the requirement of uploading to the server.
5. The federal learning based traffic flow prediction method of claim 1, wherein the server aggregates the received local model parameters using a FedAVG algorithm.
6. A traffic flow prediction system based on federal learning is characterized by comprising 1 server and
Figure 656221DEST_PATH_IMAGE030
a client, the server and each client compriseA memory and a processor, the memory having stored therein a computer program that, when executed by the processor, causes the processor of the server and the client to jointly implement the method of any one of claims 1 to 5.
7. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the method according to any one of claims 1 to 5.
CN202310006261.3A 2023-01-04 2023-01-04 Traffic flow prediction method, system and medium based on federal learning Pending CN115909746A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310006261.3A CN115909746A (en) 2023-01-04 2023-01-04 Traffic flow prediction method, system and medium based on federal learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310006261.3A CN115909746A (en) 2023-01-04 2023-01-04 Traffic flow prediction method, system and medium based on federal learning

Publications (1)

Publication Number Publication Date
CN115909746A true CN115909746A (en) 2023-04-04

Family

ID=86473049

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310006261.3A Pending CN115909746A (en) 2023-01-04 2023-01-04 Traffic flow prediction method, system and medium based on federal learning

Country Status (1)

Country Link
CN (1) CN115909746A (en)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112181971A (en) * 2020-10-27 2021-01-05 华侨大学 Edge-based federated learning model cleaning and equipment clustering method, system, equipment and readable storage medium
CN114266406A (en) * 2021-12-24 2022-04-01 北京航空航天大学 Method for predicting traffic flow state of large-scale road network based on federal learning
CN114372046A (en) * 2021-05-13 2022-04-19 青岛亿联信息科技股份有限公司 Parking flow prediction model training method based on federal learning
CN114676849A (en) * 2022-03-24 2022-06-28 支付宝(杭州)信息技术有限公司 Method and system for updating model parameters based on federal learning
CN114676838A (en) * 2022-04-12 2022-06-28 支付宝(杭州)信息技术有限公司 Method and device for jointly updating model
CN114866545A (en) * 2022-04-19 2022-08-05 郑州大学 Semi-asynchronous layered federal learning method and system based on air calculation
CN114925857A (en) * 2022-06-20 2022-08-19 东北大学秦皇岛分校 Federal learning algorithm for traffic state estimation
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

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112181971A (en) * 2020-10-27 2021-01-05 华侨大学 Edge-based federated learning model cleaning and equipment clustering method, system, equipment and readable storage medium
CN114372046A (en) * 2021-05-13 2022-04-19 青岛亿联信息科技股份有限公司 Parking flow prediction model training method based on federal learning
CN114266406A (en) * 2021-12-24 2022-04-01 北京航空航天大学 Method for predicting traffic flow state of large-scale road network based on federal learning
CN114676849A (en) * 2022-03-24 2022-06-28 支付宝(杭州)信息技术有限公司 Method and system for updating model parameters based on federal learning
CN114676838A (en) * 2022-04-12 2022-06-28 支付宝(杭州)信息技术有限公司 Method and device for jointly updating model
CN114866545A (en) * 2022-04-19 2022-08-05 郑州大学 Semi-asynchronous layered federal learning method and system based on air calculation
CN114925857A (en) * 2022-06-20 2022-08-19 东北大学秦皇岛分校 Federal learning algorithm for traffic state estimation
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

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
LU MINGMING 等: "Urban Traffic Flow Forecast Based on FastGCRNN", 《JOURNAL OF ADVANCED TRANSPORTATION》 *
LUPING WANG 等: "CMFL: Mitigating Communication Overhead for Federated Learning", 《2019 IEEE 39TH INTERNATIONAL CONFERENCE ON DISTRIBUTED COMPUTING SYSTEMS (ICDCS)》 *
吴壮 等: "车联网中基于轨迹预测的无人机动态协同优化覆盖算法", 《计算机应用研究》 *

Similar Documents

Publication Publication Date Title
Zeng et al. Parking occupancy prediction method based on multi factors and stacked GRU-LSTM
GB2599765A (en) Vehicle traffic flow prediction method with missing data
CN113065074B (en) Track destination prediction method based on knowledge graph and self-attention mechanism
CN109584552A (en) A kind of public transport arrival time prediction technique based on network vector autoregression model
Tian et al. A network traffic prediction method based on IFS algorithm optimised LSSVM
CN111768625A (en) Traffic road event prediction method based on graph embedding
CN115691137B (en) Multi-modal data prediction method based on causal Markov model
CN111081022A (en) Traffic flow prediction method based on particle swarm optimization neural network
Liu et al. A method for short-term traffic flow forecasting based on GCN-LSTM
CN110766066B (en) Tensor heterogeneous integrated vehicle networking missing data estimation method based on FNN
Wang et al. Air pollution prediction via graph attention network and gated recurrent unit
CN115204477A (en) Bicycle flow prediction method of context awareness graph recursive network
CN112950926A (en) Urban trunk road speed prediction method based on big data and deep learning
Zhang Short-term traffic flow prediction methods: A survey
Wang et al. Freeway short-term travel time prediction based on lightgbm algorithm
CN113641749A (en) Population mobility network estimation method and system based on bipartite graph
CN112562312A (en) GraphSAGE traffic network data prediction method based on fusion characteristics
CN107194505A (en) A kind of method and system that bus travel amount is predicted based on city big data
CN115909746A (en) Traffic flow prediction method, system and medium based on federal learning
Shao et al. Estimation of urban travel time with sparse traffic surveillance data
Miao et al. A queue hybrid neural network with weather weighted factor for traffic flow prediction
CN111371609B (en) Internet of vehicles communication prediction method based on deep learning
Yang et al. Vehicle information influence degree screening method based on GEP optimized RBF neural network
Li et al. Dynamic Performance Analysis of STEP System in Internet of Vehicles Based on Queuing Theory
Cheng et al. ETC Intelligent Navigation Path Planning 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