A distributed traffic flow prediction method based on personalized federated learning

Abstract

The present invention relates to the technical field of distributed traffic flow prediction, and discloses a distributed traffic flow prediction method based on personalized federated learning, comprising the following steps: establishing a federated learning framework for the purpose of cross-regional collaborative training, using local data to train a global model, and adopting a multi-task training method, under the premise of not sharing data, allowing each road area edge server equipped with a local neural network algorithm to maintain partial privatization of the neural network layer, while training to obtain a personalized prediction model that conforms to the local data characteristics, effectively improving the influence of the heterogeneity of traffic flow data on the prediction accuracy, using the last round of personalized prediction model as the local prediction model, better retaining the data characteristics of the edge servers of each road area, and better fitting non-human factor errors in the prediction work.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed traffic flow prediction, and in particular to a distributed traffic flow prediction method based on personalized federated learning. Background Art

[0002] In traffic prediction scenarios, real-time and accurate prediction methods are crucial for the control and management of traffic flow. Traditional machine learning algorithms can perform deep mining and preprocessing of traffic flow data and train traffic flow prediction models. However, this method requires a large amount of data support. When exchanging data during the cooperation and sharing process between data providers in different regions in the intelligent transportation system, there is a potential risk of privacy leakage, and it is impossible to train prediction models by sharing data across regions. Without shared data, traditional methods can no longer accurately complete the task of traffic flow prediction.

[0003] In the training process of federated learning, the local data of the client is usually not independently distributed, which will affect the convergence of the federated learning algorithm and the final global model accuracy, and the model characteristics will tend to be more weighted clients. In the traffic flow prediction of intelligent transportation systems, the problem of data differences in each region in multi-region collaborative traffic prediction is that training a single prediction model cannot meet the traffic flow prediction needs of all regions. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the existing technology, the present invention provides a distributed traffic flow prediction method based on personalized federated learning, which has the advantages of improving the accuracy of traffic flow prediction and facilitating efficient planning of traffic management. It solves the problem that traditional federated algorithms cannot train prediction models by sharing data across regions, and the data differences are large and the traffic flow prediction accuracy is low.

[0006] (II) Technical solution

[0007] To achieve the above object, the present invention provides the following technical solution: a distributed traffic flow prediction method based on personalized federated learning, comprising the following steps:

[0008] Step 1: Establish a federated learning framework for cross-regional collaborative training, use local data to train the global model, and adopt a multi-task training method to allow each road area edge server equipped with a local neural network algorithm to maintain partial privatization of the neural network layer while training a personalized prediction model that conforms to the characteristics of local data;

[0009] Step 2: Traffic flow data is collected and preprocessed through the road area edge server equipped with a local neural network algorithm, and then the preprocessed traffic flow data is uploaded to the central cloud server;

[0010] Step 3: The central cloud server equipped with the federated learning algorithm sends the initial parameters of the global model to the road area edge servers equipped with the local neural network algorithm. Each road area edge server uses local data to perform federated learning training to obtain an updated local prediction model.

[0011] Step 4: Add a multi-task algorithm to the federated learning process. After completing the local update in step 3, each road area edge server randomly samples traffic flow data for separate training based on the global model issued, and retains the obtained model parameters in the privatized recurrent neural network LN layer. The model parameters are used for local prediction model training;

[0012] The road area edge server uses the LN layer in the recurrent neural network as a patch layer to perform personalized processing of federated learning. The LN layer is a normalization processing intermediate layer in the convolutional neural network, connecting the channels of all input data and normalizing them to a normal distribution mode with a mean of 0 and a variance of 1. The local prediction model parameters updated by the road area edge server have the same characteristics as the data in the LN layer, and the local prediction model parameters are stored in the LN layer for the next round of federated training.

[0013] The recurrent neural network consists of an input layer, an output layer, and a hidden layer, and the inter-layer neurons at different times are connected through a loop structure. The LN layer is placed before the nonlinear layer. When training the recurrent neural network, the road area edge server performs standard normalization processing on all channels corresponding to the image in a batch, and performs whitening preprocessing on the input data of the LN layer. The hyperparameters recorded by the LN layer during the training process are directly applied to the next round of federated training to obtain standard distribution data with a mean of 0 and a variance of 1. The input expression of the preprocessed LN layer is as follows:

[0014]

[0015] In the formula, represents the distribution data of neurons, The minimum value introduced to prevent the variance from being 0 and causing invalid calculations. Indicates Data at a point in time, represents the mean value of the neuron, represents the variance of the neuron;

[0016] Step 5: After each road area edge server completes the local update, it uploads the new round of local prediction model obtained in step 3 to the central cloud server;

[0017] Step 6. After the central cloud server receives the model parameters uploaded by the edge servers in each road area, it uses the federated learning algorithm to aggregate and update them to obtain a new global model. If the model accuracy meets the preset requirements, the personalized prediction model of the last round of training is used as the local prediction model. If the model accuracy does not meet the preset requirements, the global model parameters are re-issued and the above steps 1 to 5 are repeated until the global model meets the preset requirements.

[0018] Preferably, in the step 2, the road area edge server equipped with a local neural network algorithm collects traffic flow data through a network connection to an intelligent device, the intelligent device includes sensors and cameras, the traffic flow data includes vehicle volume, vehicle type, vehicle speed and road occupancy rate, and the road area edge server performs data preprocessing through a Gaussian denoiser to remove data noise in the traffic flow data.

[0019] Preferably, in step 3, each road area edge server uses traffic flow data to perform federated learning training to obtain an updated local prediction model. During the training process, the parameter update formula of the local prediction model is as follows:

[0020]

[0021] In the formula, Indicates In the round of federated learning, the edge servers in multiple road areas The local prediction model of the edge server in each road area, Indicates the number of The local prediction model of the edge server in each road area, represents the learning rate, Represents the local gradient function, which performs multiple gradient descents according to the computing power and communication latency of the edge servers in different road areas. Indicates that according to the last round of federated learning The local prediction model of the edge server in each road area is gradient descended to obtain an updated local prediction model.

[0022] Preferably, in step 4, the central cloud server receives the local prediction model updated by the edge server of each road area After that, the next round of global model is obtained through aggregation formula. The edge servers in each road area randomly sample traffic flow data to update the personalized prediction model. The loss function in personalized federated learning is composed of the global model and the personalized prediction model, and its expression is as follows:

[0023]

[0024] In the formula, represents the personalized prediction model trained by the edge server in the road area, represents the global model parameters of the previous round of training. If the road area edge server is in the first round of training, ω represents the initial model sent by the central cloud server. It indicates the influence of the global model on the personalized prediction model. The influence intensity is related to the computing power of the edge server in the road area and the degree of data heterogeneity. If the local prediction model is very different from the global model, then The bigger.

[0025] Preferably, in step 4, the calculation formula of the personalized federated learning objective function is as follows:

[0026]

[0027] In the formula, represents the personalized federated learning objective function, represents the dimension of the sampled traffic flow data, It means that personalized federated learning is performed according to the loss function and the sampled traffic flow data, and the minimum value is retained.

[0028] Preferably, in step 4, the recurrent neural network LN layer performs multiple cumulative multiplications during the input and output data processes. To prevent the cumulative multiplication result from being 0, the nonlinear feature is introduced into the LN layer by setting an activation function. Assume that the activation function is , then the final LN layer output formula is as follows:

[0029]

[0030] In the formula, Represents the final neuron distribution data of the LN layer, and They are all reconstruction parameters in the neural network training process, and the reconstruction parameters help restore the characteristic distribution of the original network.

[0031] Preferably, in step five, in each round of federated learning training, traffic flow data is individually sampled for training and stored as a personalized prediction model in each road area edge server. After the final federated learning training is completed, each road area edge server uses the personalized prediction model of the last round as a local prediction model.

[0032] Preferably, in step six, the preset requirement is a prediction error threshold. If the error between the global model prediction result and the local prediction model prediction result is lower than the prediction error threshold, indicating that the model accuracy meets the preset requirement, the personalized prediction model of the last round of training is used as the local prediction model. If the error between the global model prediction result and the local prediction model prediction result is higher than the prediction error threshold, indicating that the model accuracy does not meet the preset requirement, the global model parameters are re-sent for repeated training until the global model meets the preset requirements.

[0033] Compared with the prior art, the present invention provides a distributed traffic flow prediction method based on personalized federated learning, which has the following beneficial effects:

[0034] 1. The present invention establishes a federated learning framework for the purpose of cross-regional collaborative training, uses local data to train the global model, and adopts a multi-task training method. It allows each road area edge server equipped with a local neural network algorithm to maintain partial privatization of the neural network layer while training to obtain a personalized prediction model that conforms to the characteristics of local data, effectively improving the impact of traffic flow data heterogeneity on prediction accuracy and greatly improving traffic flow prediction accuracy.

[0035] 2. The present invention collects traffic flow data by connecting smart devices to road area edge servers equipped with local neural network algorithms. The smart devices include sensors and cameras. The traffic flow data include vehicle volume, vehicle type, vehicle speed and road occupancy rate. Each road area edge server uses local data for federated learning training to obtain an updated local prediction model, and then randomly samples traffic flow data for separate training, and retains the obtained model parameters in the privatized recurrent neural network LN layer. The model parameters are used when training the local prediction model. While controlling the prediction errors of all parties within the preset requirements, the last round of personalized prediction model is used as the local prediction model, which better retains the data characteristics of each road area edge server, can better fit non-human factor errors in the prediction work, and facilitates efficient and accurate planning and management of traffic. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a step diagram of the method of the present invention;

[0037] Figure 2 Schematic diagram showing a comparison of traffic flow prediction examples using a personalized federated learning algorithm and a traditional federated algorithm. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] See also Figure 1-Figure 2 A distributed traffic flow prediction method based on personalized federated learning is implemented through a traffic flow prediction system framework. The traffic flow prediction system framework includes a central cloud server equipped with a federated learning algorithm and a road area edge server equipped with a local neural network algorithm. There are multiple road area edge servers equipped with the local neural network algorithm. The local neural network algorithm realizes that the road area edge server cooperates with the central cloud server to predict traffic flow without sharing traffic flow data, including the following steps:

[0040] Step 1: Establish a federated learning framework for cross-regional collaborative training, use local data to train the global model, and adopt a multi-task training method. Allow each road area edge server equipped with a local neural network algorithm to maintain partial privatization of the neural network layer while training a personalized prediction model that conforms to the characteristics of local data, effectively improve the impact of traffic flow data heterogeneity on prediction accuracy, and greatly improve the prediction accuracy under federated learning training. Use the last round of personalized prediction model as the local prediction model, better retain the data characteristics of the edge servers in each road area, and better fit the non-human factor errors in the prediction work;

[0041] Step 2: Traffic flow data is collected and preprocessed through the road area edge server equipped with a local neural network algorithm, and then the preprocessed traffic flow data is uploaded to the central cloud server;

[0042] The road area edge server equipped with a local neural network algorithm collects traffic flow data through network connection to smart devices, including sensors and cameras. The traffic flow data includes vehicle volume, vehicle type, vehicle speed and road occupancy rate. The road area edge server performs data preprocessing through a Gaussian denoiser to remove data noise in the traffic flow data.

[0043] Step 3: The central cloud server equipped with the federated learning algorithm sends the initial parameters of the global model to the road area edge servers equipped with the local neural network algorithm. Each road area edge server uses local data to perform federated learning training to obtain an updated local prediction model.

[0044] Each road area edge server uses traffic flow data to perform federated learning training to obtain an updated local prediction model. During the training process, the parameter update formula of the local prediction model is as follows:

[0045]

[0046] In the formula, Indicates In the round of federated learning, the edge servers in multiple road areas The local prediction model of the edge server in each road area, Indicates the number of The local prediction model of the edge server in each road area, represents the learning rate, Represents the local gradient function, which performs multiple gradient descents according to the computing power and communication latency of the edge servers in different road areas. Indicates that according to the last round of federated learning The local prediction model of the edge server in each road area is gradient descended to obtain an updated local prediction model. The traffic flow data collected by the edge servers in different road areas are quite different and have high data privacy. The federated learning model training method is flexible and suitable for application scenarios where cross-regional shared training cannot be achieved.

[0047] Step 4: Add a multi-task algorithm to the federated learning process. After completing the local update in step 3, each road area edge server randomly samples traffic flow data for separate training based on the global model issued, and retains the obtained model parameters in the privatized recurrent neural network LN layer. The model parameters are used for local prediction model training;

[0048] The central cloud server receives the local prediction model updated by the edge servers in each road area After that, the next round of global model is obtained through aggregation formula. The edge servers in each road area randomly sample traffic flow data to update the personalized prediction model. The loss function in personalized federated learning is composed of the global model and the personalized prediction model, and its expression is as follows:

[0049]

[0050] In the formula, represents the personalized prediction model trained by the edge server in the road area, represents the global model parameters of the previous round of training. If the road area edge server is in the first round of training, ω represents the initial model sent by the central cloud server. It indicates the influence of the global model on the personalized prediction model. The influence intensity is related to the computing power of the edge server in the road area and the degree of data heterogeneity. If the local prediction model is very different from the global model, then The bigger;

[0051] The calculation formula of the personalized federated learning objective function is as follows:

[0052]

[0053] In the formula, represents the personalized federated learning objective function, represents the dimension of the sampled traffic flow data, Indicates that personalized federated learning is performed based on the loss function and sampled traffic flow data, and the minimum value is retained;

[0054] The road area edge server uses the LN layer in the recurrent neural network as a patch layer to perform personalized processing of federated learning. The LN layer is a normalization processing intermediate layer in the convolutional neural network, connecting all input data channels and normalizing them to a normal distribution mode with a mean of 0 and a variance of 1. During training, any model layer can be kept private. The parameter storage cost of the LN layer is very small. While keeping the parameters of each round private, the convergence ability of the global model can also be taken into account. After the local training phase of federated learning is completed, before the model is uploaded to the cloud, each road area edge server needs to use the gradient descent algorithm to process the local prediction model. The local prediction model parameters updated by the road area edge server have the same characteristics as the data in the LN layer, and the local prediction model parameters are stored in the LN layer for the next round of federated training.

[0055] The recurrent neural network consists of an input layer, an output layer, and a hidden layer, and the inter-layer neurons at different times are connected through a loop structure. The LN layer is placed before the nonlinear layer. When training the recurrent neural network, the road area edge server performs standard normalization processing on all channels corresponding to the image in a batch, and performs whitening preprocessing on the input data of the LN layer. The weighted moving average recorded by the LN layer during the training process is directly used in the next round of federated training. The weighted moving average is the hyperparameter of the recurrent neural network for the input data, which reduces the time of model inference. Through the transformation of these two parameters, standard distribution data with a mean of 0 and a variance of 1 are obtained. The input expression of the preprocessed LN layer is as follows:

[0056]

[0057] In the formula, represents the distribution data of neurons, The minimum value introduced to prevent the variance from being 0 and causing invalid calculations. Indicates Data at a point in time, represents the mean value of the neuron, Represents the variance of the neuron. In the recurrent neural network used in the present invention, the input data and output data of the LN layer are not a simple sequence relationship. The LN layer of the recurrent neural network performs multiple multiplications during the input and output processes. Therefore, in order to prevent the multiplication result from being 0, the nonlinear feature is introduced into the LN layer by setting the activation function to enhance the learning ability of the RNN. Assume that the activation function is , then the final LN layer output formula is as follows:

[0058]

[0059] In the formula, Represents the final neuron distribution data of the LN layer, and Both are reconstruction parameters in the neural network training process. The normalization of the LN layer will affect the feature distribution of the previous layer. Introducing reconstruction parameters can help restore the feature distribution of the original network and ensure that the expression ability of the model will not decrease due to normalization.

[0060] Step 5: After each road area edge server completes the local update, it uploads the new round of local prediction model obtained in step 3 to the central cloud server;

[0061] In each round of federated learning training, traffic flow data is sampled separately for training and stored as a personalized prediction model in each road area edge server. After the final federated learning training is completed, each road area edge server uses the personalized prediction model of the last round as the local prediction model, which greatly improves the prediction accuracy of the model under federated learning training.

[0062] Step 6: After receiving the model parameters uploaded by the edge servers in each road area, the central cloud server uses the federated learning algorithm to aggregate and update to obtain a new global model. If the model accuracy meets the preset requirements, the personalized prediction model of the last round of training is used as the local prediction model. If the model accuracy does not meet the preset requirements, the global model parameters are re-issued and the above steps 1 to 5 are repeated until the global model meets the preset requirements.

[0063] The preset requirement is the prediction error threshold. If the error between the prediction result of the global model and the prediction result of the local prediction model is lower than the prediction error threshold, it means that the model accuracy meets the preset requirement. Then the personalized prediction model of the last round of training is used as the local prediction model. For example, Figure 2The horizontal axis in the middle represents the time period for recording traffic flow, and the vertical axis represents the traffic flow value at the current moment. The personalized prediction model of the last round is used as the local prediction model, which well retains the personalized characteristics of the client, enabling it to customize a proprietary personalized prediction model according to local data characteristics, and control the prediction errors of all parties within a certain range. Compared with the global model, the MAPE of the personalized prediction model is reduced by 9.79% overall. If the error between the prediction results of the global model and the prediction results of the local prediction model is higher than the prediction error threshold, indicating that the model accuracy does not meet the preset requirements, the global model parameters will be re-issued for repeated training until the global model meets the preset requirements, thereby improving the accuracy of traffic flow prediction and facilitating efficient and accurate planning and management of traffic.

[0064] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed traffic flow prediction method based on personalized federated learning, characterized by: The following steps are involved: Step 1: Establish a federated learning framework for cross-regional collaborative training, use local data to train the global model, and adopt a multi-task training method to allow each road area edge server equipped with a local neural network algorithm to maintain partial privatization of the neural network layer while training a personalized prediction model that conforms to the characteristics of local data; Step 2: Traffic flow data is collected and preprocessed through the road area edge server equipped with a local neural network algorithm, and then the preprocessed traffic flow data is uploaded to the central cloud server; Step 3: The central cloud server equipped with the federated learning algorithm sends the initial parameters of the global model to the road area edge servers equipped with the local neural network algorithm. Each road area edge server uses local data to perform federated learning training to obtain an updated local prediction model. Step 4: Add a multi-task algorithm to the federated learning process. After completing the local update in step 3, each road area edge server randomly samples traffic flow data for separate training based on the global model issued, and retains the obtained model parameters in the privatized recurrent neural network LN layer. The model parameters are used for local prediction model training; The central cloud server receives the local prediction model updated by the edge servers in each road area After that, the next round of global model is obtained through aggregation formula. The edge servers in each road area randomly sample traffic flow data to update the personalized prediction model. The loss function in personalized federated learning is composed of the global model and the personalized prediction model, and its expression is as follows: In the formula, represents the personalized prediction model trained by the edge server in the road area, Represents the global model parameters of the previous round of training. If the road area edge server is in the first round of training, represents the initial model sent by the central cloud server. It indicates the influence of the global model on the personalized prediction model. The influence intensity is related to the computing power of the edge server in the road area and the degree of data heterogeneity. If the local prediction model is very different from the global model, then The bigger; The calculation formula of the personalized federated learning objective function is as follows: In the formula, represents the personalized federated learning objective function, represents the dimension of the sampled traffic flow data, Indicates that personalized federated learning is performed based on the loss function and sampled traffic flow data, and the minimum value is retained; The road area edge server uses the LN layer in the recurrent neural network as a patch layer to perform personalized processing of federated learning. The LN layer is a normalization processing intermediate layer in the convolutional neural network, connecting the channels of all input data and normalizing them to a normal distribution mode with a mean of 0 and a variance of 1. The local prediction model parameters updated by the road area edge server have the same characteristics as the data in the LN layer, and the local prediction model parameters are stored in the LN layer for the next round of federated training. The LN layer of the recurrent neural network performs multiple multiplications during the input and output processes. To prevent the multiplication result from being 0, the nonlinear features are introduced into the LN layer by setting the activation function. Assume that the activation function is , then the final LN layer output formula is as follows: In the formula, Represents the final neuron distribution data of the LN layer, and These are reconstruction parameters in the neural network training process. The reconstruction parameters help restore the characteristic distribution of the original network. The recurrent neural network consists of an input layer, an output layer, and a hidden layer, and the inter-layer neurons at different times are connected through a loop structure. The LN layer is placed before the nonlinear layer. When training the recurrent neural network, the road area edge server performs standard normalization processing on all channels corresponding to the image in a batch, and performs whitening preprocessing on the input data of the LN layer. The hyperparameters recorded by the LN layer during the training process are directly applied to the next round of federated training to obtain standard distribution data with a mean of 0 and a variance of 1. The input expression of the preprocessed LN layer is as follows: In the formula, represents the distribution data of neurons, The minimum value introduced to prevent the variance from being 0 and causing invalid calculations. Indicates Data at a point in time, represents the mean value of the neuron, represents the variance of the neuron; Step 5: After each road area edge server completes the local update, it uploads the new round of local prediction model obtained in step 3 to the central cloud server; In each round of federated learning training, traffic flow data is sampled separately for training and stored as a personalized prediction model in each road area edge server. After the final federated learning training is completed, each road area edge server uses the personalized prediction model of the last round as the local prediction model. Step 6: After receiving the model parameters uploaded by the edge servers in each road area, the central cloud server uses the federated learning algorithm to aggregate and update to obtain a new global model. If the model accuracy meets the preset requirements, the personalized prediction model of the last round of training is used as the local prediction model. If the model accuracy does not meet the preset requirements, the global model parameters are re-issued and the above steps 1 to 5 are repeated until the global model meets the preset requirements. The preset requirement is the prediction error threshold. If the error between the prediction results of the global model and the prediction results of the local prediction model is lower than the prediction error threshold, it means that the model accuracy meets the preset requirements. The personalized prediction model of the last round of training is used as the local prediction model. If the error between the prediction results of the global model and the prediction results of the local prediction model is higher than the prediction error threshold, it means that the model accuracy does not meet the preset requirements. The global model parameters are re-sent for repeated training until the global model meets the preset requirements.

2. The distributed traffic flow prediction method based on personalized federated learning according to claim 1 is characterized by: In the step 2, the road area edge server equipped with a local neural network algorithm collects traffic flow data through a network connection to smart devices, the smart devices include sensors and cameras, the traffic flow data includes vehicle volume, vehicle type, vehicle speed and road occupancy rate, and the road area edge server performs data preprocessing through a Gaussian denoiser to remove data noise in the traffic flow data.

3. The distributed traffic flow prediction method based on personalized federated learning according to claim 2 is characterized by: In step 3, each road area edge server uses traffic flow data to perform federated learning training to obtain an updated local prediction model. During the training process, the parameter update formula of the local prediction model is as follows: In the formula, Indicates In the round of federated learning, the edge servers in multiple road areas The local prediction model of the edge server in each road area, Indicates the number of The local prediction model of the edge server in each road area, represents the learning rate, Represents the local gradient function, which performs multiple gradient descents according to the computing power and communication latency of the edge servers in different road areas. Indicates that according to the last round of federated learning The local prediction model of the edge server in each road area is gradient descended to obtain an updated local prediction model.

Images