An efficient access method for power internet federated learning based on NR-U

Abstract

The application relates to an NR-U-based power internet federal learning efficient access method, which comprises utilizing a point coordination function (PCF) and proposing a WiFi and NR-U friendly coexistence mechanism; according to the proposed WiFi and NR-U friendly coexistence mechanism, a federal learning system model running on NR-U is established, the system model comprises one central server and K power internet of things terminals, model training is carried out between the central server and the power terminals in a contention-free period (CFP), and mutual communication is completed through NR-U; according to the established system model, a joint power equipment selection and bandwidth allocation optimization method is designed, efficient model training and low-delay communication under large-scale power equipment access are realized under the guarantee of the performance requirements of WiFi networks and NU-R networks. The large-scale power internet equipment access capability is improved, the intelligent service demand of the power internet is met, low-delay communication and efficient training of the central server and the terminal power internet equipment are realized under the guarantee of the performance of NR-U and WiFi networks.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power Internet of Things, and particularly relates to a power Internet of Things federated learning efficient access method based on NR-U. BACKGROUND

[0002] In recent years, with the rapid development of the ubiquitous power Internet of Things, a large amount of data is generated by the edge devices of the power network. According to the report of Cisco, a company, by 2021, an industrial company will generate 2.4 TB of data per minute. This will lead to the data generated in the power utilization link gradually showing the characteristics of massive data, complex data, complex processing logic, and high calculation frequency. Building a shared enterprise will attract more social capital and various market subjects to participate in the construction and value mining of the energy Internet. The demand and application of external artificial intelligence of State Grid, which integrates power data and other industry data, will also increase a lot. For example, power metering, power supply, and power equipment fault detection can all be intelligently operated by using artificial intelligence technology.

[0003] As a new paradigm of distributed machine learning, federated learning is expected to solve the problems of data islands, privacy protection, and data security in the power Internet of Things, and has great development potential. However, with the popularity of wireless terminal devices such as smart meters, the number of wireless terminal devices managed by the ubiquitous power Internet of Things is increasing. On the one hand, a large number of users may simultaneously apply for access, which may easily cause access overload conflicts. On the other hand, the information transmitted by different devices is heterogeneous, and in the case of limited resources, the communication delay caused by large-scale information transmission is unacceptable in actual power application scenarios. These will cause the performance of the power system to decline or even collapse.

[0004] 5G non-licensed band new radio technology (NR-U) is unanimously considered by the industry and academia as one of the key technologies in future mobile communications. 5G NR-U technology is evolved from Long Term Evolution-Unlicensed (LTE-U) technology, aiming to deploy networks using abundant unlicensed frequency bands, and to improve the communication experience of terminal users and reduce costs by greatly increasing the communication bandwidth. Currently, some NR-U technologies have been included in the 5G standards of 3GPP Release 16, and the remaining technologies will be included in Release 17 and subsequent versions. Therefore, facing the scenario of massive device access, the power Internet of Things can use NR-U technology to improve the device access capability, thereby improving the performance and communication efficiency of federated learning training. SUMMARY

[0005] The purpose of the embodiment of the present application is to provide a power internet federal learning efficient access method based on NR-U, and the WiFi and NU-R coexistence mode can realize more access devices, and ensure the low delay communication and efficient training of the center server and the terminal power internet device under the performance of NR-U and WiFi network.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The embodiment of the present application provides a power internet federal learning efficient access method based on NR-U, comprising the following steps:

[0008] Step 1: using point coordination function (PCF), a WiFi and NR-U friendly coexistence mechanism is proposed. Each periodic access interval is composed of a beacon, a contention-free period (CFP) and a contention period (CP). NR-U transmission works in CFP to complete federal learning training, and CP continues to use the traditional carrier sense multiple access with collision avoidance (CSMA / CA) mechanism to provide normal WiFi user transmission.

[0009] Step 2: according to the mechanism proposed in step 1, a federal learning system model running on NR-U is established, and the system model comprises a center server and K power internet of things terminals. The center server and the power terminal perform model training in CFP, and complete mutual communication through NR-U.

[0010] Step 3: according to the system model established in step 2, a joint power device selection and bandwidth allocation optimization method is designed to ensure the performance requirements of WiFi network and NU-R network, efficient model training and low delay communication under large-scale power device access.

[0011] As preferred, the step 2 is specifically as follows:

[0012] Step 2.1: to implement federal learning in power internet, this step will first construct the loss function model of neural network in federal learning, and use the stochastic gradient descent method to train the data of terminal device, so as to realize the optimization of network model. First, define the set of power terminal devices in power internet as Define the power data set collected by each device k for model training as Where N k represents the sample size of the data set x ik represents the i-th data sample, and y ik represents its corresponding label. Therefore, the loss function model f k (w k ) of terminal device k can be obtained as:

[0013]

[0014] where f(w k ,x ik ,y ik ) represents the loss function in the neural network training process of the terminal device, for different tasks and scenarios, such as linear regression formula, logistic regression formula or support vector machine formula can be used instead of the loss function to realize different neural network training processes. k represents the training model weight value of the terminal device k in the neural network.

[0015] The global loss function of the data set necklace of all power devices can be represented as

[0016]

[0017] where represents the total number of training samples from the selected device. α k ∈{0,1} indicates whether the power device is selected: when α k =1, it means that the user k is selected to upload the local model to the center server; otherwise, α k =0.

[0018] Step 2.2: This step will build the model training process of the proposed federated learning. Terminal device training needs to go through multiple rounds of model updates, and each round of training needs to go through the corresponding stochastic gradient descent to update the weight of the model. Therefore, for any t-th round of training update, each power device calculates its own local model gradient t based on the received global model parameters w In addition, we further define the gradient norm (GNV) as the data importance, which can be represented as Each device needs to send its own to the center server.

[0019] Based on the of each power device, the center server will determine the set of devices S t that need to upload the local model and the resource allocation scheme of the model transmission. And feedback the relevant information to each power device.

[0020] The selected power device will send the local model to the center server according to the relevant feedback. The center server uses the stochastic gradient descent method to aggregate and update the received local model, which can be represented as:

[0021]

[0022] where λ tis the learning rate of the t-th model update. Then, the center server will broadcast the aggregated global model parameters w t+1 to the selected devices.

[0023] After a certain number of the above multi-round training processes, the optimized training weight model w* will be finally obtained.

[0024] Step 2.3: This step will establish a wireless transmission model. It is assumed that in the uplink, the selected devices will transmit the local model to the center server through the orthogonal frequency division multiple access (OFDMA) technology. Therefore, the uplink rate of device k can be expressed as:

[0025]

[0026] where B k represents the bandwidth allocated to device k, and satisfies the total bandwidth does not exceed B, i.e. P k represents the transmit power of device k. g k represents the channel gain between device k and the center server. In addition, N0 represents the noise power of the signal receiving end.

[0027] In the downlink, the center server will occupy all the available bandwidth to broadcast the aggregated updated model to the selected devices, so the downlink rate can be expressed as

[0028]

[0029] where P is the transmit power of the center server.

[0030] The model size in the training process does not change, which can be assumed to be a constant Z. Therefore, the uplink and downlink communication delay between the power device k and the center server can be expressed as:

[0031]

[0032] Since the center server has strong computing power, the time of model aggregation and update can be ignored. In addition, the time of user transmitting constant GNV can also be ignored. Therefore, the transmission delay of global model and local model in one iteration can be expressed as

[0033]

[0034] As a preferred, the step 3 will establish the efficient access optimization target and resource optimization scheme of federated learning in the proposed power Internet of Things based on the system model in step 2. The specific steps are as follows:

[0035] Step 3.1: This step will establish the efficient access optimization goal of federated learning based on NU-R. To achieve efficient access, it is necessary to select as many GNV large ground devices as possible and reduce communication latency, therefore, the objective function can be expressed as:

[0036]

[0037] where α = [α1, α2, …, α K ] T , B = [B1, B2, …, B K ] T . ρ ∈ [0, 1] is the weight factor that controls the trade-off between GNV and transmission delay.

[0038] Step 3.2: To meet the federated learning efficient access goal under the performance requirements of WiFi network and NU-R network, this step establishes an optimization scheme for joint device selection and frequency band allocation based on the training process of step 2.2. The specific steps are as follows:

[0039] Step 3.2.1: In any t-th round of model training update, the center server receives the The resource allocation scheme of each round of model update will be achieved through iterative optimization, therefore, first initialize the iteration number l = 0, and further initialize the auxiliary variables λ l and Set the termination condition of iteration σ l .

[0040] Step 3.2.2: Calculate the optimal device selection scheme The calculation formula is:

[0041]

[0042] where ν is a non-negative penalty factor, μ k , ζ k and φ k are the Lagrange multipliers respectively.

[0043] Step 3.2.3: Calculate the optimal auxiliary variable based on step 3.2.2 The calculation formula is:

[0044]

[0045] Step 3.2.4: Solve the optimal bandwidth allocation scheme It can be obtained by solving the following equation:

[0046]

[0047] wherein κ k and ψ k are Lagrange multipliers, respectively.

[0048] Step 3.2.5: Update the auxiliary variable λ l :

[0049]

[0050] Step 3.2.6: Update the auxiliary variable

[0051]

[0052] Step 3.2.7: Calculate the termination condition of iteration σ l :

[0053] σ (l) = F (l) - F (l-1)

[0054] wherein

[0055] Step 3.3: When the termination condition of iteration satisfies the threshold σ l ≥ σ min or the iteration reaches the maximum number of iterations N max , output the optimal user selection scheme optimal bandwidth allocation scheme

[0056] Compared with the prior art, the beneficial effects of the present application are: based on the background of the intelligent development demand of the power Internet of Things, aiming at the challenge that the training efficiency and communication delay of federated learning cannot be guaranteed, the present application proposes an efficient access method of federated learning under the power Internet of Things based on NR-U. First, facing the large-scale power Internet of Things device access scene, a WiFi and NR-U friendly coexistence mechanism is proposed, which effectively improves the device access capacity of the network to cope with the challenges of access overload conflict, communication delay and the like caused by a large number of power devices. And under this framework, the federated learning utilizes NR-U to transmit the training global model. In addition, the present application also proposes a joint device selection and frequency band allocation optimization scheme. Compared with the existing federated learning method, the optimization scheme jointly considers the learning and communication indicators, and provides strong guarantee for the low-delay communication and efficient federated learning of the power Internet of Things. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 : is the WiFi and NR-U friendly coexistence mechanism of the present application.

[0058] Figure 2Figure 1 is a flow chart of the present application implementing joint device selection and band allocation optimization scheme. DETAILED DESCRIPTION

[0059] Referring to Figure 1 and Figure 2 , the present application provides a high-efficiency access method for power internet federated learning based on NR-U, comprising the following steps:

[0060] Step 1: Use the point coordination function (PCF) to propose a WiFi and NR-U friendly coexistence mechanism. Each periodic access interval is composed of a beacon, a contention-free period (CFP) and a contention period (CP). NR-U transmission works in CFP to complete federated learning training, while CP continues to use the traditional carrier sense multiple access with collision avoidance (CSMA / CA) mechanism to provide normal WiFi user transmission.

[0061] Step 2: According to the mechanism proposed in step 1, a federated learning system model running on NR-U is established, which includes a central server and K power internet of things terminals. The central server and the power terminal perform model training within the CFP, and complete mutual communication through NR-U.

[0062] Step 3: According to the system model established in step 2, a joint power device selection and bandwidth allocation optimization method is designed to ensure the performance requirements of WiFi network and NU-R network, efficient model training and low latency communication under large-scale power device access.

[0063] As a preferred, the step 2 is specifically as follows:

[0064] Step 2.1: In order to implement federated learning in power internet, this step will first construct the loss function model of neural network in federated learning, and use the stochastic gradient descent method to train the data of terminal device, so as to realize the optimization of network model. First, define the set of power terminal devices in power internet as Define the power data set collected by each device k for model training as Where N k represents the sample size of the data set x ik represents the i-th data sample, and y ik represents its corresponding label. Therefore, the loss function model f k (w k ) of terminal device k can be expressed as:

[0065]

[0066] Where f(w k ,xik ,y ik ) represents the loss function in the neural network training process of the terminal device, and different loss functions such as linear regression formula, logistic regression formula or support vector machine formula can be used to replace the loss function to realize different neural network training processes for different tasks and scenes.w k represents the training model weight value of the terminal device k in the neural network.

[0067] The global loss function of the data set necklace of all power devices can be represented as

[0068]

[0069] wherein represents the total number of training samples from the selected device. α k ∈{0,1} indicates whether the power device is selected: when α k =1, it means that the user k is selected to upload the local model to the center server; otherwise, α k =0.

[0070] Step 2.2: This step will build the model training process of the proposed federated learning. The terminal device training needs to go through multiple rounds of model updates, and each round of training needs to go through the corresponding stochastic gradient descent to update the weight of the model. Therefore, for any t-th round of training update, each power device calculates the respective local model gradient t based on the received global model parameters w In addition, we further define the gradient norm (GNV) as the data importance, which can be represented as Each device needs to send its respective to the center server.

[0071] Based on the of each power device, the center server will determine the set of devices S t that need to upload the local model and the resource allocation scheme of the model transmission. And feedback the relevant information to each power device.

[0072] The selected power devices will send the local model to the center server according to the relevant feedback. The center server uses the stochastic gradient descent method to aggregate and update the received local model, which can be represented as:

[0073]

[0074] wherein λ t is the learning rate of the t-th model update. Then, the center server broadcasts the aggregated global model parameters w t+1 to the selected devices.

[0075] After a certain number of the above multi-round training processes, the optimized training weight model w* will finally be obtained.

[0076] Step 2.3: This step will establish a wireless transmission model. It is assumed that in the uplink, the selected device transmits the local model to the central server through the orthogonal frequency division multiple access (OFDMA) technology. Therefore, the uplink rate of device k can be expressed as:

[0077]

[0078] where B k represents the bandwidth allocated to device k, and satisfies that the total bandwidth does not exceed B, i.e. P k represents the transmission power of device k. g k represents the channel gain between device k and the central server. In addition, N0 represents the noise power of the signal receiving end.

[0079] In the downlink, the central server will occupy all the available bandwidth to broadcast the aggregated updated model to the selected devices, so the downlink rate can be expressed as

[0080]

[0081] where P is the transmission power of the central server.

[0082] The model size in the training process does not change, which can be assumed to be a constant Z. Therefore, the uplink and downlink communication delay between the power device k and the central server can be expressed as:

[0083]

[0084] Since the central server has strong computing power, the time for model aggregation and update can be ignored. In addition, the time for users to transmit the constant GNV can also be ignored. Therefore, the transmission delay of the global model and the local model in one iteration can be expressed as

[0085]

[0086] As preferred, the step 3 will establish the efficient access optimization target and resource optimization scheme of federated learning in the proposed power Internet of Things based on the system model in step 2. The specific steps are as follows:

[0087] Step 3.1: This step will establish the efficient access optimization target of NU-R based federated learning. In order to achieve efficient access, it is necessary to select as many devices with large GNV as possible and reduce the communication delay, therefore, the objective function can be expressed as:

[0088]

[0089] where α = [α1, α2,..., α K ] T , B = [B1, B2,..., B K ] T is a weight factor that controls the trade-off between GNV and transmission delay.

[0090] Step 3.2: To meet the goal of efficient access of federated learning under the performance requirements of WiFi network and NU-R network, this step establishes an optimization scheme for joint device selection and frequency band allocation based on the training process of step 2.2. The specific steps are as follows:

[0091] Step 3.2.1: In any t-th round of model training update, the center server receives the The resource allocation scheme of each round of model update will be achieved through iterative optimization, therefore, first initialize the iteration number l = 0, and further initialize the auxiliary variable λ l and Set the termination condition of iteration σ l .

[0092] Step 3.2.2: Calculate the optimal device selection scheme The calculation formula is:

[0093]

[0094] where ν is a non-negative penalty factor, μ k , ζ k and φ k are Lagrange multipliers respectively.

[0095] Step 3.2.3: Calculate the optimal auxiliary variable based on step 3.2.2 The calculation formula is:

[0096]

[0097] Step 3.2.4: Solve the optimal bandwidth allocation scheme It can be obtained by solving the following equation:

[0098]

[0099] where κ k and ψ k are Lagrange multipliers respectively.

[0100] Step 3.2.5: Update the auxiliary variable λ l :

[0101]

[0102] Step 3.2.6: Update the auxiliary variable

[0103]

[0104] Step 3.2.7: Calculate the termination condition σ of iteration l :

[0105] σ (l) = F (l) - F (l-1)

[0106] where

[0107] Step 3.3: When the termination condition of iteration satisfies the threshold σ l ≥ σ min or the iteration reaches the maximum iteration number N max , output the optimal user selection scheme optimal bandwidth allocation scheme

[0108] The above only describes the embodiments of the present application and is not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An efficient access method for power internet federated learning based on NR-U, characterized in that, The method comprises the following steps: Step 1: a WiFi and NR-U friendly coexistence mechanism is proposed by using a point coordination function PCF, each periodic access interval in the point coordination function PCF is composed of a beacon, a contention-free period CFP and a contention period CP, the NR-U transmission works in the contention-free period CFP to complete the federated learning training, and the contention period CP provides normal WiFi user transmission by using a traditional carrier sense multiple access mechanism with collision avoidance; Step 2: according to the WiFi and NR-U friendly coexistence mechanism proposed in step 1, a federated learning system model running on the NR-U is established, the system model comprises a center server and K power Internet of Things terminals, the model training between the center server and the power terminal is carried out in the contention-free period CFP, and mutual communication is completed through the NR-U; Step 3: according to the system model established in step 2, a joint power equipment selection and bandwidth allocation optimization method is designed to realize efficient model training and low-delay communication under the condition of large-scale power equipment access while guaranteeing the performance requirements of the WiFi network and the NU-R network; The federated learning system model running on the NR-U is established as follows: A loss function model of a neural network in the federated learning is constructed, and a random gradient descent method is used to train data of terminal equipment, so as to realize optimization of the network model; A model training process of the proposed federated learning is constructed, terminal equipment training needs to undergo multiple rounds of model updating, each round of training needs to undergo corresponding random gradient descent for model weight updating training, and after a certain number of rounds of training, an optimized training weight model is finally obtained; A wireless transmission model is established, in the uplink, the selected equipment transmits the local model to the center server through an orthogonal frequency division multiple access technology, and in the downlink, the center server occupies all available bandwidth and broadcasts the aggregated updated model to the selected equipment.

2. The NR-U based power internet federation learning efficient access method according to claim 1, characterized in that, In step 3, a joint power equipment selection and bandwidth allocation optimization method is designed as follows: An efficient access optimization target of the federated learning based on the NU-R is established; In order to meet the efficient access target of the federated learning under the performance requirements of the WiFi network and the NU-R network, an optimization scheme of joint device selection and frequency band allocation is established; When the termination condition of iteration meets a threshold value or iteration reaches a maximum iteration number, an optimal user selection scheme and an optimal bandwidth allocation scheme are output.

Images