A 5G base station energy consumption optimization method based on federated learning

Abstract

A method for optimizing 5G base station energy consumption based on federated learning. This method applies federated learning to 5G base station energy consumption optimization. By applying federated learning to base station data in different scenarios, accurate energy consumption predictions are achieved for each scenario, enabling better optimization of base station energy consumption. Base station service performance data is collected and clustering methods are used to automatically predict and determine the base station's scenario. By applying federated learning to 5G base station energy consumption optimization, base station participants no longer need to upload data, effectively addressing data communication and privacy issues.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence energy consumption optimization, and specifically to a 5G base station energy consumption optimization method based on federated learning. Background Art

[0002] With the advent of the 5G era, the construction of 5G base stations, as an infrastructure, has developed rapidly. A base station refers to a radio transceiver station that transmits information to and from mobile phone terminals through a mobile communication exchange center within a certain radio coverage area. Compared to 4G base stations, 5G base stations have a higher carrier frequency and consume a lot of power. However, the higher the frequency, the greater the attenuation of electromagnetic waves. This means that the signal coverage radius of 5G base stations is smaller than that of 4G base stations. To achieve continuous coverage in key areas and improve 5G coverage, operators have greatly increased the number of 5G base stations they are building, resulting in a sharp increase in their electricity costs. Therefore, energy consumption optimization (ECO) of 5G base stations has become an urgent issue that operators currently need to address and an important goal of sustainable development.

[0003] Base station energy consumption consists of three main components: the energy consumption of the base station's main communication equipment, the energy consumption of the ambient temperature control equipment, and the energy consumption of other auxiliary equipment. Mobile communication networks exhibit significant tidal effects, exhibiting uneven busy and idle periods in both time and space. Traditionally, energy optimization has primarily been performed by manually setting the times and durations of symbol off, channel off, carrier off, and deep sleep based on the load factor of the base station equipment. This fails to intelligently configure energy-saving plans based on the actual conditions of the base station, making it difficult to achieve energy savings while ensuring user quality. Existing energy optimization methods primarily use prediction methods based on sequence features to predict the energy consumption curve for the next phase. For example, the Autoregressive Integrated Moving Average Model (Arima), the Prophet model, and regression-based prediction methods such as the Long Short-Term Memory (LSTM) and the Gated Recurrent Unit (GRU) can provide relatively accurate predictions of energy consumption time series. The Arima model is characterized by high accuracy in predicting time series data, but requires manual determination of multiple parameters, making it difficult to train in batches. The characteristics of the Prophet model are that the model is too simple and the accuracy is low. The LSTM model is based on a sequence of recurrent neural networks and has three types of gate structures: forget gate, input gate, and output gate. The characteristic of LSTM is that it has good prediction accuracy for long-term time series data, but it has high requirements for hardware performance. The GRU model and LSTM model are both special sequences of recurrent neural networks. Compared with LSTM, GRU optimizes the three gate functions of LSTM, has a smaller number of parameters, a low burden on hardware equipment, fast model convergence, and similar accuracy. Although the above methods have been introduced into base station energy consumption optimization and achieved good performance, there are still a series of problems. (1) Base stations are distributed. If base stations from different regions upload the generated data to the central server in a unified manner, a large amount of transmission cost and storage cost will be generated. (2) Modeling in a unified manner on the central server has high requirements for data privacy. Base station information containing private data can be easily obtained by malicious persons during the upload process. The data stored in the central server may be leaked to other participants. Different base station manufacturers are reluctant to upload their own data for the generation of energy-saving models. (3) The geographical location of different base stations determines the application scenarios of the base stations. The energy consumption in different scenarios varies. If all base stations are trained uniformly, the resulting model will be difficult to converge or unsuitable for different scenarios, and personalized models cannot be generated for different scenarios.

[0004] Federated learning, proposed by Google in 2016, serves as a paradigm for distributed training of machine learning models. Federated learning allows multiple participants to collectively train a model, coordinated by a central server, without directly sharing data. Information exchanged between participants is encrypted, preventing them from accessing each other's content. Participants with similar data (or scenarios) can be combined for personalized federated learning, generating a model tailored to their needs. Therefore, applying federated learning to energy optimization can effectively address current challenges in base station energy conservation, both in terms of modeling and data privacy. Summary of the Invention

[0005] In order to overcome the deficiencies of the above technologies, the present invention provides a method for achieving accurate prediction of energy consumption in different scenarios to better optimize base station energy consumption.

[0006] The technical solution adopted by the present invention to overcome the technical problems is:

[0007] A method for optimizing energy consumption of a 5G base station based on federated learning includes the following steps:

[0008] a) Clustering base station scenarios;

[0009] b) In the base station scenario, the base station trains the base station parameters through the GRU network and uploads them to the central server. The central server performs federal averaging on the parameters and then sends them to the base station.

[0010] c) Fusing the final model parameters with the local parameters to obtain the final model of each base station;

[0011] d) Input the traffic data of base stations under different scenarios into the final model of each base station for learning, and predict the time series of base station energy consumption under different scenarios.

[0012] The method of the present invention first collects base station traffic data and uses the k-means clustering algorithm to automatically predict and determine the scenario to which the base station belongs. Next, a federated learning modeling process is performed within each scenario. Next, the local model is integrated with the global model to adapt the model to the base station's conditions. Finally, an energy-saving algorithm strategy is applied to the base station, which automatically optimizes energy consumption based on the predicted energy consumption model. The following describes the technical details in order.

[0013] Furthermore, step a) includes the following steps:

[0014] a-1) Collect base station traffic to obtain a sample data set D = {x1, x2, ..., x i ,...,x m}, where x iis the traffic data vector of the i-th base station, i∈{1,2,...,m}, m is the number of base stations, and the traffic data vector of each base station in the sample data set D is normalized to obtain the normalized sample data set D′={x′1,x′2,...,x′ i ,...,x′ m};

[0015] a-2) Randomly select k base station traffic data vectors from the normalized sample dataset D′ and name them {μ1,μ2,...,μ i ,...,μ k}, replace {μ1,μ2,...,μ i ,...,μ k} as the cluster center vector of k base station scenes, μ i is the cluster center vector of the randomly selected i-th base station, i∈{1,2,...,k};

[0016] a-3) Through formula d ji =||x′ j -u i ||2Calculate the remaining j-th base station traffic data vector x′ after random selection in step a-2) j and the cluster center vectors {μ1,μ2,...,μ i ,...,μ k} distance d ji , j∈{1,2,...,mk}, select each distance d ji The minimum value of the cluster center vector of the base station scene corresponding to the minimum value is taken as the i-th base station scene, and the minimum value is put into the set C of base station traffic data vectors of the i-th base station scene i where i∈{1,2,...,k};

[0017] a-4) Calculate the set C i The mean of all base station traffic data vectors in is used as the new cluster vector center μ′ i ;

[0018] a-5) Repeat steps a-2) to a-4, and let the cluster vector center μ′ of the tth round i and the cluster vector center μ′ in round t-1 i If it remains unchanged, the base station scene clustering is completed.

[0019] Furthermore, step b) comprises the following steps:

[0020] b-1) The normalized sample data set D′={x′1,x′2,...,x′ i ,...,x′m The normalized base station traffic data vector x′ in} i The energy consumption data is divided into segments of working days according to the number of days. After the division, n energy consumption data samples are generated. The n energy consumption data samples form the energy consumption time series dataset A. Among them, a t is the energy consumption data sample of batch learning input at time t, t∈{1,2,...,n is}, the sth base station in the i-th base station scenario has n is energy consumption data samples, where s∈{1,2,...,S i}, S i is the number of base stations in the i-th base station scenario;

[0021] b-2) Initialize the central server model parameters w0 for each base station scene;

[0022] b-3) The central server of each base station scene sends the model parameter w0 to each base station client C in the base station scene. is C is is the sth base station in the i-th base station scenario;

[0023] b-4) Base station client C participating in federated learning is After receiving the model parameter w0 from the central server, batch local training is performed. The number of base station local update training rounds is set to E rounds, the number of samples of local batch learning of the base station is B, and the number of training times required for each round is times, a total of training times, through the formula w′ is ←w is -η▽l(w is ; b) Update the neural network weights to obtain the updated local parameter w′ is , where w is is the local parameter of the sth base station in the i-base station scenario, η is the learning rate of the neural network, ▽l(·; b) is the gradient value of the objective function with respect to the weight;

[0024] b-5) The energy consumption data sample a of the batch learning input at time t t Input to the GRU neuron structure, through the formula r t =σ(W r a t +U r h t-1 +b r ) Calculate the reset gate r t , where h t-1 is the batch energy consumption data vector output at the moment before time t, W r To reset the gate weight parameters, Ur Reset the weight parameter of the hidden layer gate for the previous moment of time t, b r is the bias term, through the formula z t =σ(W z a t +U z h t-1 +b z ) Calculate the update gate z t , where σ(·) is the sigmoid activation function, W z is the weight parameter of the update gate, U z is the weight parameter of the hidden layer update gate before time t, b z is the bias term, through the formula Calculate the final output result h t , where Where tanh(·) is the tanh activation function, W h is the weight parameter of the candidate hidden layer, b h is the bias term;

[0025] b-6) The base station client of each scenario completes the After training, the local update is considered complete and the locally updated parameter w′ is set is Upload to the central server of the corresponding scene;

[0026] b-7) The central server in the base station scenario uses the formula Calculate the model parameters w′ for the next round i , n i is the total number of energy consumption data samples of the i-th base station scene;

[0027] b-8) Repeat steps b-3) to b-7) until the number of training times set in step b-4) is reached Obtain the final personalized 5G base station federation energy consumption optimization model parameter w′ suitable for each scenario i .

[0028] Furthermore, in step b-4), η=0.01.

[0029] Furthermore, in step c), the formula Calculate the final model of each base station Where W conv (·) is the convolution operation.

[0030] Furthermore, in step c), the formula Calculate the final model of each base station Where F multi(·) is the multiplication fusion operation, and λ is the weight used to balance the final model parameters and the local update parameters.

[0031] Furthermore, in step c), the formula Calculate the final model of each base station Where F add (·) is the additive fusion operation, and λ is the weight used to balance the final model parameters and the local update parameters.

[0032] The beneficial effects of the present invention are: a method for applying federated learning to optimize the energy consumption of 5G base stations. By performing federated learning on the data of base stations in different scenarios, accurate prediction of energy consumption in different scenarios is achieved to better optimize the energy consumption of base stations. The energy consumption optimization method proposed in the present invention collects base station service performance data and uses a clustering method to automatically predict and determine the scenario in which the base station is located. By applying federated learning to the energy consumption optimization of 5G base stations, base station participants do not need to upload data, which effectively solves problems in data communication and data privacy. The energy consumption optimization method proposed in the present invention performs federal modeling on the base stations in each scenario, which effectively solves the problem of personalization of the federated learning energy consumption optimization model. The energy consumption optimization method proposed in the present invention forms a model that is more in line with each base station by fusing the local model and the global model of the base station. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present invention;

[0034] Figure 2 This is the GRU neuron structure diagram of the present invention. DETAILED DESCRIPTION

[0035] The following is combined with Figure 1 , Attachment Figure 2 The present invention is further described.

[0036] A method for optimizing energy consumption of a 5G base station based on federated learning includes the following steps:

[0037] a) Clustering of base station scenarios.

[0038] b) In the base station scenario, the base station parameters are trained through the GRU network and uploaded to the central server. The central server performs federal averaging on the parameters and then sends them to the base station.

[0039] c) Fuse the final model parameters with the local parameters to obtain the final model of each base station. This model is called the FedECO model. Figure 1As shown in the figure, ① the scene center server initializes the global model, ② the scene center server sends the global model to all base stations in the scene, ③ the base stations use local data to train the neural network model, ④ the base stations upload the locally trained weights to the scene center server, ⑤ the scene center server aggregates the received weights and calculates the new global model, and repeats the above steps ② to ⑤ until the model converges or reaches the predetermined number of training rounds.

[0040] d) The base station traffic data under different scenarios is input into the final model of each base station for learning, and the time series of base station energy consumption under different scenarios are predicted. Energy-saving strategies such as symbol shutdown and carrier shutdown are then used to reduce energy consumption while ensuring user service quality and privacy.

[0041] Example 1:

[0042] Step a) comprises the following steps:

[0043] a-1) Collect base station traffic to obtain a sample data set D = {x1, x2, ..., x i ,...,x m}, where x i is the traffic data vector of the i-th base station, i∈{1,2,...,m}, m is the number of base stations, and the traffic data vector of each base station in the sample data set D is normalized to obtain the normalized sample data set D′={x′1,x′2,...,x′ i ,...,x′ m}.

[0044] a-2) Randomly select k base station traffic data vectors from the normalized sample dataset D′ and name them {μ1,μ2,...,μ i ,...,μ k}, replace {μ1,μ2,...,μ i ,...,μ k} as the cluster center vector of k base station scenes, μ i is the cluster center vector of the randomly selected i-th base station, i∈{1,2,...,k}.

[0045] a-3) Through formula d ji =||x′ j -u i ||2Calculate the remaining j-th base station traffic data vector x′ after random selection in step a-2) j and the cluster center vectors {μ1,μ2,...,μ i ,...,μ k} distance d ji, j∈{1,2,...,mk}, select each distance d ji The minimum value of the cluster center vector of the base station scene corresponding to the minimum value is taken as the i-th base station scene, and the minimum value is put into the set C of base station traffic data vectors of the i-th base station scene i where i∈{1,2,...,k}.

[0046] a-4) Calculate the set C i The mean of all base station traffic data vectors in is used as the new cluster vector center μ′ i .

[0047] a-5) Repeat steps a-2) to a-4, and let the cluster vector center μ′ of the tth round i and the cluster vector center μ′ in round t-1 i If it remains unchanged, the base station scene clustering is completed.

[0048] Example 2:

[0049] Step b) comprises the following steps:

[0050] b-1) The energy consumption prediction of 5G base stations is a time series prediction problem. The present invention uses the GRU neuron structure as a global and local neural network learner. There is a reset gate r in the GRU structure. t and update gate z t Two gating units. The former is used to control the extent to which the state of the previous hidden layer can be updated to the state of the current candidate hidden layer, and determines the amount of past information to be forgotten. The latter is used to control the extent to which the state of the previous hidden layer can be updated to the state of the current hidden layer, and determines which information to discard and which new information to add. Both are used to forget or remember the information generated by the previous training steps. Specifically, the normalized sample data set D′={x′1,x′2,...,x′ i ,...,x′ m The normalized base station traffic data vector x′ in} i The energy consumption data is divided into segments of working days according to the number of days. After the division, n energy consumption data samples are generated. The n energy consumption data samples form the energy consumption time series dataset A. Among them, a t is the energy consumption data sample of batch learning input at time t, t∈{1,2,...,n is}, the sth base station in the i-th base station scenario has n is energy consumption data samples, where s∈{1,2,...,S i}, S i is the number of base stations in the i-th base station scenario.

[0051] b-2) Initialize the central server model parameters w0 for each base station scenario.

[0052] b-3) The central server of each base station scene sends the model parameter w0 to each base station client C in the base station scene. is C is is the sth base station in the i-th base station scenario.

[0053] b-4) Training base station client local update function ClientUpdate(C is , w is ). Where C is is the sth base station of scenario i in step b-3. Base stations C participating in federated learning is After receiving the parameters from the central server, batch local training (Batch-size Training) is performed. Specifically, the base station client C participating in the federated learning is After receiving the model parameter w0 from the central server, batch local training is performed. The number of base station local update training rounds is set to E rounds, the number of samples of local batch learning of the base station is B, and the number of training times required for each round is times, a total of training times, through the formula w′ is ←w is -η▽l(w is ; b) Update the neural network weights to obtain the updated local parameter w′ is , where w is is the local parameter of the sth base station in the i-base station scenario, η is the learning rate of the neural network, and ▽l(·; b) is the gradient value of the objective function with respect to the weight.

[0054] b-5) The invention uses GRU neuron structure as global and local neural network learner. During the training process, the energy consumption data sample a of batch learning input at time t is t Input to the GRU neuron structure, through the formula r t =σ(W r a t +U r h t-1 +b r ) Calculate the reset gate r t , where h t-1 is the batch energy consumption data vector output at the moment before time t, W r To reset the gate weight parameters, U r Reset the weight parameter of the hidden layer gate for the previous moment of time t, b r is the bias term, through the formula z t =σ(W z at +U z h t-1 +b z ) Calculate the update gate z t , where σ(·) is the sigmoid activation function, W z is the weight parameter of the update gate, U z is the weight parameter of the hidden layer update gate before time t, b z is the bias term, through the formula Calculate the final output result h t , where Where tanh(·) is the tanh activation function, W h is the weight parameter of the candidate hidden layer, b h is the bias term. The sigmoid activation function is used to scale the value to between (0, 1), and the tanh activation function is used to scale the value to between [-1, 1].

[0055] b-6) The base station client of each scenario completes the After training, the local update is considered complete and the locally updated parameter w′ is set is Upload to the central server of the corresponding scene.

[0056] b-7) The central server in the base station scenario uses the formula Calculate the model parameters w′ for the next round i , n i is the total number of energy consumption data samples of the i-th base station scenario.

[0057] b-8) Repeat steps b-3) to b-7) until the number of training times set in step b-4) is reached Obtain the final personalized 5G base station federation energy consumption optimization model parameter w′ suitable for each scenario i .

[0058] Preferably, in step b-4), η=0.01.

[0059] Example 3:

[0060] The base station uploads the local update parameter w′ is When the data is sent to the central server, this value will be saved. The final model parameter w′ obtained in step b-8) i With the parameter w′ is Perform fusion to obtain the final model of each base station It can adapt to the situation of each base station in each scene. Therefore, the present invention designs three feature fusion methods: convolution fusion F conv (w is ,w′ i), multiplication fusion F multi (w is ,w′ i ), additive fusion F add (w is ,w′ i ). Convolutional fusion arranges local and global parameters into a parameter matrix and connects them, extracts them using 1*1 convolution, and adjusts the number of channels to be consistent with the global parameters. Multiplication fusion uses a learnable weight vector to multiply the parameters point by point and then add them. Additive fusion uses a learnable scalar vector, multiplies them, and then adds them. Experiments have shown that convolutional fusion works best. The formulas corresponding to the three feature fusion methods are:

[0061] (1) Through the formula Calculate the final model of each base station Where W conv (·) is the convolution operation.

[0062] (2) Through the formula Calculate the final model of each base station Where F multi (·) is the multiplication fusion operation, and λ is the weight used to balance the final model parameters and the local update parameters.

[0063] (3) Through the formula Calculate the final model of each base station Where F add (·) is the additive fusion operation, and λ is the weight used to balance the final model parameters and the local update parameters.

[0064] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A 5G base station energy consumption optimization method based on federated learning, characterized in that: The steps include: a) Clustering base station scenarios; b) In the base station scenario, the base station internally trains the base station parameters through the GRU network and uploads them to the central server. The central server then performs federal averaging on the parameters and sends them to the base station. c) Fusing the final model parameters with the local parameters to obtain the final model of each base station; d) Input the traffic data of base stations under different scenarios into the final model of each base station for learning, and predict the time series of base station energy consumption under different scenarios; Step b) comprises the following steps: b-1) The normalized sample data set D′={x′1,x′2,…,x′ i ,...,x′ m The normalized base station traffic data vector x′ in} i The energy consumption data is divided into segments of working days according to the number of days. After the division, n energy consumption data samples are generated. The n energy consumption data samples form the energy consumption time series dataset A. Among them, a t is the energy consumption data sample of batch learning input at time t, t∈{1,2,...,n is }, the sth base station in the i-th base station scenario has n is energy consumption data samples, where s∈{1,2,...,S i }, S i is the number of base stations in the i-th base station scenario; b-2) Initialize the central server model parameters w0 for each base station scene; b-3) The central server of each base station scene sends the model parameter w0 to each base station client C in the base station scene. is C is is the sth base station in the i-th base station scenario; b-4) Base station client C participating in federated learning is After receiving the model parameter w0 from the central server, batch local training is performed. The number of base station local update training rounds is set to E rounds, the number of samples of local batch learning of the base station is B, and the number of training times required for each round is times, a total of training times, through the formula Update the neural network weights to obtain the updated local parameter w′ is , where w is is the local parameter of the sth base station in the i-base station scenario, η is the learning rate of the neural network, is the gradient of the objective function with respect to the weight; b-5) The energy consumption data sample a of the batch learning input at time t t Input to the GRU neuron structure, through the formula r t =σ(W r a t +U r h t-1 +b r ) Calculate the reset gate r t , where h t-1 is the batch energy consumption data vector output at the moment before time t, W r To reset the gate weight parameters, U r Reset the weight parameter of the hidden layer gate for the previous moment of time t, b r is the bias term, through the formula z t =σ(W z a t +U z h t-1 +b z ) Calculate the update gate z t , where σ(·) is the sigmoid activation function, W z is the weight parameter of the update gate, U z is the weight parameter of the hidden layer update gate before time t, b z is the bias term, through the formula Calculate the final output result h t , where Where tanh(·) is the tanh activation function, W h is the weight parameter of the candidate hidden layer, b h is the bias term; b-6) The base station client of each scenario completes the After training, the local update is considered complete and the locally updated parameter w′ is set is Upload to the central server of the corresponding scene; b-7) The central server in the base station scenario uses the formula Calculate the model parameters w′ for the next round i , n i is the total number of energy consumption data samples of the i-th base station scene; b-8) Repeat steps b-3) to b-7) until the number of training times set in step b-4) is reached Obtain the final personalized 5G base station federation energy consumption optimization model parameter w′ suitable for each scenario i .

2. The 5G base station energy consumption optimization method based on federated learning according to claim 1 is characterized in that: Step a) comprises the following steps: a-1) Collect base station traffic to obtain a sample data set D = {x1, x2, ..., x i ,...,x m }, where x i is the traffic data vector of the i-th base station, i∈{1,2,...,m}, m is the number of base stations, and the traffic data vector of each base station in the sample data set D is normalized to obtain the normalized sample data set D′={x′1,x′2,...,x′ i ,...,x′ m }; a-2) Randomly select k base station traffic data vectors from the normalized sample dataset D′ and name them {μ1,μ2,…,μ i ,…,μ k }, {μ1,μ2,…,μ i ,…,μ k } as the cluster center vector of k base station scenes, μ i is the cluster center vector of the randomly selected i-th base station, i∈{1,2,…,k}; a-3) Through formula d ji =||x′ j -u i ||2Calculate the remaining j-th base station traffic data vector x′ after random selection in step a-2) j and the cluster center vectors {μ1,μ2,…,μ i ,…,μ k } distance d ji , j∈{1,2,...,mk}, select each distance d ji The minimum value of the cluster center vector of the base station scene corresponding to the minimum value is taken as the i-th base station scene, and the minimum value is put into the set C of the base station traffic data vector of the i-th base station scene i where i∈{1,2,...,k}; a-4) Calculate the set C i The mean of all base station traffic data vectors in is used as the new cluster vector center μ′ i ; a-5) Repeat steps a-2) to a-4, and let the cluster vector center μ′ of the tth round i and the cluster vector center μ′ in round t-1 i If it remains unchanged, the base station scene clustering is completed.

3. The 5G base station energy consumption optimization method based on federated learning according to claim 1 is characterized in that: In step b-4), η=0.

01.

4. The 5G base station energy consumption optimization method based on federated learning according to claim 1 is characterized in that: In step c), the formula Calculate the final model of each base station Where W conv (·) is the convolution operation.

5. The 5G base station energy consumption optimization method based on federated learning according to claim 1 is characterized in that: In step c), the formula Calculate the final model of each base station Where F multi (·) is the multiplication fusion operation, and λ is the weight used to balance the final model parameters and the local update parameters.

6. The 5G base station energy consumption optimization method based on federated learning according to claim 1, characterized in that: In step c), the formula Calculate the final model of each base station Where F add (·) is the additive fusion operation, and λ is the weight used to balance the final model parameters and the local update parameters.

Images