A 5G base station energy saving method based on deep learning prediction

Abstract

The present invention discloses a 5G base station energy-saving method based on deep learning prediction, and relates to the field of 5G base station energy-saving technology. The method includes: the intelligent network management center collects characteristic data related to 5G transmission in the 5G base station, and performs data cleaning on the collected characteristic data; according to the coverage of the current 5G base station cell, data labels are assigned to the cell data to obtain cell pre-processing data with labels; a 5G base station PRB utilization prediction model is constructed; the pre-processed cell base station data is input into the trained 5G base station PRB utilization prediction model to generate the PRB utilization prediction result of the cell in the energy-saving period; in the energy-saving period, if the PRB utilization prediction result is higher than the deep sleep threshold, the cell is kept in normal operation; if it is lower than the threshold, the cell is put into deep sleep, and a corresponding activation mechanism is configured. The method and system can significantly improve the energy efficiency management of 5G base stations, reduce energy consumption, and achieve green energy-saving goals.

Description

Technical Field

[0001] The present invention relates to the field of 5G base station energy saving technology, and specifically to a 5G base station energy saving method based on deep learning prediction. Background Art

[0002] With the widespread deployment of 5G (fifth generation mobile communication technology) networks, the energy consumption problem of base stations has become increasingly prominent. 5G base station equipment such as baseband processing units (BBU), remote radio units (RRU) and antenna units (AAU) consume a lot of power during operation, especially when traffic is low, the equipment remains in operation, resulting in unnecessary energy waste. Therefore, how to reduce the energy consumption of base stations through effective means while ensuring the quality of network services has become an important challenge in the field of 5G technology.

[0003] Traditional base station energy-saving methods mostly rely on simple shutdown strategies, usually shutting down services in certain cells at night or during specific periods of time to reduce energy consumption. Although this extensive energy-saving method can reduce power consumption to a certain extent, its limitations are becoming increasingly apparent as user demand scenarios become more complex. For example, downtown areas may still have high traffic demands even at night, and at transportation hubs such as train stations and airports, there are also a large number of vehicles running at night, resulting in a significant amount of traffic. Therefore, the traditional "one-size-fits-all" shutdown mode can no longer adapt to these diverse and dynamic scenario requirements, and is likely to result in reduced service quality and poor user experience.

[0004] In order to improve the refined management of energy conservation, it is particularly necessary to analyze the historical traffic data of each site and formulate personalized energy conservation strategies according to real-time needs. However, this method requires a lot of manpower investment. As the number of sites and the amount of data increase, the cost and efficiency of manual analysis are difficult to meet the needs. Operation and maintenance personnel face tremendous pressure when processing a large amount of site data, and the input-output ratio is low, which makes it impossible to achieve efficient operation and maintenance management on a large scale.

[0005] In this context, 5G operators urgently need to introduce artificial intelligence (AI) technology to realize intelligent energy-saving management. AI technology can analyze the historical data and real-time traffic volume of each base station, automatically generate dynamic energy-saving strategies, and achieve "one station, one policy". This method can not only adapt to various complex traffic scenarios, but also reduce dependence on manual operation and maintenance, reduce operating costs, and maintain a balance between user experience and key performance indicators (KPIs) while saving energy.

[0006] However, the existing 5G base station energy-saving technology still has obvious shortcomings: the response to traffic fluctuations is not timely enough, it relies on fixed rules, the activation mechanism is delayed, and it fails to make full use of historical and real-time data for optimized scheduling. This leads to limited energy-saving effects and cannot meet the ever-changing network needs. Therefore, a more intelligent, flexible and efficient energy-saving method is urgently needed to deal with the energy consumption management problems of 5G base stations in changing business scenarios and achieve more accurate energy-saving control. Summary of the invention

[0007] The present invention aims to provide a 5G base station energy-saving method based on deep learning prediction to solve the problems mentioned in the prior art, such as the lack of timely response to traffic fluctuations, reliance on manual operation and maintenance and fixed rules, delayed activation mechanism, and failure to make full use of historical and real-time data for optimized scheduling.

[0008] The present invention adopts the following technical solution: a 5G base station energy saving method based on deep learning prediction, comprising the following steps:

[0009] Step S1: The intelligent network management center collects characteristic data related to 5G transmission in the 5G base station, cleans the collected characteristic data, and obtains pre-processed data.

[0010] Step S2: According to the coverage of the current 5G base station cell, assign data labels to the cell data to obtain cell pre-processing data with labels; combine the historical data of the cell to set the deep sleep threshold and activation threshold, and divide the cell into working period, energy-saving period and sleep period based on the data labels.

[0011] Step S3: Divide the cell pre-processed data with data labels into a training set, a validation set and a test set in chronological order; take the cell pre-processed data with data labels as input, construct a neural network consisting of a feature extraction layer and a feature fusion layer, and train a 5G base station PRB utilization prediction model based on the neural network; after multiple iterations, finally select the model with the best performance on the test set as the optimal model and deploy it to the intelligent network management center.

[0012] Step S4: Input the preprocessed cell base station data into the trained 5G base station PRB utilization prediction model to generate the PRB utilization prediction result of the cell during the energy-saving period.

[0013] Step S5: During the energy-saving period, if the PRB utilization prediction result is higher than the deep sleep threshold, the cell is kept in normal operation; if it is lower than the threshold, the cell is put into deep sleep and a corresponding activation mechanism is configured to respond to the demand in a timely manner.

[0014] Furthermore, the step S1 comprises the following steps:

[0015] Step S11: The intelligent network management center collects cell characteristic data related to BBU, AAU, RRU and transmission in the 5G base station. The characteristic data include: uplink RLC layer user plane PRB utilization, downlink RLC layer user plane PRB utilization, downlink PRB average occupancy, uplink PRB average occupancy, average number of RRC connected users, and RRU model and base station model.

[0016] Step S12: The collected feature data are resampled and sorted at hourly time intervals, and the data is cleaned; all missing data are filled with zero values, and the daily data is strictly divided into hours, for a total of 24 hours. If the data of a node is missing, the mean value of the cell is used to fill it to ensure data integrity and obtain preprocessed data.

[0017] Furthermore, the step S2 includes the following steps:

[0018] Step S21: Assign data labels to each cell data according to the cell location and coverage in the pre-processed data, the labels include subway stations, industrial parks, rural areas, high-speed rail stations, large venues, schools, shopping malls, and office buildings.

[0019] Step S22: Based on the historical data of the cell, a deep sleep threshold and an activation threshold are set for each cell, and the working period, energy-saving period and sleep period of each cell are divided based on the data label.

[0020] Furthermore, the data set division in step S3 includes the following steps: dividing the cell pre-processed data with data labels into a training set, a validation set and a test set in a ratio of 7:2:1 in chronological order.

[0021] Furthermore, the feature extraction layer in step S3 includes the following steps:

[0022] Step S311: Use the training set and validation set data as the input of the neural network, expressed as , Indicates input, represents the set of real numbers, Indicates the total number of time nodes of a single base station sample divided by hour, Represents the length of the feature vector of a single time node, Indicates the number of feature vector channels.

[0023] Step S312: The feature extraction layer includes a transposition layer and three pooling layers, forming a four-layer feature pyramid structure; the transposition layer is used to Transpose to get , as the output of the first layer; the window size of each layer of the three pooling layers is 3, the step size is 2, and the padding is 1. The output of each layer is obtained using the feature pyramid method. ,in , the output of each layer is calculated as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] in, Represents the average pooling operation, the length of the feature vector output by each layer for:

[0029] ;

[0030] in, Indicates filling, the value is 1; Indicates the window size, the value is 3; Indicates the step size, the value is 2.

[0031] Furthermore, the feature fusion layer in step S3 includes the following steps:

[0032] Step S321: The feature fusion layer includes eight linear layers and three concatenation layers. The feature pyramid method is used to obtain four layers of feature pyramid outputs. The outputs of each layer are combined to form multi-scale temporal features. , four linear layers are used to transform the temporal features of the four-layer feature pyramid Mapping to intermediate features , expressed as ;

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] in Represents the linear operation of the linear layer;

[0038] Then by passing the Layer characteristics and The layer features are concatenated in the time dimension through the concatenation layer, and the first The output features are expressed as ;

[0039] when hour, ;

[0040] ;

[0041] when hour, ;

[0042] when hour, ;

[0043] in represents the fused features, represents the feature concatenation operation of the concatenation layer, Represents the output of the first layer.

[0044] Step S322: The output of the first layer The output of the entire module is obtained through a linear layer and transposed back to the input format to finally get the prediction result. :

[0045] ;

[0046] in, represents the transpose operation, express The weight matrix in the linear layer.

[0047] Furthermore, it is characterized in that the training and testing model in step S3 comprises the following steps:

[0048] Step S331: first set the total number of training cycles to 80 and the initial learning rate to 0.01; after each cycle, apply the model to the validation set and calculate its loss value; if it is found that the model performance is no longer improved, halve the learning rate to further optimize the model training.

[0049] Step S332: The root mean square error RMSE between the 5G base station PRB utilization predicted by the model and the actual PRB utilization in the test set is used as the loss function of the model, expressed as:

[0050] ;

[0051] in Represents the prediction results Predict the PRB utilization value for one time step in the future, Indicates the actual PRB utilization value of this time step; express The total number of future time steps to predict.

[0052] The Adam optimizer is used to optimize the network parameters to gradually reduce the loss value.

[0053] Step S333: Save the weight parameter corresponding to the lowest loss function value as the optimal weight parameter obtained during the training process as the optimal model.

[0054] Step S334: Apply the model to each 5G base station in the collection area, and finally obtain the PRB utilization prediction results of each base station in the next 72 time steps through the historical 72 time steps.

[0055] Furthermore, the activation mechanism in step S5 comprises the following steps:

[0056] When the user's 5G access request reaches the activation threshold, the base station will automatically wake up the cell in deep sleep to ensure the normal operation of the network; if the access request does not reach the activation threshold, the cell will continue to remain in deep sleep to save energy.

[0057] The beneficial effects of the present invention are:

[0058] (1) This invention combines deep learning with intelligent energy-saving strategies to generate personalized energy-saving solutions for each base station, realizing "one station, one policy". This method can flexibly adjust the working mode according to the actual load demand of the base station, effectively solving the shortcomings of the traditional "one-size-fits-all" shutdown strategy in complex traffic scenarios.

[0059] (2) The present invention introduces an intelligent activation mechanism. After the base station enters a deep sleep state, it can automatically wake up based on real-time traffic demand and prediction results. When the traffic exceeds the set activation threshold, the base station can quickly resume normal operation to ensure user experience and network performance. At the same time, the intelligent activation mechanism avoids frequent wake-ups and excessive sleep, further improving energy saving.

[0060] (3) By accurately predicting the PRB utilization and load conditions of 5G base stations, the present invention can intelligently adjust the power output of base stations or shut down some equipment during off-peak hours, thereby significantly reducing overall energy consumption. This method not only effectively reduces operating costs, but also reduces carbon emissions, conforming to the global trend and requirements of energy conservation and emission reduction, and has both environmental and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the process of the present invention.

[0062] Figure 2 This is a schematic diagram of the overall network structure of the 5G base station PRB utilization prediction model of the present invention.

[0063] Figure 3 This is a structural schematic diagram of the feature extraction layer of the 5G base station PRB utilization prediction model of the present invention.

[0064] Figure 4 It is a schematic diagram of the final prediction result based on the linear layer of the present invention. DETAILED DESCRIPTION

[0065] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0066] Embodiment: A 5G base station energy-saving technology based on deep learning prediction is implemented in a 5G base station whose coverage area is a teaching building.

[0067] like Figure 1 As shown, the present invention provides a 5G base station energy saving method based on deep learning prediction, the method comprising the following steps:

[0068] Step S1: The intelligent network management center collects characteristic data related to 5G transmission in the 5G base station, cleans the collected characteristic data, and obtains pre-processed data.

[0069] Step S11: The intelligent network management center collects cell characteristic data related to BBU, AAU, RRU and transmission in the 5G base station. The characteristic data include: uplink RLC (Radio Link Control) layer user plane PRB (Physical Resource Block) utilization of the current cell, downlink RLC layer user plane PRB utilization, downlink PRB average occupancy, uplink PRB average occupancy, average RRC (Radio Resource Control) connected users, and RRU model and base station model.

[0070] Step S12: The collected feature data are resampled and sorted at hourly intervals, and data cleaning is performed on them. All missing data are filled with zero values, and the daily data is strictly divided into hours, for a total of 24 hours. If the data of a node is missing, the mean value of the cell is used to fill it in to ensure data integrity and obtain preprocessed data.

[0071] Step S2: According to the coverage of the current 5G base station cell, assign data tags to it to obtain cell pre-processing data with tags. Combined with the historical data of the cell, set the deep sleep threshold and activation threshold, and divide the cell into working period, energy-saving period and sleep period based on the data tags.

[0072] Step S21: Assign data labels to each cell data according to the cell location and coverage in the preprocessed data. The labels include subway stations, industrial parks, rural areas, high-speed rail stations, large venues, schools, shopping malls, office buildings, etc. In this embodiment, the data label is "teaching building".

[0073] Step S22: Based on the historical data of the cell, a deep sleep threshold and an activation threshold are set for each cell, and the working period, energy-saving period and sleep period of each cell are divided based on the data label.

[0074] In this embodiment, the deep sleep threshold is set for each cell as the 5th percentile of the PRB utilization historical data, the activation threshold is set as the 10th percentile of the PRB utilization historical data, and then based on the data label "teaching building", the cell is divided into working hours 08~12, 14~22, energy-saving hours 07~08, 12~14, 22~00, and sleep hours 00~07. During the working hours, the 5G base station must operate normally; during the sleep hours, the base station should be in a deep sleep state; and during the energy-saving period, whether the 5G base station takes energy-saving measures is dynamically determined by the method of the present invention.

[0075] Step S3: Divide the cell pre-processed data with data labels into training set, validation set and test set in chronological order. Take the cell pre-processed data with data labels as input, build a neural network consisting of feature extraction layer and feature fusion layer, and train the 5G base station PRB utilization prediction model based on the neural network. After multiple iterations, the model with the best performance on the test set is finally selected as the optimal model and deployed to the intelligent network management center.

[0076] Furthermore, the cell pre-processed data with data labels are divided into training set, validation set and test set in a ratio of 7:2:1 in chronological order to ensure the continuity and integrity of the training data.

[0077] Step S31: Design a feature extraction layer of a 5G base station PRB utilization prediction model based on a neural network, such as Figure 2 As shown, the specific steps are as follows:

[0078] Step S311: Use the training set and validation set data as the input of the neural network, expressed as , Indicates input, represents the set of real numbers, Indicates the total number of time nodes of a single base station sample divided by hour, Represents the length of the feature vector of a single time node, Indicates the number of feature vector channels. In this example, is 72, is 8, is 1.

[0079] Step S312: The feature extraction layer includes a transposition layer and three pooling layers, forming a four-layer feature pyramid structure. The transposition layer is used to Transpose to get , as the output of the first layer. The kernel size of each layer of the three-layer pooling layer is 3, the stride is 2, and the padding is 1. The present invention uses the feature pyramid method (FPN, Feature pyramid Network), and the output of each layer is ,in , the output of each layer is calculated as follows:

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] in, Represents the average pooling operation, the length of the feature vector output by each layer for:

[0085] ;

[0086] in, Indicates filling, the value is 1; Indicates the window size, the value is 3; Indicates the step size, the value is 2.

[0087] The downsampling operation of the above pooling layer captures a wider range of contextual information, improves the global feature extraction capability of the model, and helps the model focus on key features, ignores irrelevant information, and improves the accuracy of prediction.

[0088] Step S32: Design a feature fusion layer of a 5G base station PRB utilization prediction model based on a neural network. The specific steps are as follows:

[0089] Step S321: The feature fusion layer includes eight linear layers and three concatenation layers. The feature pyramid method of step S312 is used to obtain four layers of feature pyramid output, and each layer output is combined to form a multi-scale temporal feature. ,like Figure 3 As shown, four linear layers are used to transform the temporal features of the four-layer feature pyramid Mapping to intermediate features , expressed as ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] in Represents the linear operation of a linear layer.

[0095] Then by passing the Layer characteristics and The layer features are concatenated in the time dimension through the concatenation layer, and the first The output features are expressed as .

[0096] when hour, ;

[0097] ;

[0098] when hour, ;

[0099] when hour, ;

[0100] in represents the fused features, represents the feature concatenation operation of the concatenation layer, Represents the output of the first layer.

[0101] Step S322: The output of the first layer The output of the entire module is obtained through a linear layer and transposed back to the input format, such as Figure 2 The right half and Figure 4 As shown, the final prediction result is :

[0102] ;

[0103] in, represents the transpose operation, express The weight matrix in the linear layer.

[0104] Step S33: train and test the model. The specific steps are as follows:

[0105] Step S331: First set the total number of training epochs to 80 and the initial learning rate to 0.01. After each epoch, apply the model to the validation set and calculate its loss. If it is found that the model performance is no longer improved, the learning rate is halved to further optimize the model training.

[0106] Step S332: The root mean square error RMSE between the 5G base station PRB utilization predicted by the model and the actual PRB utilization in the test set is used as the loss function of the model, expressed as:

[0107] ;

[0108] in Represents the prediction results Predict the PRB utilization value for one time step in the future, Indicates the actual PRB utilization value of this time step; express The total number of future time steps to predict.

[0109] The Adam optimizer is used to optimize the network parameters to gradually reduce the loss value.

[0110] Step S333: Save the weight parameter corresponding to the lowest loss function value as the optimal weight parameter obtained during the training process as the optimal model.

[0111] Step S334: Apply the model to each 5G base station in the collection area, such as Figure 4 As shown, through the historical 72 time steps, the PRB utilization prediction results of each base station in the future 72 time steps are finally obtained.

[0112] Step S4: Input the preprocessed cell base station data into the trained 5G base station PRB utilization prediction model to generate the PRB utilization prediction result of the cell during the energy-saving period.

[0113] Step S5: During the energy-saving period, if the PRB utilization prediction result is higher than the deep sleep threshold, the cell is kept in normal operation; if it is lower than the threshold, the cell is put into deep sleep and a corresponding activation mechanism is configured to respond to the demand in a timely manner.

[0114] Furthermore, the activation mechanism is as follows Figure 1 As shown in the right half, when the user's 5G access request reaches the activation threshold, for example, if 5G access requests from different users greater than the activation threshold are received in the past 5 minutes, the base station will automatically wake up the cell in deep sleep state to ensure the normal operation of the network; if the access request does not reach the activation threshold, the cell will continue to maintain a deep sleep state to save energy.

[0115] It should be understood that the deep sleep threshold is not limited to the 5th percentile of the above-mentioned PRB utilization historical data, the activation threshold is not limited to the 10th percentile of the above-mentioned PRB utilization historical data, and the energy-saving period division is not limited to the above-mentioned period division, and no limitation is made here.

[0116] The above description is only a typical embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A 5G base station energy saving method based on deep learning prediction, characterized in that: The method comprises the following steps: Step S1: The intelligent network management center collects characteristic data related to 5G transmission in the 5G base station, cleans the collected characteristic data, and obtains pre-processed data; Step S2: According to the coverage of the current 5G base station cell, assign data labels to the cell data to obtain cell pre-processing data with labels; combine the historical data of the cell to set the deep sleep threshold and activation threshold, and divide the cell into working period, energy-saving period and sleep period based on the data labels; Step S3: Divide the cell pre-processed data with data labels into a training set, a validation set, and a test set in chronological order; take the cell pre-processed data with data labels as input, construct a neural network consisting of a feature extraction layer and a feature fusion layer, and train a 5G base station PRB utilization prediction model based on the neural network; after multiple iterations, finally select the model with the best performance on the test set as the optimal model and deploy it to the intelligent network management center; Step S4: input the pre-processed cell base station data into the trained 5G base station PRB utilization prediction model to generate a PRB utilization prediction result for the cell during the energy-saving period; Step S5: During the energy-saving period, if the predicted PRB utilization result is higher than the deep sleep threshold, the cell is kept in normal operation; if it is lower than the threshold, the cell is put into deep sleep and a corresponding activation mechanism is configured to respond to the demand in a timely manner; The feature extraction layer in step S3 includes the following steps: Step S311: Use the training set and validation set data as the input of the neural network, expressed as , Indicates input, represents the set of real numbers, Indicates the total number of time nodes of a single base station sample divided by hour, Represents the length of the feature vector of a single time node, Indicates the number of feature vector channels; Step S312: The feature extraction layer includes a transposition layer and three pooling layers, forming a four-layer feature pyramid structure; the transposition layer is used to Transpose to get , as the output of the first layer; the window size of each layer of the three pooling layers is 3, the step size is 2, and the padding is 1. The output of each layer is obtained using the feature pyramid method. ,in , the output of each layer is calculated as follows: ； ； ； ； in, Represents the average pooling operation, the length of the feature vector output by each layer for: ； in, Indicates filling, the value is 1; Indicates the window size, the value is 3; Indicates the step length, the value is 2; The feature fusion layer in step S3 includes the following steps: Step S321: The feature fusion layer includes eight linear layers and three concatenation layers. The feature pyramid method is used to obtain four layers of feature pyramid outputs. The outputs of each layer are combined to form multi-scale temporal features. , four linear layers are used to transform the temporal features of the four-layer feature pyramid Mapping to intermediate features , expressed as ; ； ； ； ； in Represents the linear operation of the linear layer; Then by passing the Layer characteristics and The layer features are concatenated in the time dimension through the concatenation layer, and the first The output features are expressed as ; when hour, ; ； when hour, ; when hour, ; in represents the fused features, represents the feature concatenation operation of the concatenation layer, represents the output of the first layer; Step S322: The output of the first layer The output of the entire module is obtained through a linear layer and transposed back to the input format to finally get the prediction result. : ； in, represents the transpose operation, express The weight matrix in the linear layer; The training and testing model in step S3 comprises the following steps: Step S331: first set the total number of training cycles to 80 and the initial learning rate to 0.01; after each cycle, apply the model to the validation set and calculate its loss value; if it is found that the model performance is no longer improved, halve the learning rate to further optimize the model training; Step S332: The root mean square error RMSE between the 5G base station PRB utilization predicted by the model and the actual PRB utilization in the test set is used as the loss function of the model, expressed as: ； in Represents the prediction results Predict the PRB utilization value for one time step in the future, Indicates the actual PRB utilization value of this time step; express Predict the total number of future time steps; Use Adam optimizer to optimize the network parameters to gradually reduce the loss value; Step S333: saving the weight parameter corresponding to the lowest loss function value as the optimal weight parameter obtained during the training process as the optimal model; Step S334: Apply the model to each 5G base station in the collection area, and finally obtain the PRB utilization prediction results of each base station in the next 72 time steps through the historical 72 time steps.

2. The 5G base station energy saving method based on deep learning prediction according to claim 1 is characterized in that: The step S1 comprises the following steps: Step S11: The intelligent network management center collects cell characteristic data related to BBU, AAU, RRU and transmission in the 5G base station, and the characteristic data includes: uplink RLC layer user plane PRB utilization rate, downlink RLC layer user plane PRB utilization rate, downlink PRB average occupancy rate, uplink PRB average occupancy rate, average number of RRC connected users, and RRU model and base station model; Step S12: The collected feature data are resampled and sorted at hourly time intervals, and the data is cleaned; all missing data are filled with zero values, and the daily data is strictly divided into hours, for a total of 24 hours. If the data of a node is missing, the mean value of the cell is used to fill it to ensure data integrity and obtain preprocessed data.

3. The 5G base station energy saving method based on deep learning prediction according to claim 1 is characterized in that: The step S2 includes the following steps: Step S21: assigning data labels to each cell data according to the cell location and coverage in the pre-processed data, the labels include subway stations, industrial parks, rural areas, high-speed rail stations, large venues, schools, shopping malls, and office buildings; Step S22: Based on the historical data of the cell, a deep sleep threshold and an activation threshold are set for each cell, and the working period, energy-saving period and sleep period of each cell are divided based on the data label.

4. The 5G base station energy saving method based on deep learning prediction according to claim 1 is characterized in that: The data set division in step S3 includes the following steps: dividing the cell pre-processed data with data labels into a training set, a validation set and a test set in a ratio of 7:2:1 in chronological order.

5. The 5G base station energy saving method based on deep learning prediction according to claim 1 is characterized in that: The activation mechanism in step S5 comprises the following steps: When the user's 5G access request reaches the activation threshold, the base station will automatically wake up the cell in deep sleep to ensure the normal operation of the network; if the access request does not reach the activation threshold, the cell will continue to remain in deep sleep to save energy.

Images