Base station load prediction method and device combining deep learning and statistical method, and electronic equipment

Abstract

The invention discloses a base station load prediction method and device combining deep learning and a statistics method, and electronic equipment, and belongs to the field of mobile communication. The method comprises the following steps: defining a load as the number of users entering a coverage range of a base station, and decomposing historical load data Y (t) of the base station into two parts: a part Y1 (t) which follows a long-term trend and changes along with time according to a specific rule, and a load change Y2 (t) caused by short-term random fluctuation; accurate prediction is carried out according to the inherent rule of Y1 (t) changing along with time, and an LSTM network is trained to complete prediction; according to the fact that Y2 (t) has high burstiness and real-time performance, a statistical method is used for prediction; and finally, predicting the load at a plurality of moments in the future, and performing base station dormancy control according to the load. According to the method, the problem of how to minimize the error of load prediction in a high mobility scene is solved, the loads at multiple moments in the future are predicted through the loads of the base station in a past period of time and at the current moment, and a base station dormancy control strategy is formulated.

Description

Technical Field

[0001] The present invention belongs to the field of mobile communications, and more specifically, relates to a base station load prediction method, device and electronic equipment combining deep learning and statistical methods. Background Art

[0002] In recent years, mobile communication networks have been densely deployed with a large number of base stations to improve performance. In order to ensure communication quality, base station deployment is usually planned based on the number of users and the peak of business demand, with a certain degree of redundancy. However, the distribution of users has a tidal effect, that is, it changes significantly over time. If all base stations remain in working condition at all times, a large number of base stations will have low utilization and serious energy waste.

[0003] To meet this challenge, base station sleep control has become a key strategy to reduce energy consumption. When there are fewer users, some base stations can enter a sleep state, and users within their coverage area switch to neighboring base stations. With the increase in high-mobility terminals (such as vehicles and drones), these terminals will pass through the coverage of multiple base stations in a short period of time and perform multiple base station switches, causing the base station load to change rapidly in a short period of time. This makes accurate load prediction essential for effective base station sleep control.

[0004] In existing research, base station load prediction is usually modeled as a time series prediction problem. However, in high mobility scenarios, the load of high-frequency mobile communication networks follows both long-term laws (tidal effects) and short-term fluctuations. Existing solutions do not take both into account. Summary of the invention

[0005] In response to the need for improvement in the prior art, the present invention provides a base station load prediction method, device and electronic device that combine deep learning and statistical methods, thereby solving the problem of base station load prediction in high mobility scenarios. In addition, the predicted load data can be used to solve the problem of intelligent base station sleep control strategy.

[0006] To achieve the above object, according to a first aspect of the present invention, a base station load prediction method combining deep learning and statistical methods is provided, comprising:

[0007] S1, according to the definition of base station load, the base station load historical data Y(t) is divided into two parts: Y1(t) that follows the long-term trend and changes according to a specific law over time, and Y2(t) that fluctuates randomly in the short term; Y1(t) is the responsible data obtained by averaging the past load data in a sliding window, and Y2(t) is the difference between Y(t) and Y1(t);

[0008] Among them, the communication system is a multi-user and multi-base station mobile communication system. The base station load is defined as the number of users within the coverage area of ​​each base station. The load of the base station in the high-frequency band mobile communication network is affected by two factors: the load changes over a long time dimension caused by the tidal effect of user distribution, and the short-term fluctuations caused by the random movement of users.

[0009] S2, according to the inherent law of Y1(t) changing over time, a long short-term memory neural network (LSTM) is trained to obtain the trained LSTM neural network to complete the first load forecast and obtain the first load forecast data;

[0010] S3, according to the strong burstiness and real-time nature of Y2(t), a statistical method is used to perform a second load prediction to obtain second load prediction data;

[0011] S4, since the frequency of base station sleep control update is much lower than the frequency of load prediction update, the base station needs to predict the load at multiple moments in the future. According to the first load prediction data and the second load prediction data at multiple moments in the future, the base station sleep control problem is modeled as an MDP and the MDP is solved, and the base station sleep control is performed accordingly.

[0012] Preferably, step S2 comprises:

[0013] T1. Determine the structure and initial parameters of the LSTM neural network;

[0014] T2, normalize Y1(t) and divide it into training data and test data for training and testing respectively;

[0015] T3. Update the weights of the LSTM neural network by minimizing the loss function that represents the difference between the output of the LSTM neural network and the actual data. Use training data and test data to complete training and testing, respectively, to obtain the first load prediction data.

[0016] Preferably, the LSTM neural network structure is as follows: 1) the input of the input layer is the sequence within the current moment and the previous sliding window; 2) after the input layer is a fully connected layer with an activation function of ReLu, which is used for feature extraction; 3) further followed by several recursive layers based on the LSTM structure, which are used to analyze and identify the correlation within the input sequence; 4) the output layer is a softmax function, in which the output value of each neuron represents the probability that the load will be at the value corresponding to the neuron at the next moment.

[0017] Preferably, the step S3 comprises:

[0018] R1, use kernel density estimation to obtain short-term prediction value y1 and long-term prediction value y2 for Y2(t);

[0019] R2, use the weighted moving average method to obtain the predicted value y3 for Y2(t);

[0020] R3. Since the accuracy of different prediction values ​​varies with the measurement period, the dynamic weight is used to combine the above three prediction values ​​to obtain the final prediction value, that is, the second load prediction data.

[0021] Preferably, the step R3:

[0022] y=αy1+βy2+γy3

[0023] Among them, y is the final prediction value, α, β and γ are dynamic weights, α+β+γ=1 and are all non-negative values.

[0024] According to a second aspect of the present invention, a base station load prediction device combining deep learning and statistical methods is provided, comprising:

[0025] The base station load data acquisition module obtains the load of the base station in the past period and at the current moment according to the number of users entering the base station coverage area as the base station load history data;

[0026] The data preprocessing module is used to divide the collected base station load historical data into two parts: Y1(t), which follows the long-term trend and changes according to a specific law over time, and Y2(t), which is a short-term random fluctuation, and to perform preprocessing operations such as cleaning and normalization on the original data;

[0027] The load prediction module consists of two sub-modules:

[0028] The LSTM neural network prediction submodule is used to train the LSTM neural network according to Y1(t), obtain the trained LSTM neural network to complete the first load prediction, and obtain the first load prediction data to capture and predict the long-term trend part of the load.

[0029] The statistical method prediction submodule is used to perform a second load prediction based on Y2(t) using an appropriate statistical method (such as exponential weighted moving average) to obtain second load prediction data, and to model and predict the short-term random fluctuation part of the load.

[0030] Multi-step prediction integration module: Since the frequency of base station sleep control update is much lower than the frequency of load prediction update, this module is used to integrate the first load prediction data and the second load prediction data at multiple future moments to generate load prediction values ​​at multiple future moments.

[0031] The base station sleep control module is used to model the base station sleep control problem as an MDP and solve it based on the integrated load prediction data, and perform base station sleep control accordingly.

[0032] According to a third aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in the first aspect when executing the computer program.

[0033] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0034] 1. The base station load prediction method based on deep learning proposed in this application analyzes the impact of different factors on the long-term change trend and short-term fluctuation of the load, divides the load into two parts, and uses the LSTM neural network to predict the load changes in a longer time dimension caused by the tidal effect of user distribution. The accuracy is better than other methods.

[0035] 2. The statistical-based base station load prediction method proposed in this application uses kernel density estimation to obtain the short-term prediction value y1 and the long-term prediction value y2 according to the short-term fluctuations caused by random movement of users, and uses the weighted moving average method to obtain the prediction value y3. By combining these three prediction values ​​through dynamic weights, the prediction error can be effectively reduced to achieve high-precision load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of a scenario provided by an embodiment of the present invention;

[0037] Figure 2 A system block diagram provided for an embodiment of the present invention;

[0038] Figure 3 A flowchart of a base station load prediction method combining deep learning and statistical methods provided in an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0041] In existing research, base station load prediction is usually modeled as a time series prediction problem. Prediction methods based on ARIMA models and Markov models focus on predicting load changes caused by fluctuations in the short term, while prediction methods based on models such as LSTM focus on learning the long-term trend of load changes. In high-mobility scenarios, the load of high-frequency mobile communication networks follows both long-term laws (tidal effects) and is affected by short-term fluctuations.

[0042] Based on this, an embodiment of the present invention provides a base station load prediction method that combines deep learning and statistical methods, which is used for the base station load prediction end and can solve the problem of difficult load prediction in high mobility scenarios and minimize load prediction errors.

[0043] The communication system model is a multi-user and multi-base station mobile communication system, such as Figure 1 As shown in the figure, in a target area, multiple high-band base stations jointly provide services to multiple users. Each user may be within the coverage of one or more base stations and connect to one of the base stations in working state according to certain criteria (such as: the highest received signal strength). There is a network controller in the area responsible for collecting various information in the network and making decisions on base station sleep.

[0044] The load prediction scheme aims to predict the load at the next moment based on the load of the base station in the past period of time and the current moment. In order to reflect the utilization rate of the base station in the working state, the load defined here refers to the number of users entering the coverage area of ​​the base station. The proposed prediction model can be expressed as follows:

[0045] Y(t+Δ)=f(Y(t),Y(t-Δ),Y(t-2Δ),...)

[0046] Where Y(t) represents the load of the base station at time t, Δ is the time interval between two consecutive data records, and f(·) is the function that maps the past and current loads to the load at the next moment. Figure 2 As shown in Figure 1, the load of high-frequency mobile communication network base stations is affected by two factors: the load changes over a long time dimension caused by the tidal effect of user distribution, and the short-term fluctuations caused by random user movement. Therefore, Y(t) can be decomposed into two parts: Y1(t) is the part where the load follows a long-term trend and changes according to a specific law over time, and Y2(t) is the change in load caused by short-term random fluctuations.

[0047] The flow chart of the base station load prediction method combining deep learning and statistical methods proposed in the embodiment of the present invention is as follows: Figure 3 As shown, the steps S1-S4 included in the load prediction method are described in detail below:

[0048] S1, according to the definition of base station load, the base station load historical data Y(t) is divided into two parts: Y1(t) that follows the long-term trend and changes according to a specific law over time, and Y2(t) that fluctuates randomly in the short term;

[0049] Specifically, the communication system is a multi-user and multi-base station mobile communication system. The base station load is defined as the number of users within the coverage area of ​​each base station. The load of the high-frequency band mobile communication network base station is affected by two factors: the load changes over a longer time dimension caused by the tidal effect of user distribution, and the short-term fluctuations caused by the random movement of users.

[0050] In order to extract Y1(t) and Y2(t) from Y(t), the data of Y(t) can be filtered by averaging to achieve the effect of "denoising". The "smoothed" data obtained after filtering will become the Y1(t) data, and the remaining part will become the Y2(t) data. Specifically, Y1(t) and Y2(t) are obtained in the following way:

[0051]

[0052] Y2(t)=Y(t)-Y1(t)

[0053] Y1(t) is the data obtained by averaging the past data in a sliding window, which can reduce the disturbance of short-term emergencies (such as vehicles and drones quickly passing through the coverage area of ​​the base station) to the data and highlight the inherent laws of the data in the time domain. The remaining part Y2(t) corresponds to the fluctuation of the load in a short period of time (which can be a negative value).

[0054] S2, according to the inherent law of Y1(t) changing over time, use historical data to train an LSTM neural network to complete the first load prediction and obtain the first load prediction data.

[0055] The load prediction model is a sequential model, and the predicted value depends only on historical load data. According to the base station load prediction model, the following LSTM neural network is constructed: 1) The input layer is the sequence of the current moment and the previous sliding window; 2) After the input layer is a fully connected layer with an activation function of ReLu, which is used for feature extraction; 3) After that are several recursive layers based on the LSTM structure, which are used to analyze and identify the correlation within the input sequence; 4) The output layer is a softmax function, in which the output value of each neuron represents the probability that the load at the next moment is in the value (or range) corresponding to the neuron.

[0056] The training process of the LSTM neural network involves hyperparameter tuning which is performed offline. Before the data is fed into the neural network, it is normalized as follows:

[0057]

[0058] in is the sample average load of the input data. The normalized data is then split into 80% for training and 20% for testing.

[0059] The training of LSTM neural network is achieved by back propagation through time, that is, updating the weights of LSTM neural network by minimizing the loss function L that represents the difference between the output of the neural network and the actual data. Finally, the loss of root mean square error (RMSE) is monitored, and training is stopped when L cannot be further reduced.

[0060] After the training is completed, the trained LSTM network is used for online prediction.

[0061] S3, according to the strong burstiness and real-time nature of Y2(t), a statistical method is used to predict the second load prediction data to obtain the second load prediction data.

[0062] Specifically, a hybrid statistical approach to modeling was used, which combined three predictors.

[0063] First, the kernel density estimation (KDE) method is used to approximate its probability density function (PDF). This method assumes that the load distribution shows a centralized trend around the mean, that is, the closer the load is to its mean, the higher the corresponding probability density; conversely, when it is far away from the mean, the density gradually decreases. On this basis, a sampling value prediction method based on the kernel probability density function is proposed to dynamically capture the time-varying characteristics of the base station load. However, since the probability density function of the base station load will be updated in real time with the continuous influx of new measurement data, it makes it challenging to accurately characterize the real load distribution at a specific time point. Therefore, in order to improve the prediction accuracy and the adaptability of the model, it is necessary to evaluate the performance of different kernel functions to select the optimal solution. After systematic research and comparative analysis, the results show that the use of the Epanechnikov kernel function for load measurement has significant advantages over other methods. Specifically, for the probability density value of a given load x, the following formula is used to calculate:

[0064]

[0065] where x i is the ith data point, n is the number of samples, h is the bandwidth parameter, and K is the Epanechnikov kernel function. Using the kernel function, we get the probability density function with x as a continuous variable: This function depends on n and the choice of sample set (i.e. x1, x2, ..., x n). In order to capture the trend of load variation over time, a sliding window technique is used to extract data from the measurement trace to calculate To ensure that the most recent samples contain more valuable information. Based on this, two predictors are calculated separately: the short-term predictor y1 and the long-term predictor y2. These two predictors are defined according to the given sampling range, aiming to integrate the impact of short-term fluctuations and long-term trends to improve forecasting accuracy. The short-term predictor focuses on recent data patterns, while the long-term predictor takes into account a wider range of historical information. This dual predictor approach not only reflects the immediate dynamics of change, but also maintains an understanding of the overall trend, making the forecasting model more robust and accurate.

[0066] However, the above sampling-based methods cannot fully capture the dependencies in the collected load data (i.e., autocorrelations between samples). Therefore, an exponentially weighted moving average (EWMA) method is used to enhance them:

[0067] y i =y i-1 +a(x i -y i-1 )

[0068] where y i is the predicted value at time i, a=2n+1 is the weight parameter. Let y3 represent the prediction factor obtained by the EWMA method.

[0069] From the measurements, it can be observed that the accuracy of different predictors varies with the measurement period. For example, when the load fluctuates greatly, the short-term predictor performs better than the other two predictors. This is because the recent load has a greater impact on the current load than the long-term load, and the fluctuation further reduces the accuracy of the EWMA method. Since y1, y2, and y3 are relatively independent in the prediction, dynamic weights (weights can be adjusted according to their relative importance) α, β, and γ can be used to combine these three predictors and reduce the prediction error by optimizing the weight distribution. The final prediction value can be expressed as:

[0070] y=αy1+βy2+γy3

[0071] Among them, α+β+γ=1, and all weights are non-negative. This dynamic weight allocation mechanism allows the model to flexibly adjust the contribution ratio of each prediction factor according to the actual situation to ensure the best performance of the prediction results.

[0072] In order to improve the prediction accuracy in various dynamic environments, the coefficients α, β, and γ need to be updated over time. Specifically, at each time step i, a potential coefficient set containing N sets of coefficient vectors is first set, and the possible prediction value y is calculated i,1 ,y i,2,...,y i,j ,...,y i,J , as shown below:

[0073]

[0074] Among them (α n ,β n ,γ n ) is the possible coefficient vector (α, β, γ) at time i. Then, calculate y i,n and the actual delay x i The error vector between i,1 ,E i,2 ,...,E i,N ] and find out what makes E i,N The smallest n′∈{1, 2, …, N}:

[0075]

[0076] Finally, update {α i+1 ,β i+1 ,γ i+1}={α n′ ,β n′ ,γ n′}Make the next prediction.

[0077] S4. Since the frequency of base station sleep control update is much lower than the frequency of load prediction update, the base station needs to predict the load at multiple moments in the future. According to the first load prediction data and the second load prediction data at multiple moments in the future, the base station sleep control problem is modeled as an MDP and the MDP is solved, and the base station sleep control is performed based on this.

[0078] The base station sleep control problem is modeled as an MDP. The network controller that determines the sleep control of base stations in the area acts as an agent that attempts to solve the MDP. The system state in the MDP includes the load interval of all base stations at the current moment, the channel state of all users (whether blocked), and the transmission rate interval of all users. The agent's behavior is the working / sleeping decision of all base stations. The reward at each moment is defined as the negative value of the cost function at that moment. The agent's goal is to maximize the sum of the discounted rewards at the current moment and all subsequent moments.

[0079] The above two methods are used to predict the base station load at multiple moments in the future as the state of the intelligent agent to perform a solution based on deep reinforcement learning and formulate a base station sleep strategy.

[0080] The base station load prediction device combining deep learning and statistical methods provided by the present invention is described below. The base station load prediction device combining deep learning and statistical methods described below and the base station load prediction method combining deep learning and statistical methods described above can be referenced to each other.

[0081] An embodiment of the present invention provides a base station load prediction device combining deep learning and statistical methods, including:

[0082] The base station load data acquisition module obtains the load of the base station in the past period and at the current moment according to the number of users entering the base station coverage area as the base station load history data;

[0083] The data preprocessing module is used to divide the collected base station load historical data into two parts: Y1(t), which follows the long-term trend and changes according to a specific law over time, and Y2(t), which is a short-term random fluctuation, and to perform preprocessing operations such as cleaning and normalization on the original data;

[0084] The load prediction module consists of two sub-modules:

[0085] The LSTM neural network prediction submodule is used to train the LSTM neural network according to Y1(t), obtain the trained LSTM neural network to complete the first load prediction, and obtain the first load prediction data to capture and predict the long-term trend part of the load.

[0086] The statistical method prediction submodule is used to perform a second load prediction based on Y2(t) using an appropriate statistical method (such as exponential weighted moving average) to obtain second load prediction data, and to model and predict the short-term random fluctuation part of the load.

[0087] Multi-step prediction integration module: Since the frequency of base station sleep control update is much lower than the frequency of load prediction update, this module is used to integrate the first load prediction data and the second load prediction data at multiple future moments to generate load prediction values ​​at multiple future moments.

[0088] The base station sleep control module is used to model the base station sleep control problem as an MDP and solve it based on the integrated load prediction data, and perform base station sleep control accordingly.

[0089] The device provided by the embodiment of the present invention is applicable to multi-user multi-base station mobile communication systems in all high mobility scenarios.

[0090] An embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in any of the above embodiments when executing the computer program.

[0091] The embodiment of the present invention provides a base station load prediction system 300 combining deep learning and statistical methods, such as Figure 4 As shown, it includes: a computer readable storage medium 320 and a processor 310;

[0092] The computer-readable storage medium is used to store executable instructions;

[0093] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the method described in any of the above embodiments.

[0094] Specifically, the processor 310 may include, for example, a general-purpose microprocessor, an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 310 may also include an onboard memory for cache purposes. The processor 310 may be a processor for executing the above-mentioned Figure 1-Figure 3 The described method flow according to the embodiments of the present disclosure may be performed by a single processing unit or multiple processing units for different actions.

[0095] Computer readable storage medium 320, for example, can be any medium capable of containing, storing, conveying, propagating or transmitting instructions. For example, readable storage media can include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices, devices or propagation media. Specific examples of readable storage media include: magnetic storage devices, such as magnetic tapes or hard disks (HDD); optical storage devices, such as compact disks (CD-ROMs); memory, such as random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0096] The computer-readable storage medium 320 may include a computer program 321 , which may include code / computer-executable instructions, which when executed by the processor 310 enable the processor 310 to perform, for example, the method flow described in any of the above embodiments and any variations thereof.

[0097] The computer program 321 may be configured to have, for example, a computer program code including a computer program module. For example, in an exemplary embodiment, the code in the computer program 321 may include one or more program modules, for example, including module 321A, module 321B, ... It should be noted that the division method and number of modules are not fixed, and those skilled in the art may use appropriate program modules or program module combinations according to actual conditions. When these program module combinations are executed by the processor 310, the processor 310 may execute, for example, the method flow described in any of the above embodiments and any variations thereof.

[0098] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A base station load prediction method combining deep learning and statistical methods, characterized in that: include: S1, divide the base station load history data Y(t) into two parts: Y1(t) that follows the long-term trend and changes according to a specific law over time, and Y2(t) that fluctuates randomly in the short term; Y1(t) is the responsible data obtained by averaging the past load data in a sliding window, and Y2(t) is the difference between Y(t) and Y1(t); S2, training the LSTM neural network according to Y1(t), obtaining the trained LSTM neural network to complete the first load prediction, and obtaining the first load prediction data; S3, performing a second load prediction using a statistical method according to Y2(t) to obtain second load prediction data; S4, based on the first load prediction data and the second load prediction data at multiple future moments, the base station sleep control problem is modeled as an MDP and the MDP is solved, thereby performing base station sleep control.

2. The method according to claim 1, characterized in that The step S2 comprises: T1. Determine the structure and initial parameters of the LSTM neural network; T2, normalize Y1(t) and divide it into training data and test data for training and testing respectively; T3. Update the weights of the LSTM neural network by minimizing the loss function that represents the difference between the output of the LSTM neural network and the actual data. Use training data and test data to complete training and testing, respectively, to obtain the first load prediction data.

3. The method according to claim 2, characterized in that The LSTM neural network structure is as follows: 1) The input of the input layer is the sequence in the current moment and the previous sliding window; 2) After the input layer is a fully connected layer with an activation function of ReLu, which is used for feature extraction; 3) Further back are several recursive layers based on the LSTM structure, which are used to analyze and identify the correlation within the input sequence; 4) The output layer is a softmax function, in which the output value of each neuron represents the probability that the load is at the value corresponding to the neuron at the next moment.

4. The method according to claim 1, characterized in that The step S3 comprises: R1, use kernel density estimation to obtain short-term prediction value y1 and long-term prediction value y2 for Y2(t); R2, use the weighted moving average method to obtain the predicted value y3 for Y2(t); R3. Use dynamic weights to combine the above three prediction values ​​to obtain the final prediction value, i.e., the second load prediction data.

5. The method according to claim 4, characterized in that Step R3: y=αy1+βy2+γy3 Among them, y is the final prediction value, α, β and γ are dynamic weights, α+β+γ=1 and are all non-negative values.

6. A base station load prediction device combining deep learning and statistical methods, characterized in that: include: The base station load data acquisition module obtains the load of the base station in the past period and at the current moment according to the number of users entering the base station coverage area as the base station load history data; The data preprocessing module is used to divide the collected historical data of base station load into two parts: the long-term trend that changes according to a specific law over time Y1(t) and the short-term random fluctuation Y2(t); The load prediction module includes two sub-modules: The LSTM neural network prediction submodule is used to train the LSTM neural network according to Y1(t), obtain the trained LSTM neural network to complete the first load prediction, and obtain the first load prediction data; A statistical method prediction submodule, used to perform a second load prediction based on Y2(t) using a statistical method to obtain second load prediction data; A multi-step prediction integration module, used to integrate the first load prediction data and the second load prediction data at multiple future moments; The base station sleep control module is used to model the base station sleep control problem as an MDP and solve it based on the integrated load prediction data, and perform base station sleep control accordingly.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

Images