5G base station power consumption prediction method and system based on conditional generative adversarial network

Abstract

The present invention relates to a 5G base station power consumption prediction method based on a conditional generative adversarial network, and belongs to the field of 5G technology. The method of the present invention generates conditional data based on the historical power consumption data of a base station to be predicted and a plurality of spatially associated base stations at a plurality of moments before a specified moment, takes the historical power consumption data of the base station to be predicted at a specified moment as real data, and forms a set of training data based on the conditional data, the real data and random noise data, and trains an improved conditional generative adversarial network based on a training set formed by a plurality of sets of training data; uses the generative model in the trained improved conditional generative adversarial network to predict the power consumption of the base station to be predicted at the time to be predicted based on random noise data and the conditional data for prediction. The method of the present invention realizes accurate prediction of the power consumption of a 5G base station based on the spatiotemporal characteristics of the power consumption of the 5G base station, provides an accurate basis for 5G network planning and deployment, and is of great significance to improving the resource utilization rate of the distribution network and ensuring the safety of the power grid.

Description

Technical Field

[0001] The present invention belongs to the field of 5G technology, and in particular relates to a 5G base station power consumption prediction method and system based on a conditional generative adversarial network. Background Art

[0002] With the rapid development of 5G mobile communication technology, the large-scale application of high-energy-consuming 5G base stations has increased the operating costs of base stations and exacerbated problems such as imbalance between supply and demand in distribution networks. Accurate analysis and prediction of 5G base station power consumption characteristics is the key basis for the rational planning and deployment of distribution networks, and is of great significance to improving the efficiency of power grid resource utilization and ensuring power grid security.

[0003] At present, the prediction of 5G base station power consumption is generally divided into two categories: traditional prediction methods and intelligent prediction methods. Traditional prediction methods mainly include linear methods such as trend extrapolation, time series, and regression models, and nonlinear methods such as generalized multifractal wavelets and autoregressive fractional integrated moving average models. Traditional prediction methods are difficult to accurately model the complex random and nonlinear factors affecting the power consumption of 5G base stations, and the prediction accuracy is low. Intelligent prediction methods usually include: BP neural network, support vector machine SVM, neural network CNN, RNN, etc. The existing intelligent prediction methods lack the study of the relationship between deep learning models and the spatiotemporal distribution characteristics of 5G base station power consumption, and the prediction results are still far from the actual power consumption. Summary of the invention

[0004] In view of the above analysis, the present invention aims to provide a 5G base station power consumption prediction method and system based on a conditional generative adversarial network. By considering the spatiotemporal characteristics of 5G base station power consumption, based on the historical power consumption data of the base station to be predicted and multiple spatially associated base stations with the base station to be predicted, and based on an improved conditional generative adversarial network, accurate prediction of the base station power consumption of the base station to be predicted is achieved.

[0005] On the one hand, the present invention provides a 5G base station power consumption prediction method based on a conditional generative adversarial network, which specifically includes the following steps:

[0006] Generate conditional data based on historical power consumption data of a base station to be predicted and multiple spatially associated base stations at multiple times before a specified time, use the historical power consumption data of the base station to be predicted at the specified time as real data, form a set of training data based on the conditional data, the real data and random noise data, and train an improved conditional generative adversarial network based on multiple sets of training data to form a training set;

[0007] Generate condition data for prediction based on power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple times before the time to be predicted;

[0008] The trained generative model in the improved conditional generative adversarial network is used to predict the power consumption of the base station to be predicted at the time to be predicted based on random noise data and conditional data for prediction.

[0009] Furthermore, the generating condition data based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple times before the specified time includes:

[0010] Constructing a power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted at each moment in the multiple moments and the multiple spatially associated base stations;

[0011] Forming a matrix sequence based on the power consumption matrices at the multiple moments;

[0012] Extracting features from the matrix sequence to generate conditional data.

[0013] Further, the constructing of a power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted at each moment in the multiple moments and the multiple spatially associated base stations includes:

[0014] Selecting base stations within a circular range with the base station to be predicted as the center and r as the radius as the multiple spatially associated base stations;

[0015] With the coordinates (m, n) of the base station to be predicted as the origin, divide the x-axis and y-axis directions into (2r+1)×(2r+1) square areas of the same size;

[0016] Based on the power consumption data of the spatially associated base stations corresponding to each of the square areas, a power consumption matrix corresponding to each time is constructed, which is expressed as:

[0017]

[0018] Wherein, (m,n) represents the coordinates of the base station to be predicted, (mr,nr), (mr,n), (mr,n+r), (m,n+r), (m,n+r), (m+r,nr), (m+r,n+r) represent the coordinates of the center points of the corresponding square areas respectively; is the power consumption data of the base stations in each of the square areas at the historical time t; Represents the power consumption matrix corresponding to the base station to be predicted at the historical time t.

[0019] Further, the corresponding power consumption data in each of the square areas is determined include:

[0020] For each of the spatially associated base stations, respectively calculating the Euclidean distance to the center point of each square area, and determining the square area to which the spatially associated base station belongs based on the shortest Euclidean distance;

[0021] Determine the power consumption data corresponding to the square area based on the power consumption data of each of the spatially associated base stations in each square area

[0022] When the number of spatially associated base stations in the square area is 0,

[0023] When the number of spatially associated base stations in the square area is 1, the power consumption data of the spatially associated base station is Get value;

[0024] When there are multiple spatially associated base stations in the square area, the average power consumption data of the multiple associated base stations is taken as Get the value.

[0025] Furthermore, the improved conditional generative adversarial network includes a generative model and a discriminative model;

[0026] The generation model includes a condition generation module and a prediction generation module; the condition generation module is used to extract features from the matrix sequence to generate condition data; the prediction generation module is used to generate prediction data based on the condition data and noise data;

[0027] The discriminant model is used to discriminate the real data and the predicted data to obtain a discriminant result.

[0028] Furthermore, the improved conditional generative adversarial network based on multiple sets of training data to form a training set includes:

[0029] In each iteration:

[0030] Based on the conditional data and random noise data of each set of training data in the training set, the prediction data is obtained using the generative model of the improved conditional generative adversarial network;

[0031] Based on the predicted data and the real data, the discrimination result is obtained based on the discriminant model and the loss function value of this round of iteration is calculated;

[0032] Based on the loss function value, determine whether to end the training. If so, end the training. If not, adjust the generator and discriminator for the next round of iteration based on the discrimination result.

[0033] Furthermore, an improved loss function of a conditional generative adversarial network is constructed based on the deviation between the predicted data and the real data, the feature loss of the discriminant model, the loss function of the generative model and the loss function of the discriminant model.

[0034] Furthermore, the loss function of the improved conditional generative adversarial network is expressed as:

[0035]

[0036] Among them, G represents the generative model; D represents the discriminative model; L w Indicates the deviation between the predicted data and the actual data; L loss (G,x) represents the feature loss of the discriminative model; λ w and λ loss Respectively represent the corresponding weights;

[0037] represents the loss function of the conditional generative adversarial network CGAN, Based on the loss function of the generative model and the loss function of the discriminative model,

[0038]

[0039] Among them, x, y, and z represent real data, conditional data, and predicted data, respectively; E x,y and E z,y (·) represents the expected value of the distribution of x, y and the expected value of the distribution of z, y respectively; G(·) is the output value of the generative model, and D(·) is the output value of the discriminative model.

[0040] On the other hand, the present invention also provides a 5G base station power consumption prediction system based on a conditional generative adversarial network, comprising:

[0041] A data processing module, used to generate conditional data based on power consumption data at multiple times before a specified time in historical power consumption data of a base station to be predicted and multiple base stations in a specified surrounding area, use the power consumption data of the base station to be predicted at a specified time as real data, form a set of training data based on the conditional data, the real data and random noise data, form a training set based on multiple sets of training data, and generate conditional data for prediction based on power consumption data at multiple times before a predicted time of the base station to be predicted and multiple base stations in a specified surrounding area;

[0042] A training module for training an improved conditional generative adversarial network based on a training set;

[0043] The prediction module is used to use the trained improved conditional generative adversarial network generator to predict the power consumption of the base station to be predicted at the time to be predicted based on random noise data and conditional data for prediction.

[0044] Furthermore, the loss function of the improved conditional generative adversarial network is expressed as:

[0045]

[0046] Among them, G represents the generative model; D represents the discriminative model; L w Indicates the deviation between the predicted data and the actual data; L loss(G,x) represents the feature loss of the discriminative model; λ w and λ loss Respectively represent the corresponding weights;

[0047] represents the loss function of the conditional generative adversarial network CGAN, Based on the loss function of the generative model and the loss function of the discriminative model,

[0048]

[0049] Among them, x, y, and z represent real data, conditional data, and predicted data, respectively; E x,y and E z,y (·) represents the expected value of the distribution of x, y and the expected value of the distribution of z, y respectively; G(·) is the output value of the generative model, and D(·) is the output value of the discriminative model.

[0050] The present invention can achieve at least one of the following beneficial effects:

[0051] By analyzing the time dependence of 5G base station power consumption, that is, the periodic fluctuation characteristics and long-term stable regularity characteristics over time, as well as the spatial correlation characteristics of 5G base station power consumption, we select corresponding data training conditions to generate an adversarial network based on the spatiotemporal characteristics of base station power consumption, and use the trained improved conditional generative adversarial network to select data that is associated with out-of-control at the time to be predicted to achieve accurate prediction of the base station power consumption of the base station to be predicted.

[0052] By adopting a multi-layer three-dimensional convolutional (Conv3D) neural network structure to construct an improved conditional generative adversarial network, the spatiotemporal characteristics of 5G base station power consumption can be effectively captured. This has significant advantages over the traditional Conv2D structure that can only extract single time or spatial scale features.

[0053] The loss function of the conditional generative adversarial network is improved by comprehensively considering the evaluation effects of the L1 loss function and the L2 loss function and the feature loss of the discriminant model, which further improves the accuracy of the training of the improved conditional generative adversarial network and thus improves the accuracy of the prediction.

[0054] Other features and advantages of the present invention will be described in the following description, and some advantages may become apparent from the description, or may be understood by practicing the present invention. The purpose and other advantages of the present invention may be realized and obtained through the contents particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 A schematic diagram of the structure of the improved conditional generative adversarial network of the present invention;

[0058] Figure 3 This is a 24-hour power consumption variation trend diagram of a 5G base station of the present invention;

[0059] Figure 4 This is a 24×7 hour power consumption change trend diagram of a 5G base station of the present invention;

[0060] Figure 5 The autocorrelation coefficient diagram calculated by the present invention for a certain time point of a time-based sequence;

[0061] Figure 6 A partial correlation coefficient graph calculated for a certain time point of a time-based sequence according to the present invention;

[0062] Figure 7 It is the correlation coefficient matrix of the present invention. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0064] Method Embodiment

[0065] A specific embodiment of the present invention discloses a 5G base station power consumption prediction method based on a conditional generative adversarial network. By considering the spatiotemporal characteristics of the 5G base station power consumption and based on the historical power consumption data of the base station to be predicted and multiple spatially associated base stations with the base station to be predicted, an accurate prediction of the base station power consumption of the base station to be predicted is achieved based on the conditional generative adversarial network.

[0066] The method of the present invention specifically includes steps S01 to S03.

[0067] Step S01, generate conditional data based on the historical power consumption data of the base station to be predicted and multiple spatially associated base stations at multiple times before a specified time, take the historical power consumption data of the base station to be predicted at the specified time as real data, and form a group of training data based on the conditional data, the real data and random noise data, and form a training set based on multiple groups of training data to train an improved conditional generative adversarial network.

[0068] Specifically, generating conditional data based on historical power consumption data of the base station to be predicted and multiple spatially associated base stations at multiple times before a specified time includes:

[0069] Constructing a power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted at each moment in the multiple moments and the multiple spatially associated base stations;

[0070] Forming a matrix sequence based on the power consumption matrices at the multiple moments;

[0071] Extracting features from the matrix sequence to generate conditional data.

[0072] Further, constructing a power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at each moment in the multiple moments includes:

[0073] Selecting base stations within a circular range with the base station to be predicted as the center and r as the radius as the multiple spatially associated base stations;

[0074] With the coordinates (m, n) of the base station to be predicted as the origin, divide the x-axis and y-axis directions into (2r+1)×(2r+1) square areas of the same size;

[0075] Based on the power consumption data of the spatially associated base stations corresponding to each of the square areas, a power consumption matrix corresponding to each time is constructed, which is expressed as:

[0076]

[0077] Wherein, (m,n) represents the coordinates of the base station to be predicted, (mr,nr), (mr,n), (mr,n+r), (m,n+r), (m,n+r), (m+r,nr), (m+r,n+r) represent the coordinates of the center points of the corresponding square areas respectively; The power consumption data corresponding to each base station in the square area at the historical time t; Represents the power consumption matrix corresponding to the base station to be predicted at the historical time t.

[0078] Further, the corresponding power consumption data in each of the square areas is determined include:

[0079] For each of the spatially associated base stations, the Euclidean distances to the center points of the square areas are calculated respectively, and the square area to which the spatially associated base station belongs is determined based on the shortest Euclidean distance; it should be noted that, in practical applications, the multiple spatially associated base stations are usually not evenly and equidistantly distributed, therefore, there may be multiple spatially associated base stations in each square area, or there may be one or zero spatially associated base stations; in addition, since the base stations are selected within a circular range of a radius of r around the base station to be predicted, in the large square area composed of (2r+1)×(2r+1) square areas, the number of base stations in the square areas at the four corners of the large square area must be 0;

[0080] Determine the power consumption data corresponding to the square area based on the power consumption data of each of the spatially associated base stations in each square area

[0081] When the number of spatially associated base stations in the square area is 0,

[0082] When the number of spatially associated base stations in the square area is 1, the power consumption data of the spatially associated base station is Get value;

[0083] When there are multiple spatially associated base stations in the square area, the average power consumption data of the multiple associated base stations is taken as Get the value.

[0084] Specifically, the power consumption data of the base station to be predicted at the historical time t+1 is used as the real data.

[0085] Specifically, in the method of the present invention, the improved conditional generative adversarial network includes a generative model and a discriminative model (such as Figure 2 shown);

[0086] The generation model includes a condition generation module and a prediction generation module; the condition generation module is used to extract features from the matrix sequence to generate condition data; the prediction generation module is used to generate prediction data based on the condition data and noise data;

[0087] The discriminant model is used to discriminate the real data and the predicted data to obtain a discriminant result.

[0088] Further, such as Figure 2As shown in Figure 1, the conditional generation module includes L layers of neural networks, each of which includes three-dimensional convolution (Conv3D), batch normalization (Bath Normalization, BN) and Relu activation function. Among them, Conv3D can extract high-dimensional abstract features of the 5G base station power consumption sequence, which can improve the prediction accuracy of the model. The output of the l∈{1,…,L}th layer at the tth moment in the conditional generation module is expressed as in, represents the output of the l-1th layer, The table shows the three-dimensional convolution kernel of the lth layer, and * indicates the convolution operation. After L layers of convolution operations, the abstract feature representation of the historical power consumption sequence is obtained as

[0089] Furthermore, The conditional data y of the improved conditional generative adversarial network is concatenated with the noise data z, and the concatenated result is input into the prediction generation module, such as Figure 2 As shown in (a), the prediction generation module includes three layers of "Conv+BN+ReLu" neural network; the output of the third layer of the upgraded network is calculated by the Relu activation function to obtain the prediction data at time t+1

[0090] Further, Input to Figure 2 In the discriminant model shown in (b) (the discriminant model contains 4 layers of "Conv+BN+LealyReLu" neural network, and 1 fully connected layer, i.e., a hidden layer neural network), the discriminant model outputs the true or false probability between the real data and the prediction.

[0091] Furthermore, the improved conditional generative adversarial network based on multiple sets of training data to form a training set includes:

[0092] In each iteration:

[0093] Based on the conditional data and random noise data of each set of training data in the training set, the prediction data is obtained using the generative model of the improved conditional generative adversarial network;

[0094] Based on the predicted data and the real data, the discrimination result is obtained based on the discriminant model and the loss function value of this round of iteration is calculated;

[0095] Based on the loss function value, determine whether to end the training. If so, end the training. If not, adjust the generator and discriminator for the next round of iteration based on the discrimination result.

[0096] Furthermore, an improved loss function of a conditional generative adversarial network is constructed based on the deviation between the predicted data and the real data, the feature loss of the discriminant model, the loss function of the generative model and the loss function of the discriminant model.

[0097] It should be noted that the loss function of the improved conditional generative adversarial network constructed in the present invention is obtained by improving the loss function of the conditional generative adversarial network CGAN by considering the deviation between the predicted data and the real data and the feature loss of the discriminant model.

[0098] The loss function L(G, D) of the improved conditional generative adversarial network is expressed as:

[0099]

[0100] Among them, G represents the generative model; D represents the discriminative model; represents the loss function of the conditional generative adversarial network CGAN; L w Indicates the deviation between the predicted data and the actual data; L loss (G,x) represents the feature loss of the discriminative model; λ w and λ loss Represent the corresponding weights respectively; during training, the smaller the loss value of the loss function L(G,D), the better the performance of the improved conditional generative adversarial network.

[0101] Furthermore, It is obtained based on the loss function of the generation model and the loss function of the discriminant model.

[0102] Specifically, the generative model first needs to determine how close the predicted data distribution is to the real data distribution, and secondly, whether the G(z|y) calculated by the generative model is the same as the probability distribution of the conditional data y. Therefore, the loss function L of the generative model in the conditional generative adversarial network CGAN is determined. G And the loss function L of the discriminant model D They are:

[0103] L G =-E z,y (D(G(z|y)|y));

[0104] L D =-E z,y (D(G(x|y)))+E z,y (D(G(z|y)|y));

[0105] The generative model hopes to increase the output value of the predicted data G(z|y) in the discriminative model, while the discriminative model hopes to reduce the output value of the predicted data G(z|y) and increase the output value of the real data. Therefore, Expressed as

[0106]

[0107] Among them, x, y, and z represent real data, conditional data, and predicted data, respectively; E x,y and E z,y (·) represents the expected value of the distribution of x, y and the expected value of the distribution of z, y respectively; G(·) is the output value of the generative model, and D(·) is the output value of the discriminative model.

[0108] Furthermore, L w It represents the deviation between the predicted data and the real data, which is obtained based on the L1 loss function and the L2 loss function and is expressed as:

[0109] L w =w1E x,y,z (||xG(z|y)||)+w2E x,y,z (||xG(z|y)|| 2 );

[0110] Among them, ||·|| represents the calculation of absolute deviation, w1 is the weight of L1 loss function, and w2 is the weight of L2 loss function.

[0111] It should be noted that the L1 loss function is called the mean absolute error function, that is, the absolute value of the difference between the predicted value and the true value; the L2 loss function is called the mean squared error function, that is, the square of the difference between the predicted value and the true value. Since L2 uses the square of the error, if the absolute value of the error is greater than 1, the error will be amplified a lot. If L2 is used as the loss function, the model will be more sensitive to abnormal samples, which will sacrifice many normal samples. At this time, L1 will be more effective. The problem with L1 is that its gradient will jump at the extreme point, and a small difference will also bring a large gradient, which is not conducive to model learning. Therefore, the method of the present invention takes into account the advantages and disadvantages of L1 and L2, and determines the weighted mixed loss function L w .

[0112] Furthermore, L loss (G,x) represents the feature loss of the discriminant model, expressed as: L per (G,x)=E x,y,z (λ i P i (G,x)); where P i (G,x) is the predicted data generated by the generative model and the real data in the hidden layer h i The deviation value, h i is the hidden layer of the discriminant model; h i (G) is the generative model in the hidden layer h i The predicted data ofi is the weight parameter; h i (x) is the hidden layer h i The real data of C i , H i and W i The hidden layers h i The number of channels, height, and width of the

[0113] Step S02: Generate condition data for prediction based on the power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple time points before the time point to be predicted.

[0114] Specifically, when predicting the power consumption of the predicted base station at the predicted time t'+1, it is necessary to obtain power consumption data of multiple times before the predicted time of the predicted base station and power consumption data of multiple spatially associated base stations at multiple times before the predicted time.

[0115] Furthermore, the specific steps of generating the conditional data y' for prediction are the same as step S01, which will not be described again here.

[0116] Step S03: Use the generative model in the trained improved conditional generative adversarial network to predict the power consumption of the base station to be predicted at the time to be predicted based on the random noise data z' and the conditional data y' used for prediction.

[0117] The following is an explanation of the principle of considering the spatiotemporal characteristics of 5G base station power consumption data when generating conditional data in step S01 of this embodiment:

[0118] The present invention analyzes and studies the actual power consumption data of multiple 5G base stations in multiple regions, and takes multiple macro base stations in one region as an example to illustrate the spatiotemporal characteristics of the power consumption of 5G base stations.

[0119] Specifically, a 5G macro base station is selected to study the time characteristics of base station power consumption. Figure 3 The following is a 24-hour power consumption trend chart of the 5G base station. Figure 4 This is the 24×7 hour power consumption trend chart of the 5G base station. Figure 3 It can be seen that the power consumption of 5G base stations gradually decreases from 0:00 to 5:00, and then gradually increases after 5:00, which lasts until about 10:00. From 10:00 to 17:00, the power consumption shows a fluctuating upward trend and reaches the maximum, and then gradually decreases after 17:00. Figure 4 It can be seen that within a 24×7 cycle, 5G power consumption shows obvious periodic fluctuation characteristics. Figure 3 and Figure 4It shows that the power consumption of 5G base stations has obvious daily and weekly time series characteristics, and the power consumption has dependence characteristics on the time scale. The actual power consumption data of multiple 5G base stations in multiple regions studied by this invention are similar to Figure 3 and Figure 4 daily and weekly cycle characteristics.

[0120] Furthermore, the autocorrelation coefficient and partial correlation coefficient of the 5G base station power consumption data for one week are calculated based on the time series, such as Figure 5 , Figure 6 The figure shows the autocorrelation coefficient and partial correlation coefficient calculated for a time-based sequence at a certain time point. The autocorrelation coefficient represents the correlation between the time-based 5G base station power consumption data series at time points t and tk, where k is the lag order (indicating the number of time intervals traced back from the current time point); Figure 5 It can be seen that the autocorrelation coefficient has a periodic decay law with the lag order k, but it fluctuates continuously and will not decay to 0, indicating that the power consumption data of 5G base stations has time-dependent characteristics. The partial correlation coefficient measures the correlation between the power consumption data at a certain moment in the time-based sequence and the power consumption data at a previous moment, describing the independent influence of the kth lag moment on the current moment after considering the influence of all power consumption data before the k-1 moment; Figure 6 It can be seen that the partial autocorrelation coefficient of the time-based series of 5G base station power consumption data decays with the increase of the lag order, indicating that the time-based series of 5G base station power consumption data within this period is stable, that is, the weekly periodic data has a stable time dependence.

[0121] Specifically, 10 5G macro base stations are selected to study the spatial characteristics of base station power consumption. The base station power consumption data of the 10 5G macro base stations at multiple historical moments are selected to form 10 power consumption data series, and the correlation coefficients of the 10 base station power consumption data series are calculated. For example, the Spearman correlation coefficient can be used. Figure 7 The figure shows the correlation coefficient matrix, where u1, u2, ..., u10 represent the serial numbers of 10 5G macro base stations. Figure 7 It can be seen that the correlation coefficient of the base station power consumption data series of base stations u3 and u4 is 0.95, and the correlation coefficient of base stations u3 and u5 is 0.98, both of which are strongly correlated; while the correlation coefficient between u8 and u9 is 0.04, and the correlation coefficient between u5 and u9 is 0.23, that is, they are weakly correlated.

[0122] Furthermore, the direct spatial distances between the 10 base stations are calculated, and the rule is determined: the closer the distance between two base stations, the higher the correlation of the power consumption data series, and the farther the distance, the lower the correlation of the power consumption data series. That is, the power consumption of a certain base station will have a spatial correlation with multiple other base stations in close proximity.

[0123] To sum up, the power consumption data of 5G base stations has obvious spatiotemporal characteristics. Therefore, in the method of this embodiment, conditional data is generated based on the historical power consumption data of the base station to be predicted and multiple spatially associated base stations at multiple times before the specified time, which is used to train the conditional generation adversarial network; and conditional data for prediction is generated based on the power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple times before the predicted time, so as to make full use of the spatiotemporal characteristics of the 5G base station power consumption data and improve the accuracy of the prediction.

[0124] The present embodiment discloses a 5G base station power consumption prediction method based on a conditional generative adversarial network. By analyzing the time dependence of the 5G base station power consumption, that is, the periodic fluctuation characteristics and long-term stable regularity characteristics over time, and analyzing the spatial correlation characteristics of the 5G base station power consumption, corresponding data is selected based on the spatiotemporal characteristics of the base station power consumption to train an improved conditional generative adversarial network, and the trained improved conditional generative adversarial network is used to select data that is associated with out-of-control at the time to be predicted, so as to achieve accurate prediction of the base station power consumption of the base station to be predicted.

[0125] When constructing an improved conditional generative adversarial network, by adopting a multi-layer three-dimensional convolutional (Conv3D) neural network structure, the spatiotemporal characteristics of 5G base station power consumption can be effectively captured. This has significant advantages over the traditional Conv2D structure that can only extract single time or spatial scale features.

[0126] When determining the loss function, the evaluation effects of the L1 loss function and the L2 loss function and the feature loss of the discriminant model are comprehensively considered, and the loss function of the conditional generative adversarial network CGAN is improved, which further improves the accuracy of the improved conditional generative adversarial network training and thus improves the accuracy of the prediction.

[0127] System Example

[0128] A 5G base station power consumption prediction system based on conditional generative adversarial network includes a data processing module, a training module and a prediction module.

[0129] Specifically, the data processing module is used to generate conditional data based on the power consumption data at multiple times before a specified time in the historical power consumption data of the base station to be predicted and multiple base stations in a specified surrounding area, take the power consumption data of the base station to be predicted at the specified time as real data, form a group of training data based on the conditional data, the real data and random noise data, form a training set based on multiple groups of training data, and generate conditional data for prediction based on the power consumption data at multiple times before a predicted time of the base station to be predicted and multiple base stations in the specified surrounding area.

[0130] Specifically, the training module is used to train an improved conditional generative adversarial network based on a training set.

[0131] Specifically, the prediction module is used to use the trained improved conditional generative adversarial network generator to predict the power consumption of the base station to be predicted at the time to be predicted based on random noise data and conditional data for prediction.

[0132] Furthermore, the loss function L(G, D) of the improved conditional generative adversarial network is expressed as:

[0133]

[0134] Among them, G represents the generative model; D represents the discriminative model; The loss function of the conditional generative adversarial network CGAN; L w Indicates the deviation between the predicted data and the actual data; L loss (G,x) represents the feature loss of the discriminative model; λ w and λ loss Respectively represent the corresponding weights;

[0135] Based on the loss function of the generative model and the loss function of the discriminative model,

[0136]

[0137] Among them, x, y, and z represent real data, conditional data, and predicted data, respectively; E x,y and E z,y (·) represents the expected value of the distribution of x, y and the expected value of the distribution of z, y respectively; G(·) is the output value of the generative model, and D(·) is the output value of the discriminative model.

[0138] Compared with the prior art, the beneficial effects of a 5G base station power consumption prediction system based on a conditional generative adversarial network provided in this embodiment are basically the same as the beneficial effects provided by the method embodiment, which will not be elaborated here one by one.

[0139] It should be noted that the above embodiments are based on the same inventive concept and parts not described repeatedly can be used as reference for each other.

[0140] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A 5G base station power consumption prediction method based on conditional generative adversarial network, characterized in that: The steps include: Generate conditional data based on historical power consumption data of a base station to be predicted and multiple spatially associated base stations at multiple times before a specified time, use the historical power consumption data of the base station to be predicted at the specified time as real data, form a set of training data based on the conditional data, the real data and random noise data, and train an improved conditional generative adversarial network based on multiple sets of training data to form a training set; The generating conditional data based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple moments before the specified moment includes: constructing a power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at each moment in the multiple moments; forming a matrix sequence based on the power consumption matrices at the multiple moments; extracting features from the matrix sequence to generate conditional data; the constructing of the power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at each moment in the multiple moments includes: selecting base stations within a circular range with the base station to be predicted as the center and r as the radius as the multiple spatially associated base stations; taking the base station to be predicted as the center coordinate (m, n) as the origin, dividing (2r+1)×(2r+1) square areas of the same size in the x-axis and y-axis directions; constructing a power consumption matrix corresponding to each moment based on the power consumption data of the corresponding spatially associated base stations in each of the square areas, expressed as: Wherein, (m,n) represents the coordinates of the base station to be predicted, (mr,nr), (mr,n), (mr,n+r), (m,n+r), (m,n+r), (m+r,nr), (m+r,n+r) represent the coordinates of the center points of the corresponding square areas respectively; is the power consumption data of the base stations in each of the square areas at the historical time t; Represents the power consumption matrix corresponding to the base station to be predicted at the historical time t; Determine the power consumption data corresponding to each of the square areas The method comprises: for each of the spatially associated base stations, respectively calculating the Euclidean distance to the center point of each square area, and determining the square area to which the spatially associated base station belongs based on the shortest Euclidean distance; and determining the power consumption data corresponding to the square area based on the power consumption data of each of the spatially associated base stations in each square area. When the number of spatially associated base stations in the square area is 0, When the number of spatially associated base stations in the square area is 1, the power consumption data of the spatially associated base station is When there are multiple spatially associated base stations in the square area, the average power consumption data of the multiple associated base stations is taken as Get value; Generate condition data for prediction based on power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple times before the time to be predicted; Using the generative model in the trained improved conditional generative adversarial network, based on random noise data and conditional data for prediction, the power consumption of the base station to be predicted at the time to be predicted is predicted.

2. The power consumption prediction method according to claim 1, characterized in that: The improved conditional generative adversarial network includes a generative model and a discriminative model; The generation model includes a condition generation module and a prediction generation module; The condition generation module is used to extract features from the matrix sequence to generate condition data; the prediction generation module is used to generate prediction data based on the condition data and noise data; The discriminant model is used to discriminate the real data and the predicted data to obtain a discriminant result.

3. The power consumption prediction method according to claim 2, characterized in that: The improved conditional generative adversarial network based on multiple sets of training data to form a training set training includes: In each iteration: Based on the conditional data and random noise data of each set of training data in the training set, the prediction data is obtained using the generative model of the improved conditional generative adversarial network; Based on the predicted data and the real data, the discrimination result is obtained based on the discriminant model and the loss function value of this round of iteration is calculated; Based on the loss function value, determine whether to end the training. If so, end the training. If not, adjust the generation model and the discrimination model based on the discrimination result for the next round of iteration.

4. The power consumption prediction method according to claim 3, characterized in that: An improved loss function of a conditional generative adversarial network is constructed based on the deviation between the predicted data and the real data, the feature loss of the discriminant model, the loss function of the generative model and the loss function of the discriminant model.

5. The power consumption prediction method according to claim 4, characterized in that: The loss function of the improved conditional generative adversarial network is expressed as: Among them, G represents the generative model; D represents the discriminative model; L w Indicates the deviation between the predicted data and the actual data; L loss (G,x) represents the feature loss of the discriminative model; λ w and λ loss Respectively represent the corresponding weights; represents the loss function of the conditional generative adversarial network CGAN, Based on the loss function of the generative model and the loss function of the discriminative model, Among them, x, y, and z represent real data, conditional data, and predicted data, respectively; E x,y and E z,y (·) represents the expected value of the distribution of x, y and the expected value of the distribution of z, y respectively; G(·) is the output value of the generative model, and D(·) is the output value of the discriminative model.

6. A 5G base station power consumption prediction system based on conditional generative adversarial network, characterized in that: include: A data processing module, used to generate conditional data based on power consumption data at multiple times before a specified time in historical power consumption data of a base station to be predicted and multiple base stations in a specified surrounding area, use the power consumption data of the base station to be predicted at a specified time as real data, form a set of training data based on the conditional data, the real data and random noise data, form a training set based on multiple sets of training data, and generate conditional data for prediction based on power consumption data at multiple times before a predicted time of the base station to be predicted and multiple base stations in a specified surrounding area; The generating conditional data based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at multiple moments before the specified moment includes: constructing a power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at each moment in the multiple moments; forming a matrix sequence based on the power consumption matrices at the multiple moments; extracting features from the matrix sequence to generate conditional data; the constructing of the power consumption matrix corresponding to each moment based on the historical power consumption data of the base station to be predicted and the multiple spatially associated base stations at each moment in the multiple moments includes: selecting base stations within a circular range with the base station to be predicted as the center and r as the radius as the multiple spatially associated base stations; taking the base station to be predicted as the center coordinate (m, n) as the origin, dividing (2r+1)×(2r+1) square areas of the same size in the x-axis and y-axis directions; constructing a power consumption matrix corresponding to each moment based on the power consumption data of the corresponding spatially associated base stations in each of the square areas, expressed as: Wherein, (m,n) represents the coordinates of the base station to be predicted, (mr,nr), (mr,n), (mr,n+r), (m,n+r), (m,n+r), (m+r,nr), (m+r,n+r) represent the coordinates of the center points of the corresponding square areas respectively; is the power consumption data of the base stations in each of the square areas at the historical time t; Represents the power consumption matrix corresponding to the base station to be predicted at the historical time t; Determine the power consumption data corresponding to each of the square areas The method comprises: for each of the spatially associated base stations, respectively calculating the Euclidean distance to the center point of each square area, and determining the square area to which the spatially associated base station belongs based on the shortest Euclidean distance; and determining the power consumption data corresponding to the square area based on the power consumption data of each of the spatially associated base stations in each square area. When the number of spatially associated base stations in the square area is 0, When the number of spatially associated base stations in the square area is 1, the power consumption data of the spatially associated base station is When there are multiple spatially associated base stations in the square area, the average power consumption data of the multiple associated base stations is taken as Get value; A training module for training an improved conditional generative adversarial network based on a training set; The prediction module is used to use the generative model of the trained improved conditional generative adversarial network to predict the power consumption of the base station to be predicted at the time to be predicted based on random noise data and conditional data for prediction.

7. The power consumption prediction system according to claim 6, characterized in that: The loss function of the improved conditional generative adversarial network is expressed as: Among them, G represents the generative model; D represents the discriminative model; L w Indicates the deviation between the predicted data and the actual data; L loss (G,x) represents the feature loss of the discriminative model; λ w and λ loss Respectively represent the corresponding weights; represents the loss function of the conditional generative adversarial network CGAN, Based on the loss function of the generative model and the loss function of the discriminative model, Among them, x, y, and z represent real data, conditional data, and predicted data, respectively; E x,y and E z,y (·) represents the expected value of the distribution of x, y and the expected value of the distribution of z, y respectively; G(·) is the output value of the generative model, and D(·) is the output value of the discriminative model.

Images