Multi-scale hybrid base station flow prediction system based on STL decomposition

Abstract

The multi-scale hybrid base station traffic prediction system based on STL decomposition comprises the following modules: a multi-scale time decomposition module which is used for performing multi-scale decomposition on base station traffic data; the function of the STL-based time mixing module is to perform STL decomposition on the multi-scale sub-sequence of the high-dimensional space obtained by the multi-scale time decomposition module, decompose the multi-scale sub-sequence into three parts of a seasonal term, a trend term and a residual term, and perform modeling on the time sequence through different sampling methods; the future multi-prediction hybrid module is used for integrating prediction results of different time scales and calculating a final prediction result through dynamic weight adjustment; according to the system, through a multi-scale time decomposition technology, the prediction precision, the real-time performance and the calculation efficiency can be effectively improved in a large-scale communication network environment.

Description

Technical Field

[0001] The invention relates to a multi-scale hybrid base station traffic prediction system based on STL decomposition, belonging to the field of information technology, in particular to the field of base station traffic prediction technology. Background Art

[0002] Base station traffic prediction is a method to predict future traffic by inputting historical traffic data of base stations. Through accurate prediction, base station resource allocation can be optimized, network performance can be improved, and energy consumption can be reduced.

[0003] The Chinese patent with application number 202310249218.X pre-processes base station traffic data through Kmeans clustering and EMD decomposition, and uses the BiLSTM model for prediction. Although EMD can decompose non-stationary signals, it cannot simultaneously process the characteristics of multiple time scales in the data, especially the changes in shorter periods, which may be ignored. The combination of Kmeans clustering and EMD will lead to a significant increase in computational complexity, especially in large-scale base station networks, where the computational load of the model is large, affecting real-time performance. Although the BiLSTM model can capture time series dependencies, its generalization ability for future data is limited, especially when processing data with noise and nonlinear periodic changes, the model behaves unstable.

[0004] The Chinese patent with application number 202310388134.4 uses EMD decomposition and MWOA algorithm to optimize the LSTM model for traffic prediction. Although MWOA can optimize the hyperparameters of LSTM, LSTM itself still has limitations when dealing with long-term dependency problems and is prone to losing important information in long sequence data. This method fails to fully consider the traffic correlation between multiple base station cells, and only makes independent predictions for a single base station cell, failing to capture the traffic transfer and correlation characteristics between multiple cells. STM is prone to overfitting in overly complex time series, especially when the amount of data is insufficient or the noise is too large, the prediction effect of the model is unstable.

[0005] The Chinese patent with application number 202410798641.X uses the Gray Wolf algorithm to optimize the Transformer model for communication base station traffic prediction. Although the Transformer can capture multi-dimensional features through the multi-head attention mechanism, the technology also has the following defects: the number of heads in the multi-head attention layer has a significant impact on the accuracy of the model, but manually selecting the number of heads is often difficult to achieve the best effect, and too many heads may increase the complexity of the model and increase the risk of overfitting. Although the Gray Wolf algorithm was introduced for optimization, it is difficult for the algorithm to effectively balance global search and local optimization in some scenarios, and premature convergence may occur, resulting in failure to find the optimal solution. The Transformer model may have prediction bias when processing traffic data with extreme changes, especially when facing abnormal fluctuations, the model's robustness is weak.

[0006] In summary, how to capture the characteristic changes of base station traffic on different time scales, reduce computational complexity, adapt to the real-time requirements of large-scale base station traffic prediction, and improve the overall accuracy and robustness of the prediction are technical problems that urgently need to be solved in the field of base station traffic prediction technology. Summary of the invention

[0007] In view of this, the purpose of the present invention is to decompose the time series of base station traffic into seasonality, trend and residual, so that the characteristics of different time scales can be processed independently, the computational complexity can be reduced, and the technical goals of real-time, accurate and robust prediction of large-scale base station traffic can be achieved.

[0008] In order to achieve the above object, the present invention proposes a multi-scale hybrid base station traffic prediction system based on STL decomposition, and the system includes the following modules:

[0009] Multi-scale time decomposition module: The function of this module is to decompose the base station traffic data into multiple scales;

[0010] The specific contents are as follows:

[0011] First, the long time series x(t) of base station traffic data is analyzed in the frequency domain. The discrete Fourier transform is used to convert the long time series x(t) from the time domain to the frequency domain. The first M highest peaks {f1,f2,…,f M}, the periods corresponding to these frequencies are These cycles represent the main modes of change in the series, and contain multi-scale features from short-term to long-term;

[0012] Secondly, after determining the main periods {T1, T2, …, T M}, these periods are used to perform multi-scale decomposition on the aforementioned long time series x(t), that is, the long time series x(t) of the base station traffic data is decomposed according to the period Tm Slice and average the corresponding positions of each segment to obtain the downsampled time series:

[0013]

[0014]

[0015] STL-based time mixing module: The function of this module is to convert the multi-scale subsequence X′ in the high-dimensional space obtained by the multi-scale time decomposition module m Perform STL decomposition into three parts: seasonal term, trend term, and residual term, and model the time series through different sampling methods;

[0016] The specific contents are as follows:

[0017] Through STL decomposition, the multi-scale subsequence X′ in the high-dimensional space is m Decomposition into seasonal terms Trending Items Residual Based on the different characteristic parts obtained by decomposition, hybrid operations are applied to the seasonal term, trend term and residual term respectively to achieve information interaction between multiple scales; the operation formula is as follows:

[0018]

[0019]

[0020] In the above formula, STLDecomp() represents the STL decomposition operation; FeedForward(·) contains two linear layers and the middle GELU activation function, which is used to perform information interaction operations between channels. S-Mix(·), T-Mix(·) and R-Mix(·) represent the mixing operations of seasonal terms, trend terms and residual terms, respectively.

[0021] Future multi-forecast hybrid module: The function of this module is to integrate the forecast results of different time scales and calculate the final forecast results through dynamic weight adjustment; the specific contents are as follows:

[0022] The final prediction result is calculated according to the following formula

[0023]

[0024] In the above formula, w m represents the weight of the prediction result of the mth scale, Represents the prediction result of the mth scale;

[0025] w m Calculate according to the following formula:

[0026]

[0027] Where Score() represents the scoring function, which is implemented through attention mechanism or other learning mechanism;

[0028] Calculate according to the following formula:

[0029]

[0030] In the above formula, They represent the seasonal, trend and residual prediction results of the mth scale, respectively, and are calculated according to the following formulas:

[0031]

[0032] In the above formula,

[0033] MLP() stands for Multi-layer Perceptron;

[0034] Linear1() represents the operation of the first fully connected layer, i.e., the linear layer. Its function is to perform a linear transformation on the input features to extract higher-dimensional deep features from the input embedded features.

[0035] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation, which is used to enhance the nonlinear ability of the model. After completing the linear transformation of Linear1, GELU will perform nonlinear processing on these deep features to capture the complex patterns in the input features.

[0036] Linear2() represents the operation of the second fully connected layer, i.e., the linear layer, which is used to project the features after GELU activation into the target space for future predictions.

[0037] The specific content of the S-Mix (·) seasonal term mixing operation is as follows: in the seasonal term mixing operation, a bottom-up approach is adopted to pass the information from the lower-level fine-scale time series upward to supplement the detailed information of the coarser-scale seasonal modeling; for the multi-scale seasonal part In the residual mode, the bottom-up seasonal information interaction operation is realized by using the bottom-up mixing layer Bottom-Up-Mixing according to the following formula:

[0038]

[0039] In the above formula,

[0040] Conv1D() means to input Perform a one-dimensional convolution operation to extract its local features and possibly reduce the length of the time series to better capture local correlations;

[0041] Linear1() means performing the first linear layer operation on the convolutional features, mapping the features to a new dimension and changing the feature space;

[0042] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation;

[0043] Linear2() means to perform a second linear layer operation on the features after nonlinear activation, and further project the results to Same dimensions;

[0044] The specific content of the mixing operation of the T-Mix (·) trend term is: adopting a top-down mixing method to use the macroscopic knowledge of the coarser scale to guide the trend modeling of the finer scale; for the multi-scale trend term In the residual mode, the top-down mixing layer Top-Down-Mixing is used according to the following formula to realize the top-down trend information interaction operation:

[0045]

[0046] In the above formula,

[0047] Linear1() represents the trend information at a coarser scale Perform the first linear layer operation and project it into a new feature space for extracting deep features;

[0048] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation;

[0049] Linear2() means to perform the second linear layer operation on the features after nonlinear activation, projecting them to Same feature dimensions;

[0050] The specific contents of the mixing operation of the R-Mix (·) residual term are as follows: for the residual terms, since they contain the unpredictable part of the time series, they are of great significance at all scales; in order to effectively utilize this information, a bottom-up mixing method is adopted; for the multi-scale residual terms According to the following formula, information interaction is realized through residual method:

[0051]

[0052] In the above formula,

[0053] Conv1D() means to input Perform one-dimensional convolution operation to extract local residual information and perform dimensionality reduction to highlight key features;

[0054] Linear1() means performing the first linear layer operation on the convolutional features and projecting the features to a new feature dimension;

[0055] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation, which performs nonlinear processing on the features after linear transformation;

[0056] Linear2() means performing a second linear layer operation on the features after nonlinear activation, and further projecting the features to the current residual information. Consistent dimensions.

[0057] The beneficial effects of the present invention are as follows: the system of the present invention can accurately capture the short-term and long-term periodic characteristics in the base station traffic data through multi-scale time decomposition technology, and further decompose the data into seasonal, trend and residual parts through STL decomposition, thereby realizing accurate processing of multi-scale features; through dynamic adjustment of weights, the ability to capture traffic fluctuations is significantly enhanced, especially in large-scale communication network environments, which can effectively improve prediction accuracy, real-time performance and computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a functional module diagram of the multi-scale hybrid base station traffic prediction system based on STL decomposition proposed by the present invention. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings.

[0060] See also Figure 1 , the present invention introduces a multi-scale hybrid base station traffic prediction system based on STL (Seasonal-Trend decomposition using LOESS) decomposition, the system includes the following modules:

[0061] Multi-scale time decomposition module: The function of this module is to decompose the base station traffic data into multiple scales;

[0062] The specific contents are as follows:

[0063] First, the long time series x(t) of base station traffic data is analyzed in the frequency domain. The discrete Fourier transform is used to convert the long time series x(t) from the time domain to the frequency domain. The first M highest peaks {f1,f2,…,fM}, the periods corresponding to these frequencies are These cycles represent the main modes of change in the series, and contain multi-scale features from short-term to long-term;

[0064] Secondly, after determining the main periods {T1, T2, …, T M}, these periods are used to perform multi-scale decomposition on the aforementioned long time series x(t), that is, the long time series x(t) of the base station traffic data is decomposed according to the period T m Slice and average the corresponding positions of each segment to obtain the downsampled time series:

[0065]

[0066]

[0067] In the experiment of the embodiment, the data used by the inventors include hourly data of base station traffic. Multi-scale decomposition analysis is performed on the data of the first 48 hours. First, discrete Fourier transform (FFT) is performed on the data of the first 48 hours. The specific values ​​are as follows:

[0068] The data subset is [0.77,0.52,0.41,0.12,0.10,0.19,0.19,0.68,0.51,1.13,1.11,1.64,5.04,2.03,1.41,1.60,3.78,3.07,3.56,2.21,2.62,2.86,1.84,1.03,0.88,2.51,0.26,0.19,0.07,0.09,0.23,0.90,0.75,0.72,1.93,1.50,1.62,1.49,1.87,1.45,1.74,2.56,3.08,2.90,4.20,2.72,1.75,1.19].

[0069] The amplitude spectrum obtained by FFT is:

[0070] ,4.92,1.27,2.61,8.26,1.08,2.55,6.31,8.06,7.36,4.83,4.88,4.87,7.58,4.69,8.31,5.78,31.28,4.62].

[0071] Find several significant peaks in the amplitude spectrum, and their corresponding frequencies are:

[0072] [0.0417,0.0833,0.125,0.1667,0.2292,0.3125,0.375,0.4583].

[0073] These frequencies represent different periodic variation characteristics. The inventor selected the first three significant frequencies to determine the main periodic characteristics:

[0074] The corresponding periods are 3 hours, 12 hours and 24 hours.

[0075] According to the extracted period, the uplink traffic data is decomposed into multiple scales:

[0076] The traffic data is divided into multiple 24-hour segments with a 24-hour period. For example, the first segment is [0.77, 0.52, 0.41, 0.12, ..., 1.03].

[0077] For all these 24-hour segments, the flow at each time point is averaged to obtain a downsampled sequence representing a 24-hour period.

[0078] Through this embedding operation, the system can better capture the complex change patterns in the time series, thereby improving the accuracy of the forecast. The subsequences with longer periods mainly reflect the long-term trend, while the subsequences with shorter periods capture short-term fluctuations. The multi-scale time series representation finally constructed can analyze and utilize the characteristics of the time series at different scales.

[0079] Through the above-mentioned periodic decoupling method, the original time series is decomposed into subsequences at different scales, each of which reflects the main change pattern of the corresponding period in the time series. Subsequences with longer periods mainly reflect long-term trends, while subsequences with shorter periods capture short-term fluctuations. The multi-scale time series representation constructed in this way can analyze and utilize the characteristics of time series at different scales.

[0080] STL-based time mixing module: The function of this module is to convert the multi-scale subsequence X′ in the high-dimensional space obtained by the multi-scale time decomposition module m Perform STL decomposition into three parts: seasonal term, trend term, and residual term, and model the time series through different sampling methods;

[0081] The specific contents are as follows:

[0082] Through STL decomposition, the multi-scale subsequence X′ in the high-dimensional space is m Decomposition into seasonal terms Trending Items Residual Based on the different characteristic parts obtained by decomposition, hybrid operations are applied to the seasonal term, trend term and residual term respectively to achieve information interaction between multiple scales; the operation formula is as follows:

[0083]

[0084]

[0085] In the above formula, STLDecomp() represents the STL decomposition operation; FeedForward(·) contains two linear layers and the middle GELU activation function, which is used to perform information interaction operations between channels. S-Mix(·), T-Mix(·) and R-Mix(·) represent the mixing operations of seasonal terms, trend terms and residual terms, respectively.

[0086] Assume that there is a segment of base station traffic data, such as a 24-hour traffic record every day. This segment of data can reflect the change in the number of service bytes of the base station. First, the data is decomposed by STL, and the data is decomposed into seasonal terms, trend terms, and residual terms. In this process, the uplink and downlink data are divided into different periods respectively. The division of each period is mainly based on the significant frequency obtained by spectrum analysis. According to these periods, the data is divided into the following time series of different scales. In order to achieve multi-scale information fusion, different mixing operations are performed on each component (season, trend, residual):

[0087] Seasonal Mixing (S-Mix): Bottom-Up Mixing is used to gradually merge short-term seasonal information upward into longer periods. This ensures that detailed information is fully captured and reflected in coarse-scale seasonal modeling.

[0088] Trend Term Mixing (T-Mix): Top-Down Mixing is used to pass information from longer cycles downward to ensure that macro trends can guide the modeling process of shorter cycles.

[0089] Residual Mix (R-Mix): A certain bottom-up fusion can be performed to ensure that detail information is fully utilized.

[0090] In order to predict the traffic volume for the next 24 hours, the seasonal term, trend term and residual term of each scale are predicted separately. The specific steps are as follows: For each scale of seasonal term, trend term and residual term, an MLP predictor is designed to predict the changes in the next 24 hours. Each MLP contains two linear layers and a GELU activation function: Conv1d is used for local feature extraction and downsampling: the number of input channels in_channels = 1, the number of output channels out_channels = 1. The convolution kernel size kernel_size = 3, the step size stride = 2, and the padding padding = 1. This means that after each convolution operation, the length of the feature is reduced by half, but the same number of output channels is maintained.

[0091] Linear layer Linear 1: The input dimension is input_dim / / 2+1, which is the feature dimension after convolution. The output dimension is consistent with the input to maintain the integrity of the features.

[0092] Linear layer Linear 2: The dimensions of input and output are the same, and further feature transformation is performed.

[0093] The activation function uses GELU to perform nonlinear processing on the input features and capture complex patterns.

[0094] The input data first passes through Conv1d, and the dimension of the result after convolution will be reduced.

[0095] Adjust the shape, pass it to the linear layer Linear 1, and then pass it through the GELU activation function.

[0096] Through Linear 2, the feature dimension is mapped back to the original space.

[0097] Through the above multi-scale mixing module, the fusion results corresponding to each sub-sequence are finally obtained. These fused results can provide rich multi-scale information input for further time series prediction, which helps to improve the accuracy of the prediction.

[0098] Future multi-forecast hybrid module: The function of this module is to integrate the forecast results of different time scales and calculate the final forecast results through dynamic weight adjustment; the specific contents are as follows:

[0099] The final prediction result is calculated according to the following formula

[0100]

[0101] In the above formula, w m represents the weight of the prediction result of the mth scale, Represents the prediction result of the mth scale;

[0102] w m Calculate according to the following formula:

[0103]

[0104] Where Score() represents the scoring function, which is implemented through attention mechanism or other learning mechanism;

[0105] Calculate according to the following formula:

[0106]

[0107] In the above formula, They represent the seasonal, trend and residual prediction results of the mth scale, respectively, and are calculated according to the following formulas:

[0108]

[0109]

[0110]

[0111] In the above formula,

[0112] MLP() stands for Multi-layer Perceptron;

[0113] Linear1() represents the operation of the first fully connected layer, i.e., the linear layer. Its function is to perform a linear transformation on the input features to extract higher-dimensional deep features from the input embedded features.

[0114] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation, which is used to enhance the nonlinear ability of the model. After completing the linear transformation of Linear1, GELU will perform nonlinear processing on these deep features to capture the complex patterns in the input features.

[0115] Linear2() represents the operation of the second fully connected layer, i.e., the linear layer, which is used to project the features after GELU activation into the target space for future predictions.

[0116] In the embodiment, the inventors used multi-scale decomposition to decompose the data into different periods, such as 12 hours, 24 hours, 163 hours, etc. to capture short-term fluctuations and long-term trends in traffic. In the future multi-forecast hybrid module, the goal is to integrate the seasonality, trend, and residual prediction results at each scale to generate the final future traffic forecast. Each component (seasonality, trend, and residual) is predicted by an MLP, and then the prediction results of each scale are integrated together.

[0117] For each scale of data, MLP is used to predict seasonality (S), trend (T) and residual (R).

[0118] The predicted value of each component (e.g. ) is calculated by MLP:

[0119]

[0120] Finally, the predicted values ​​of each component are added together to get the total prediction result for that scale:

[0121]

[0122] Dynamic weights are used to integrate forecast results at different time scales, ultimately resulting in a forecast for the next 24 hours.

[0123] Weight (w m ) is calculated through a scoring function to indicate the importance of each scale.

[0124] MLP parameters: For each component (seasonality, trend, and residual), the input dimension is the feature length of the original mixture.

[0125] Linear 1: For period t, the input dimension is t and the output dimension is 2t.

[0126] Linear 2: Convert the nonlinearly mapped features into predictions for the next 24 hours.

[0127] Finally, the prediction results of each period are weighted and summed by dynamic weights to obtain the prediction for the next 24 hours.

[0128] The specific content of the S-Mix (·) seasonal term mixing operation is as follows: in the seasonal term mixing operation, a bottom-up approach is adopted to pass the information from the lower-level fine-scale time series upward to supplement the detailed information of the coarser-scale seasonal modeling; for the multi-scale seasonal part In the residual mode, the bottom-up seasonal information interaction operation is realized by using the bottom-up mixing layer Bottom-Up-Mixing according to the following formula:

[0129]

[0130] In the above formula,

[0131] Conv1D() means to input Perform a one-dimensional convolution operation to extract its local features and possibly reduce the length of the time series to better capture local correlations;

[0132] Linear1() means performing the first linear layer operation on the convolutional features, mapping the features to a new dimension and changing the feature space;

[0133] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation;

[0134] Linear2() means to perform a second linear layer operation on the features after nonlinear activation, and further project the results to Same dimensions;

[0135] The specific content of the mixing operation of the T-Mix (·) trend term is: adopting a top-down mixing method to use the macroscopic knowledge of the coarser scale to guide the trend modeling of the finer scale; for the multi-scale trend term In the residual mode, the top-down mixing layer Top-Down-Mixing is used according to the following formula to realize the top-down trend information interaction operation:

[0136]

[0137] In the above formula,

[0138] Linear1() represents the trend information at a coarser scale Perform the first linear layer operation and project it into a new feature space for extracting deep features;

[0139] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation;

[0140] Linear2() means to perform the second linear layer operation on the features after nonlinear activation, projecting them to Same feature dimensions;

[0141] The specific contents of the mixing operation of the R-Mix (·) residual term are as follows: for the residual terms, since they contain the unpredictable part of the time series, they are of great significance at all scales; in order to effectively utilize this information, a bottom-up mixing method is adopted; for the multi-scale residual terms According to the following formula, information interaction is realized through residual method:

[0142]

[0143] In the above formula,

[0144] Conv1D() means to input Perform one-dimensional convolution operation to extract local residual information and perform dimensionality reduction to highlight key features;

[0145] Linear1() means performing the first linear layer operation on the convolutional features and projecting the features to a new feature dimension;

[0146] GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation, which performs nonlinear processing on the features after linear transformation;

[0147] Linear2() means to perform the second linear layer operation on the features after nonlinear activation, and further project the features to the current residual information r l m Consistent dimensions.

[0148] In an embodiment, the multi-layer perceptron (MLP) structure technical parameters are as follows:

[0149] Linear 1:

[0150] Input dimension: The data length of different periods is different, and the input dimension corresponds to the feature length of the original mixture. For example, the input of period_24 is 24, the input of period_163 is 163, and the input of period_12 is 12.

[0151] Output dimension: The output dimension is usually set to twice the original input dimension to increase the capacity and expressiveness of the model.

[0152] For period_24, the output dimension is 24*2=48.

[0153] For period_163, the output dimension is 163*2=326.

[0154] For period_12, the output dimension is 12*2=24.

[0155] Activation function: GELU uses GELU (Gaussian Error Linear Unit) as the activation function, which can enhance the nonlinear ability of the model and capture the complex patterns of input features. Compared with other activation functions (such as ReLU and Tanh), GELU has a smoother activation effect, allowing the model to perform more flexible nonlinear mapping in different input ranges.

[0156] Linear 2:

[0157] Input dimension: Accepts the features processed by the first linear layer and activation function. The input dimension is the output of Linear 1.

[0158] Output dimension: Set to the number of future time steps to be predicted, for example 24 in this scenario, i.e. the prediction results for the next 24 hours.

[0159] The inventors have conducted a large number of experiments on the system proposed in the present invention, and the experimental results are as follows:

[0160] The experiment used a base station traffic data set, including uplink and downlink traffic data of multiple base stations at different time points. The data covers 60 days of 1,080 base stations, with 24 hours of observation data per day. The experiment analyzed the periodicity of various time scales, including short, medium and long periods, and evaluated the prediction accuracy of seasonality, trend and residual terms at each time scale. The seasonality, trend and residual terms of each scale were predicted for the next 24 hours, and the prediction results were fused to finally obtain the prediction results for the traffic in the next 24 hours.

[0161] In order to verify the stability and consistency of the experimental results, the experiment was repeated several times on different base stations and different time periods. The results of repeated experiments show that the proposed system has higher prediction accuracy and stability in most cases. A variety of evaluation indicators were used in the experiment to comprehensively evaluate the prediction performance. The experiment not only evaluated the overall prediction accuracy, but also conducted a detailed comparative analysis of the prediction performance on different time scales, different prediction time periods (short-term and long-term), and different characteristics (such as seasonality, trend, residual terms), proving that multi-scale fusion has advantages in improving the prediction accuracy of different time ranges. In the experiment, several typical base stations were selected, and the prediction results of their uplink and downlink traffic were analyzed in detail. The experimental results show that the system of the present invention has better applicability and advantages in dealing with different types of base stations and different traffic patterns. The experimental results show that the fusion of information at different time scales can effectively capture short-term fluctuations and long-term trends, and the prediction accuracy and stability are further improved through the dynamic weight adjustment module.

[0162] Experimental results prove that the method proposed in the present invention is effective.

Claims

1. A multi-scale hybrid base station traffic prediction system based on STL decomposition, characterized by: The system includes the following modules: Multi-scale time decomposition module: The function of this module is to decompose the base station traffic data into multiple scales; The specific contents are as follows: First, the long time series x(t) of base station traffic data is analyzed in the frequency domain. The discrete Fourier transform is used to convert the long time series x(t) from the time domain to the frequency domain. The first M highest peaks {f1,f2,…,f M }, the periods corresponding to these frequencies are These cycles represent the main modes of change in the series, and contain multi-scale features from short-term to long-term; Secondly, after determining the main periods {T1, T2, …, T M }, these periods are used to perform multi-scale decomposition on the aforementioned long time series x(t), that is, the long time series x(t) of the base station traffic data is decomposed according to the period T m Slice and average the corresponding positions of each segment to obtain the downsampled time series: Finally, the multi-scale subsequence is projected into a high-dimensional feature space through the embedding network layer, which is specifically expressed as: Among them, X′ m It is the multi-scale subsequence representation after being mapped to a high-dimensional space; Embed() represents the operation of embedding the network layer; STL-based time mixing module: The function of this module is to convert the multi-scale subsequence X′ in the high-dimensional space obtained by the multi-scale time decomposition module m Perform STL decomposition into three parts: seasonal term, trend term, and residual term, and model the time series through different sampling methods; The specific contents are as follows: Through STL decomposition, the multi-scale subsequence X′ in the high-dimensional space is m Decomposition into seasonal terms Trending Items Residual Based on the different characteristic parts obtained by decomposition, hybrid operations are applied to the seasonal term, trend term and residual term respectively to achieve information interaction between multiple scales; the operation formula is as follows: In the above formula, STLDecomp( ) represents the STL decomposition operation; FeedForward(·) contains two linear layers and the middle GELU activation function, which is used to perform information interaction operations between channels. S-Mix(·), T-Mix(·) and R-Mix(·) represent the mixing operations of seasonal terms, trend terms and residual terms, respectively. Future multi-forecast hybrid module: The function of this module is to integrate the forecast results of different time scales and calculate the final forecast results through dynamic weight adjustment; the specific contents are as follows: The final prediction result is calculated according to the following formula In the above formula, w m represents the weight of the prediction result of the mth scale, Represents the prediction result of the mth scale; w m Calculate according to the following formula: Where Score( ) represents the scoring function, which is implemented through attention mechanism or other learning mechanism; Calculate according to the following formula: In the above formula, They represent the seasonal, trend and residual prediction results of the mth scale, respectively, and are calculated according to the following formulas: In the above formula, MLP() stands for Multi-layer Perceptron; Linear1() represents the operation of the first fully connected layer, i.e., the linear layer. Its function is to perform a linear transformation on the input features to extract higher-dimensional deep features from the input embedded features. GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation, which is used to enhance the nonlinear ability of the model. After completing the linear transformation of Linear1, GELU will perform nonlinear processing on these deep features to capture the complex patterns in the input features. Linear2() represents the operation of the second fully connected layer, i.e., the linear layer, which is used to project the features after GELU activation into the target space for future predictions.

2. The multi-scale hybrid base station traffic prediction system based on STL decomposition according to claim 1 is characterized by: The specific content of the S-Mix (·) seasonal term mixing operation is as follows: in the seasonal term mixing operation, a bottom-up approach is adopted to pass the information from the lower-level fine-scale time series upward to supplement the detailed information of the coarser-scale seasonal modeling; for the multi-scale seasonal part In the residual mode, the bottom-up seasonal information interaction operation is realized by using the bottom-up mixing layer Bottom-Up-Mixing according to the following formula: In the above formula, Conv1D() means to input Perform a one-dimensional convolution operation to extract its local features and possibly reduce the length of the time series to better capture local correlations; Linear1() means performing the first linear layer operation on the convolutional features, mapping the features to a new dimension and changing the feature space; GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation; Linear2() means to perform a second linear layer operation on the features after nonlinear activation, and further project the results to Same dimensions; The specific content of the mixing operation of the T-Mix (·) trend term is: adopting a top-down mixing method to use the macroscopic knowledge of the coarser scale to guide the trend modeling of the finer scale; for the multi-scale trend term In the residual mode, the top-down mixing layer Top-Down-Mixing is used according to the following formula to realize the top-down trend information interaction operation: In the above formula, Linear1() represents the trend information at a coarser scale Perform the first linear layer operation and project it into a new feature space for extracting deep features; GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation; Linear2() means to perform the second linear layer operation on the features after nonlinear activation, projecting them to Same feature dimensions; The specific contents of the mixing operation of the R-Mix (·) residual term are as follows: for the residual terms, since they contain the unpredictable part of the time series, they are of great significance at all scales; in order to effectively utilize this information, a bottom-up mixing method is adopted; for the multi-scale residual terms According to the following formula, information interaction is realized through residual method: In the above formula, Conv1D() means to input Perform one-dimensional convolution operation to extract local residual information and perform dimensionality reduction to highlight key features; Linear1() means performing the first linear layer operation on the convolutional features and projecting the features to a new feature dimension; GELU() represents the Gaussian Error Linear Unit nonlinear activation function operation, which performs nonlinear processing on the features after linear transformation; Linear2() means performing a second linear layer operation on the features after nonlinear activation, and further projecting the features to the current residual information. Consistent dimensions.

Images