A method, device, and readable storage medium for predicting cellular traffic data

Abstract

The application discloses a kind of cellular traffic data prediction method, device, equipment and readable storage medium, it is related to communication technical field, to improve the accuracy of the predicted cellular traffic data obtained.The method comprises: obtaining historical cellular traffic data;According to the historical cellular traffic data, obtain a plurality of univariate time series;For the first univariate time series in a plurality of univariate time series, obtain the feature map of the first univariate time series;According to the feature map, obtain the predicted value of each first univariate time series;According to the predicted value of each first univariate time series, obtain predicted cellular traffic data.The embodiment of the application can improve the accuracy of the predicted cellular traffic data obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, and particularly relates to a cellular traffic data prediction method and device, equipment and a readable storage medium. BACKGROUND

[0002] Cellular traffic data is a typical structured time series data. Traditional prediction models based on statistical learning and machine learning are difficult to effectively handle the complex nonlinear relationship of cellular traffic data, and are difficult to model and predict in the long term, with low prediction accuracy. In recent years, methods based on deep learning have been widely used in cellular network traffic prediction problems to cope with these complexity challenges.

[0003] Among them, common modeling methods include methods based on recurrent neural network (RNN) and methods based on temporal convolutional network (TCN) with convolution structure.

[0004] However, the existing modeling method makes the expression ability of the model insufficient, thereby leading to inaccurate predicted cellular traffic data. SUMMARY

[0005] The embodiments of the present application provide a cellular traffic data prediction method, device, equipment and readable storage medium to improve the accuracy of the obtained predicted cellular traffic data.

[0006] In a first aspect, the embodiments of the present application provide a cellular traffic data prediction method, comprising:

[0007] obtaining historical cellular traffic data;

[0008] obtaining a plurality of univariate time series according to the historical cellular traffic data;

[0009] for a first univariate time series in the plurality of univariate time series, obtaining a feature map of the first univariate time series;

[0010] obtaining a predicted value of each of the first univariate time series according to the feature map;

[0011] obtaining predicted cellular traffic data according to the predicted value of each of the first univariate time series.

[0012] Optionally, the obtaining of the feature map of the first univariate time series comprises:

[0013] The steps S1-S3 are executed cyclically N times, and an output of a converter network obtained by the Nth execution of S1-S3 is taken as a feature map of the first univariate time series, where N is an integer greater than or equal to 2:

[0014] S1: divide the first target object into a plurality of patches to obtain a patch sequence, wherein when N = 1, the first target object comprises the first univariate time series, or when N ≥ 2, the first target object comprises an intermediate feature map obtained by the N-1th execution of steps S1-S3, and the number of patches used when the first target object is divided into a plurality of patches is determined according to the patch parameters used when the first target object is divided into a plurality of patches in the N-1th execution of steps S1-S3;

[0015] S2: positionally encode the patch sequence to obtain a positionally encoded result;

[0016] S3: take the positionally encoded result as an input of a converter network and run the converter network to obtain an intermediate feature map of the first univariate time series.

[0017] Optionally, when N ≥ 2, the step of dividing the first target object into a plurality of patches to obtain a patch sequence comprises:

[0018] dividing the first target object into a plurality of patches and linearly mapping the plurality of patches to obtain the patch sequence.

[0019] Optionally, the step of positionally encoding the patch sequence to obtain a positionally encoded result comprises:

[0020] linearly mapping each patch in the patch sequence to obtain a high-dimensional feature of the patch sequence;

[0021] positionally encoding based on position information of each patch in the patch sequence and the high-dimensional feature to obtain the positionally encoded result.

[0022] Optionally, the step of taking the positionally encoded result as an input of a converter network and running the converter network to obtain an intermediate feature map of the first univariate time series comprises:

[0023] mapping the positionally encoded result into a plurality of subspaces using a multi-head attention mechanism to obtain an attention output result;

[0024] processing the attention output result through a normalization operation and a feedforward neural network to obtain the intermediate feature map.

[0025] Optionally, when N ≥ 2, the number of patches used when the first target object is divided into a plurality of patches is determined in the following manner:

[0026]

[0027] wherein, N m denotes the number of blocks when the step S1 to S3 is executed for the mth time, N m-1 denotes the number of blocks when the step S1 to S3 is executed for the (m-1)th time, P m-1 denotes the length of block when the step S1 to S3 is executed for the (m-1)th time, S m-1 denotes the step of non-overlapping area between two consecutive blocks when the step S1 to S3 is executed for the (m-1)th time, m is an integer, 2≤m≤N.

[0028] Optionally, the obtaining, according to the feature map, of the prediction value of each first univariate time series comprises:

[0029] linearly predicting the feature map to obtain the prediction value of each first univariate time series.

[0030] In a second aspect, an embodiment of the present application provides a device for predicting cellular traffic data, comprising:

[0031] a first obtaining module configured to obtain historical cellular traffic data;

[0032] a second obtaining module configured to obtain a plurality of univariate time series according to the historical cellular traffic data;

[0033] a third obtaining module configured to obtain, for a first univariate time series in the plurality of univariate time series, a feature map of the first univariate time series;

[0034] a fourth obtaining module configured to obtain, according to the feature map, a prediction value of each first univariate time series;

[0035] a fifth obtaining module configured to obtain predicted cellular traffic data according to the prediction value of each first univariate time series.

[0036] Optionally, the third obtaining module is further configured to execute the steps S1 to S3 for N times in a loop, and take the output of the converter network obtained by executing the steps S1 to S3 for the Nth time as the feature map of the first univariate time series, wherein N is an integer greater than or equal to 2:

[0037] S1: divide the first target object into a plurality of patches to obtain a patch sequence, wherein when N = 1, the first target object comprises the first univariate time series, or when N ≥ 2, the first target object comprises an intermediate feature map obtained by performing steps S1 to S3 for the (N-1)th time, and the number of patches used when dividing the first target object into a plurality of patches is determined according to the patch parameters used when dividing the first target object into a plurality of patches in the (N-1)th time of performing steps S1 to S3;

[0038] S2: position encoding is performed on the patch sequence to obtain a position encoding result;

[0039] S3: the position encoding result is taken as an input of a transformer network, and the transformer network is run to obtain an intermediate feature map of the first univariate time series.

[0040] Optionally, the third acquisition module is further configured to, when N ≥ 2, divide the first target object into a plurality of patches and perform linear mapping on the plurality of patches to obtain the patch sequence.

[0041] Optionally, the third acquisition module is further configured to:

[0042] linearly map each patch in the patch sequence to obtain a high-dimensional feature of the patch sequence;

[0043] perform position encoding based on position information of each patch in the patch sequence and the high-dimensional feature to obtain the position encoding result.

[0044] Optionally, the third acquisition module is further configured to:

[0045] map the position encoding result into a plurality of subspaces using a multi-head attention mechanism to obtain an attention output result;

[0046] perform processing on the attention output result through a normalization operation and a feedforward neural network to obtain the intermediate feature map.

[0047] Optionally, when N ≥ 2, the number of patches used when dividing the first target object into a plurality of patches is determined in the following manner:

[0048]

[0049] wherein N m represents the number of patches in the mth time of performing steps S1 to S3, N m-1 represents the number of patches in the (m-1)th time of performing steps S1 to S3, P m-1 represents the patch length in the (m-1)th time of performing steps S1 to S3, and S m-1denotes a step size of a non-overlapping region between two consecutive patches when the steps S1-S3 are performed for the m-1th time, m is an integer, 2≤m≤N.

[0050] Optionally, the fourth obtaining module is further configured to:

[0051] perform linear prediction head processing on the feature map to obtain a predicted value of each of the first univariate time series.

[0052] In a third aspect, an embodiment of the present application further provides a cellular traffic data prediction device, comprising: a processor and a transceiver, wherein the processor is configured to:

[0053] obtain historical cellular traffic data;

[0054] obtain a plurality of univariate time series according to the historical cellular traffic data;

[0055] for a first univariate time series in the plurality of univariate time series, obtain a feature map of the first univariate time series;

[0056] obtain a predicted value of each of the first univariate time series according to the feature map;

[0057] obtain predicted cellular traffic data according to the predicted value of each of the first univariate time series.

[0058] Optionally, the processor is further configured to:

[0059] perform steps S1-S3 for N times, and take an output of a converter network obtained by performing S1-S3 for the Nth time as the feature map of the first univariate time series, wherein N is an integer greater than or equal to 2:

[0060] S1: divide a first target object into a plurality of patches to obtain a patch sequence, wherein when N=1, the first target object comprises the first univariate time series, or when N≥2, the first target object comprises an intermediate feature map obtained by performing steps S1-S3 for the N-1th time, and a patch number adopted when the first target object is divided into a plurality of patches is determined according to a patch parameter adopted when the first target object is divided into a plurality of patches for the N-1th time;

[0061] S2: positionally encode the patch sequence to obtain a positionally encoded result;

[0062] S3: take the positionally encoded result as an input of a converter network and run the converter network to obtain an intermediate feature map of the first univariate time series.

[0063] Optionally, the processor is further configured to: when N≥2, divide the first target object into a plurality of patches, and perform linear mapping on the plurality of patches to obtain the patch sequence.

[0064] Optionally, the processor is further configured to:

[0065] perform linear mapping on each patch in the patch sequence to obtain a high-dimensional feature of the patch sequence;

[0066] perform position encoding based on position information of each patch in the patch sequence and the high-dimensional feature to obtain the position encoding result.

[0067] Optionally, the processor is further configured to:

[0068] map the position encoding result into a plurality of subspaces using a multi-head attention mechanism to obtain an attention output result;

[0069] perform processing on the attention output result through a normalization operation and a feedforward neural network to obtain the intermediate feature map.

[0070] Optionally, when N≥2, the number of patches used when dividing the first target object into a plurality of patches is determined in the following manner:

[0071]

[0072] wherein, N m represents the number of patches when the steps S1-S3 are performed for the mth time, N m-1 represents the number of patches when the steps S1-S3 are performed for the (m-1)th time, P m-1 represents the patch length when the steps S1-S3 are performed for the (m-1)th time, S m-1 represents a step size of a non-overlapping region between two consecutive patches when the steps S1-S3 are performed for the (m-1)th time, and m is an integer, 2≤m≤N.

[0073] Optionally, the processor is further configured to:

[0074] perform processing on the feature map using a linear prediction head to obtain a predicted value of each first univariate time series.

[0075] In a fourth aspect, an embodiment of the present application further provides a communication device, including a memory, a processor, and a program stored in the memory and executable on the processor, and the processor implements the steps in the cellular traffic data prediction method when executing the program.

[0076] In a fifth aspect, the embodiments of the present application further provide a readable storage medium, which stores a program. The program is executed by a processor to implement the steps of the method for predicting cellular traffic data.

[0077] In the embodiments of the present application, for each of the plurality of univariate time series obtained from the historical cellular traffic data, a corresponding feature map is obtained, so that the predicted value of each univariate time series is obtained based on the feature map, and then the predicted cellular traffic data can be obtained based on the obtained plurality of predicted values. Since the cellular traffic data has characteristics such as time series data, in the scheme of the embodiments of the present application, these characteristics of the cellular traffic data can be fully considered through the feature map, and thus the accuracy of the prediction of the cellular traffic data can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 is one of the flowcharts of the method for predicting cellular traffic data provided by the embodiments of the present application;

[0079] Figure 2 is the second flowchart of the method for predicting cellular traffic data provided by the embodiments of the present application;

[0080] Figure 3 is one of the structural diagrams of the device for predicting cellular traffic data provided by the embodiments of the present application;

[0081] Figure 4 is the second structural diagram of the device for predicting cellular traffic data provided by the embodiments of the present application. DETAILED DESCRIPTION

[0082] In the embodiments of the present application, the term "and / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that there are three cases of A alone, A and B together, and B alone. The character " / " generally represents an "or" relationship between the associated objects before and after it.

[0083] In the embodiments of the present application, the term "a plurality of" means two or more, and other quantifiers are similar.

[0084] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0085] Referring to Figure 1 , Figure 1 is the flowchart of the method for predicting cellular traffic data provided by the embodiments of the present application, asFigure 1 As shown, the method comprises the following steps:

[0086] Step 101, obtaining historical cellular traffic data.

[0087] Here, the historical cellular traffic data of a wireless cell can be obtained, denoted as:

[0088]

[0089] wherein L is the sequence length of the historical cellular traffic data, M is the number of features, L and M are integers greater than or equal to 1. X can be understood as a multivariate time series. For each time slice t (1≤t≤L), there is x t which can be understood as the multivariate cellular data at a certain time. The historical cellular traffic data can be understood as a time series.

[0090] Step 102, obtaining a plurality of univariate time series according to the historical cellular traffic data.

[0091] Here, the multivariate time series X=(x1,x2,…,x L ) is divided into M univariate time series wherein i=1,…M.

[0092] Step 103, for a first univariate time series in the plurality of univariate time series, obtaining a feature map of the first univariate time series.

[0093] wherein the first univariate time series can be any univariate time series.

[0094] In the embodiments of the present application, the feature map is mainly obtained based on a Transformer network. Here, a pyramid structure is introduced into the Transformer network architecture, which can generate a feature map for modeling the multi-period features of the cellular traffic data, so as to improve the prediction accuracy and reduce the computational consumption of the model. At the same time, in the embodiments of the present application, the model can support better learning of time series feature representation from longer historical traffic data sequences, improving the generality and prediction accuracy of the model.

[0095] The implementation of the transformer network architecture can include multiple stages, each stage having the same model structure and being capable of generating feature maps of different scales, respectively. The model structure can include a patch division module, a position encoding module, and a multi-layer transformer encoder module. The patch division module divides the input into multiple patches to obtain a patch sequence. The position encoding module performs position encoding based on the patch sequence. The multi-layer transformer encoder module obtains the feature map of each stage based on the result of the position encoding.

[0096] Specifically, in the process of obtaining the feature map of the first univariate time series, N stages can be included, where N is an integer greater than or equal to 2. Optionally, N = 4. At this time, four scale feature maps {F1, F2, F3, F4} can be obtained according to patches of different lengths, and the prediction accuracy of the model can be significantly improved through the four feature maps.

[0097] That is, in the embodiments of the present application, steps S1 to S3 are executed N times in a loop, and the output of the transformer network obtained by the Nth execution of S1 to S3 is taken as the feature map of the first univariate time series. Wherein,

[0098] S1: divide the first target object into multiple patches to obtain a patch sequence, wherein when N = 1, the first target object includes the first univariate time series, or when N ≥ 2, the first target object includes an intermediate feature map obtained by the N-1th execution of steps S1 to S3, and the number of patches used when the first target object is divided into multiple patches is determined according to the patch parameters used when the first target object is divided into multiple patches in the N-1th execution of steps S1 to S3. The patch parameters can include patch length, patch number, step length, etc.

[0099] When N = 1, i.e., the process is executed for the first time, the input first univariate time series is divided into overlapping or non-overlapping patches, each patch has a length P, and the non-overlapping area between two consecutive patches has a step length S. After the patching operation, a patch sequence is generated: where N is the number of patches, and the calculation method is as follows:

[0100]

[0101] By introducing patches, the input length of the model can be reduced from L to about This makes the time complexity and memory consumption of attention mechanism calculation decrease by a factor of S square. Therefore, the patch-based design can increase the length of the historical sequence of model input, thereby significantly improving the prediction performance. At the same time, based on the patch division of historical cellular traffic data, the local semantic information of the time series can be preserved, thereby enhancing the local features of the time series, capturing comprehensive semantic information, and reducing the time complexity of the model.

[0102] When N≥2, the first target object is the intermediate feature map obtained by performing steps S1-S3 last time. At the same time, the patch parameters used for patching at this time are determined according to the patch parameters used for performing steps S1-S3 last time.

[0103] The number of patches used when dividing the first target object into multiple patches is determined in the following manner:

[0104]

[0105] Wherein, N m represents the number of patches when performing steps S1-S3 for the mth time, N m-1 represents the number of patches when performing steps S1-S3 for the (m-1)th time, P m-1 represents the patch length when performing steps S1-S3 for the (m-1)th time, S m-1 represents the step length of the non-overlapping area between two consecutive patches when performing steps S1-S3 for the (m-1)th time, and m is an integer, 2≤m≤N.

[0106] In this case, the first target object is divided into multiple patches, and linear mapping is performed on the multiple patches to obtain the patch sequence.

[0107] S2: Position encoding is performed on the patch sequence to obtain a position encoding result.

[0108] For a time series, the position of data at different time points in the entire sequence is important. If the order of data points in the time series is disturbed, the semantic information of the entire time series will change. Therefore, in the embodiments of the present application, position encoding is introduced to explicitly encode the position information in the time series.

[0109] Specifically, in this step, each patch in the patch sequence is linearly mapped to obtain high-dimensional features of the patch sequence, and position encoding is performed based on the position information of each patch in the patch sequence and the high-dimensional features to obtain the position encoding result.

[0110] Wherein, the position encoding result can be expressed as:

[0111]

[0112] wherein, is a trainable linear mapping that maps each chunk to a D (D>2, typically 32, 64, 128, etc., is a tunable hyperparameter) dimensional vector space for better learning of its high-dimensional feature representation; is the final input data for the Transformer encoder model; is the triangular position encoding, which is calculated as follows:

[0113]

[0114] wherein, pos is the position of the current chunk in the univariate time series it belongs to, and l is the vector dimension.

[0115] S3: input the position encoding result into the transformer network and run the transformer network to obtain the intermediate feature map of the first univariate time series.

[0116] The data after the above grouping operation, high-dimensional mapping and position encoding can be input into the backbone network based on the Transformer encoder to learn its time characteristics.

[0117] In this step, the position encoding result is mapped into multiple subspaces by using the multi-head attention mechanism to obtain the attention output result, and the attention output result is processed by normalization operation and feedforward neural network to obtain the intermediate feature map.

[0118] In the embodiments of the present application, the Transformer encoder is composed of the multi-head attention mechanism and the feedforward neural network, and the calculation process of a standard Transformer block is as follows:

[0119] z′ l = z l-1 + MSA(LN(z l-1 ))

[0120] z l = z′ l + MLP(LN(z′ l ))

[0121] wherein, z l-1 and z l are the input and output of the l-th layer Transformer block, z′ l represents an intermediate variable, MSA represents the multi-head attention mechanism, LN represents the Layer Normalization, and MLP is the feedforward neural network.

[0122] Here, the multi-head attention mechanism is adopted to map the time series data into multiple subspaces, so that the model pays attention to different aspects of information, and then the obtained information is spliced to improve the model's representation ability for cellular data. For each head First, the input is mapped to Query, Key and Value matrices through linear transformation, and the calculation method is as follows:

[0123]

[0124]

[0125]

[0126] wherein, and are learnable parameters (d k represents the dimension of the hidden vector). After linear transformation, the attention output between blocks is learned through the scaled dot-product attention operation The calculation formula is as follows:

[0127]

[0128] For the same input The multi-head attention mechanism defines multiple sets of different and generate different and learn different parameters to improve the modeling ability of the model. After the output of the multi-head attention mechanism is normalized and passed through a feedforward neural network, the feature map F of the current stage is generated i If the current stage is the last stage of the entire processing, F i can be used as the input of step 104, otherwise, the above process can be returned to step S1 and executed again.

[0129] As can be seen from the above description, the feature pyramid structure is introduced into the Transformer model, and the multi-scale feature map is learned through multiple length blocks, and then the multi-scale features of the time series are learned, and the influence size between different blocks is learned through the attention mechanism, which can fully capture the time dimension features and enhance the fitting ability of the model for cellular traffic data.

[0130] Step 104, obtaining the predicted value of each first univariate time series according to the feature map.

[0131] In this step, the feature map is processed by a linear prediction head to obtain the predicted value for each of the first univariate time series.

[0132] Specifically, the obtained feature map is passed through a linear prediction head to output the final prediction result. In this process, the mean squared error (MSE) between the predicted and actual values ​​is used as the loss function, and it is minimized through backpropagation. The calculation formula is as follows:

[0133]

[0134] in, Represents the true value. Indicates the predicted value. This represents the overall expected value.

[0135] Step 105: Obtain predicted cellular traffic data based on the predicted values ​​of each of the first univariate time series.

[0136] In this step, the predicted values ​​of each of the first univariate time series are concatenated to obtain the predicted cellular traffic data.

[0137] In this embodiment, for each univariate time series obtained from multiple univariate time series data obtained from historical cellular traffic data, a corresponding feature map is obtained. Based on this feature map, a predicted value for each univariate time series is obtained, and thus, predicted cellular traffic data can be derived from the multiple predicted values. Since cellular traffic data possesses characteristics such as being time series data, the solution in this embodiment can fully consider these characteristics of cellular traffic data through the feature map, thereby improving the accuracy of cellular traffic data prediction.

[0138] The following, combined with Figure 2 Describe in detail the process of obtaining feature maps. The example given is N=4.

[0139] A0: Obtain historical cellular traffic data and convert it into multiple univariate time series.

[0140] A1: For the first target object in the input, divide it into overlapping or non-overlapping patches. The length of each patch is P, and the step size of the non-overlapping region between two consecutive patches is S. After the patching operation, a patch sequence is generated:

[0141] When N=1, the first target object is the first univariate time series, which is any one of multiple univariate time series.

[0142] A2: position encoding the block sequence to obtain a position encoding result.

[0143] In this step, the block sequence is mapped to a high-dimensional feature, and the position encoding is added to retain the position information. The specific process can refer to the description of S2.

[0144] A3: taking the position encoding result as the input of the transformer network and running the transformer network to obtain the intermediate feature map of the first univariate time series.

[0145] A4: determining whether N is equal to 4, if N = 4, performing A5; otherwise, determining the block parameter for the N+1th time, adding 1 to the value of N, and returning to A1 to re-execute A1-A4.

[0146] wherein the block parameter for the N+1th time is determined as follows:

[0147] When generating the feature map, the embodiment of the application proposes a progressive shrinking strategy, which controls the size of the generated feature map by controlling the size of the block.

[0148] Suppose that when N = m, the length of the block is P m , and the number of blocks is N m . When m = 1, the input is the original T x M-dimensional first univariate time series data, and after the transformer encoder, the feature map is generated. In the subsequent mth stage, the feature map F m-1 obtained in the m-1th stage is used as input, and it is divided into blocks, and each block is mapped to D m dimensions through linear mapping. After the block operation, the feature map of each stage is approximately P m times smaller than the feature map of the previous stage. In this way, the feature map scale can be flexibly adjusted at each stage, thereby constructing a feature pyramid for the backbone network based on the transformer encoder and learning the multi-scale features of the time series.

[0149] A5: obtaining the predicted cellular traffic data according to the feature map of each first univariate time series.

[0150] As can be seen from the above description, the scheme of the embodiment of the application is simple, has strong generalization, and can be applied to a longer input sequence, learn the long-term dependence relationship of the cellular traffic data, and significantly improve the prediction accuracy when applied to the cellular traffic prediction problem.

[0151] Referring to Figure 3 , Figure 3Figure 1 is a structural diagram of a cellular traffic data prediction device provided by an embodiment of the present application. As shown in Figure 1, the cellular traffic data prediction device comprises: Figure 3

[0152] A first obtaining module 301 is configured to obtain historical cellular traffic data.

[0153] A second obtaining module 302 is configured to obtain a plurality of univariate time series according to the historical cellular traffic data.

[0154] A third obtaining module 303 is configured to obtain a feature map of a first univariate time series in the plurality of univariate time series.

[0155] A fourth obtaining module 304 is configured to obtain a predicted value of each of the first univariate time series according to the feature map.

[0156] A fifth obtaining module 305 is configured to obtain predicted cellular traffic data according to the predicted value of each of the first univariate time series.

[0157] Optionally, the third obtaining module is further configured to perform steps S1-S3 for N times in a loop, and take an output of a converter network obtained by performing S1-S3 for the Nth time as the feature map of the first univariate time series, where N is an integer greater than or equal to 2.

[0158] S1: divide a first target object into a plurality of patches to obtain a patch sequence, where when N=1, the first target object comprises the first univariate time series, or when N≥2, the first target object comprises an intermediate feature map obtained by performing steps S1-S3 for the N-1th time, and a patch number adopted when the first target object is divided into a plurality of patches is determined according to a patch parameter adopted when the first target object is divided into a plurality of patches in the N-1th time.

[0159] S2: positionally encode the patch sequence to obtain a positionally encoded result.

[0160] S3: take the positionally encoded result as an input of the converter network and run the converter network to obtain an intermediate feature map of the first univariate time series.

[0161] Optionally, when N≥2, the third obtaining module is further configured to divide the first target object into a plurality of patches and perform linear mapping on the plurality of patches to obtain the patch sequence.

[0162] Optionally, the third obtaining module is further configured to:

[0163] ​Linearly mapping each patch in the patch sequence to obtain a high-dimensional feature of the patch sequence;

[0164] Encoding a position based on position information of each patch in the patch sequence and the high-dimensional feature to obtain a position encoding result.

[0165] Optionally, the third obtaining module is further configured to:

[0166] Mapping the position encoding result into multiple subspaces using a multi-head attention mechanism to obtain an attention output result;

[0167] Processing the attention output result through a normalization operation and a feedforward neural network to obtain the intermediate feature map.

[0168] Optionally, when N is greater than or equal to 2, the number of patches used when the first target object is divided into multiple patches is determined by the following method:

[0169]

[0170] wherein, N m represents the number of patches when the steps S1 to S3 are performed for the mth time, N m-1 represents the number of patches when the steps S1 to S3 are performed for the (m-1)th time, P m-1 represents the patch length when the steps S1 to S3 are performed for the (m-1)th time, S m-1 represents the step length of the non-overlapping region between two consecutive patches when the steps S1 to S3 are performed for the (m-1)th time, and m is an integer and 2≤m≤N.

[0171] Optionally, the fourth obtaining module is further configured to:

[0172] Processing the feature map through a linear prediction head to obtain a predicted value of each first univariate time series.

[0173] The apparatus provided by the embodiments of the present application can execute the method embodiments, and the implementation principles and technical effects are similar, and will not be described here again.

[0174] Referring to Figure 4 , Figure 4 is a structural diagram of a cellular traffic data prediction apparatus provided by the embodiments of the present application. As shown in Figure 4 , the cellular traffic data prediction apparatus includes a processor 401 and a transceiver 402, wherein the processor 401 is configured to:

[0175] obtain historical cellular traffic data;

[0176] obtain a plurality of univariate time series according to the historical cellular traffic data;

[0177] For a first univariate time series in the plurality of univariate time series, obtain a feature map of the first univariate time series;

[0178] According to the feature map, obtain a predicted value of each of the first univariate time series;

[0179] According to the predicted value of each of the first univariate time series, obtain predicted cellular traffic data.

[0180] Optionally, the processor 401 is further configured to:

[0181] Steps S1 to S3 are executed cyclically N times, and the output of the transformer network obtained by the Nth execution of S1 to S3 is taken as the feature map of the first univariate time series, where N is an integer greater than or equal to 2:

[0182] S1: divide a first target object into a plurality of patches to obtain a patch sequence, wherein when N = 1, the first target object includes the first univariate time series, or when N ≥ 2, the first target object includes an intermediate feature map obtained by the N-1th execution of steps S1 to S3, and the number of patches used when the first target object is divided into a plurality of patches is determined according to the patch parameters used when the first target object is divided into a plurality of patches in the N-1th execution of steps S1 to S3;

[0183] S2: positionally encode the patch sequence to obtain a positionally encoded result;

[0184] S3: take the positionally encoded result as the input of a transformer network and run the transformer network to obtain an intermediate feature map of the first univariate time series.

[0185] Optionally, the processor 401 is further configured to, when N ≥ 2, divide a first target object into a plurality of patches and linearly map the plurality of patches to obtain the patch sequence.

[0186] Optionally, the processor 401 is further configured to:

[0187] Linearly map each patch in the patch sequence to obtain a high-dimensional feature of the patch sequence;

[0188] Positionally encode based on the position information of each patch in the patch sequence and the high-dimensional feature to obtain the positionally encoded result.

[0189] Optionally, the processor 401 is further configured to:

[0190] Map the positionally encoded result into a plurality of subspaces using a multi-head attention mechanism to obtain an attention output result;

[0191] The attention output result is processed through a normalization operation and a feedforward neural network to obtain the intermediate feature map.

[0192] Optionally, when N≥2, the number of blocks used when the first target object is divided into blocks is determined by the following method:

[0193]

[0194] wherein, N m represents the number of blocks when the steps S1-S3 are executed for the mth time, N m-1 represents the number of blocks when the steps S1-S3 are executed for the (m-1)th time, P m-1 represents the block length when the steps S1-S3 are executed for the (m-1)th time, S m-1 represents the step length of the non-overlapping area between two consecutive blocks when the steps S1-S3 are executed for the (m-1)th time, and m is an integer, 2≤m≤N.

[0195] Optionally, the processor 401 is further configured to:

[0196] The feature map is processed by a linear prediction head to obtain a predicted value of each first univariate time series.

[0197] The device provided in the embodiments of the present application can execute the above-mentioned method embodiments, and the implementation principles and technical effects are similar, and the present embodiment will not be described here.

[0198] It should be noted that the division of units in the embodiments of the present application is illustrative, and is only a logical functional division. In actual implementation, another division method can be used. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present alone, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0199] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a processor-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0200] The embodiment of the present application provides a communication device, including: a memory, a processor and a program stored in the memory and executable on the processor; the processor is used for reading the program in the memory to realize the steps in the cellular traffic data prediction method.

[0201] The embodiment of the present application also provides a readable storage medium, and the readable storage medium stores a program. The program is executed by a processor to realize the processes of the cellular traffic data prediction method, and the same technical effects can be achieved. To avoid repetition, details are not described herein. The readable storage medium can be any available medium or data storage device that can be accessed by the processor, including but not limited to a magnetic memory (for example, a floppy disk, a hard disk, a magnetic tape, a magneto-optical disk (MO) and the like), an optical memory (for example, a CD, a DVD, a BD, a HVD and the like), and a semiconductor memory (for example, a ROM, an EPROM, an EEPROM, a non-volatile memory (NAND FLASH), a solid state disk (SSD)) and the like.

[0202] It should be noted that, in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed, or further include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0203] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and necessary general hardware platform, of course, also can be realized by hardware, but in many cases, the former is a better embodiment. According to such understanding, the technical scheme of the present application can be embodied in the form of software product, and the computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk), including a plurality of instructions to make a terminal (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) execute the method described in various embodiments of the present application.

[0204] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative, not limiting, and those skilled in the art can make many forms without departing from the purpose of the present application and the scope protected by the claims under the inspiration of the present application, which all belong to the protection of the present application.

Claims

1. A method for predicting cellular traffic data, characterized in that, The method comprises: obtaining historical cellular traffic data; obtaining a plurality of univariate time series according to the historical cellular traffic data; for a first univariate time series in the plurality of univariate time series, obtaining a feature map of the first univariate time series; obtaining a predicted value of each of the first univariate time series according to the feature map; obtaining predicted cellular traffic data according to the predicted value of each of the first univariate time series; wherein the obtaining of the feature map of the first univariate time series comprises: cyclically performing steps S1 to S3 for N times, and taking the output of the transformer network obtained by the Nth time of performing S1 to S3 as the feature map of the first univariate time series: S1: dividing a first target object into a plurality of patches to obtain a patch sequence, wherein when N = 1, the first target object comprises the first univariate time series, or when N ≥ 2, the first target object comprises an intermediate feature map obtained by the N-1th time of performing steps S1 to S3, and the number of patches used when the first target object is divided into a plurality of patches is determined according to the patch parameters used when the first target object is divided into a plurality of patches in the N-1th time of performing steps S1 to S3; S2: positionally encoding the patch sequence to obtain a positionally encoded result; S3: taking the positionally encoded result as the input of a transformer network and running the transformer network to obtain an intermediate feature map of the first univariate time series.

2. The method of claim 1, wherein, When N ≥ 2, the dividing of the first target object into a plurality of patches to obtain a patch sequence comprises: dividing the first target object into a plurality of patches and performing linear mapping on the plurality of patches to obtain the patch sequence.

3. The method of claim 1, wherein, The positionally encoding of the patch sequence to obtain a positionally encoded result comprises: performing linear mapping on each patch in the patch sequence to obtain a high-dimensional feature of the patch sequence; positionally encoding based on the position information of each patch in the patch sequence and the high-dimensional feature to obtain the positionally encoded result.

4. The method of claim 1, wherein, The taking of the positionally encoded result as the input of a transformer network and the running of the transformer network to obtain an intermediate feature map of the first univariate time series comprises: mapping the positionally encoded result into a plurality of subspaces using a multi-head attention mechanism to obtain an attention output result; processing the attention output result through a normalization operation and a feedforward neural network to obtain the intermediate feature map.

5. The method of claim 1, wherein, When N ≥ 2, the number of patches used when the first target object is divided into a plurality of patches is determined in the following manner: ； wherein, denotes the number of blocks at the mth execution of steps S1 to S3, denotes the number of blocks at the m-1th execution of steps S1 to S3, denotes the block length at the m-1th execution of steps S1 to S3, denotes the step size of the non-overlapping region between two consecutive blocks at the m-1th execution of steps S1 to S3, m is an integer, 2≤m≤N.

6. The method of claim 1, wherein, The obtaining of a predicted value of each of the first univariate time series according to the feature map comprises: processing the feature map through a linear prediction head to obtain the predicted value of each of the first univariate time series.

7. A cellular traffic data prediction apparatus characterized by comprising: The method comprises: a first obtaining module configured to obtain historical cellular traffic data; a second obtaining module configured to obtain a plurality of univariate time series according to the historical cellular traffic data; a third obtaining module configured to, for a first univariate time series in the plurality of univariate time series, obtain a feature map of the first univariate time series; a fourth obtaining module, configured to obtain a predicted value of each of the first univariate time series according to the feature map; a fifth obtaining module, configured to obtain predicted cellular traffic data according to the predicted value of each of the first univariate time series; wherein the third obtaining module is further configured to: execute steps S1 to S3 for N times in a loop, and take the output of the converter network obtained by executing S1 to S3 for the Nth time as the feature map of the first univariate time series; S1: divide a first target object into a plurality of patches to obtain a patch sequence, wherein when N = 1, the first target object comprises the first univariate time series, or when N ≥ 2, the first target object comprises an intermediate feature map obtained by executing steps S1 to S3 for the (N-1)th time, and the number of patches used when dividing the first target object into a plurality of patches is determined according to the patch parameter used when dividing the first target object into a plurality of patches in the (N-1)th execution of steps S1 to S3; S2: perform position encoding on the patch sequence to obtain a position encoding result; S3: take the position encoding result as the input of the converter network and run the converter network to obtain an intermediate feature map of the first univariate time series.

8. A cellular traffic data prediction apparatus characterized by comprising: comprise: a processor and a transceiver, wherein the processor is configured to: obtain historical cellular traffic data; obtain a plurality of univariate time series according to the historical cellular traffic data; for a first univariate time series in the plurality of univariate time series, obtain a feature map of the first univariate time series; obtain a predicted value of each of the first univariate time series according to the feature map; obtain predicted cellular traffic data according to the predicted value of each of the first univariate time series; wherein the processor is further configured to: execute steps S1 to S3 for N times in a loop, and take the output of the converter network obtained by executing S1 to S3 for the Nth time as the feature map of the first univariate time series; S1: divide a first target object into a plurality of patches to obtain a patch sequence, wherein when N = 1, the first target object comprises the first univariate time series, or when N ≥ 2, the first target object comprises an intermediate feature map obtained by executing steps S1 to S3 for the (N-1)th time, and the number of patches used when dividing the first target object into a plurality of patches is determined according to the patch parameter used when dividing the first target object into a plurality of patches in the (N-1)th execution of steps S1 to S3; S2: perform position encoding on the patch sequence to obtain a position encoding result; S3: take the position encoding result as the input of the converter network and run the converter network to obtain an intermediate feature map of the first univariate time series.

9. A communication device comprising: a memory, a processor, and a program stored in the memory and capable of running on the processor; and the processor is configured to read the program in the memory to implement the steps in the cellular traffic data prediction method according to any one of claims 1 to 6.

10. A readable storage medium for storing a program, characterized in that, The program, when executed by the processor, implements the steps in the method for predicting cellular traffic data according to any one of claims 1 to 6.

Images