Cellular traffic prediction method and system based on multilayer meta-learning model

Abstract

The invention relates to the technical field of communication, and particularly discloses a cellular traffic prediction method and system based on a multilayer meta-learning model. S01, a multi-layer element learning model is constructed, and the first layer is a main model and is used for generating weights according to frequency domain information of cellular network flow loads; the second layer is a weight distribution model and is used for mapping a corresponding sub-model according to the weight; the third layer is a sub-model and is used for predicting the traffic load of the next time period; s02, collecting cellular network flow load data, preprocessing the cellular network flow load data, and training the main model and the sub-model to obtain a trained multi-layer meta-learning model; and S03, performing cellular network flow load prediction by using the trained multilayer meta-learning model, calculating a flow load prediction value, and obtaining prediction load information.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular, to a cellular traffic prediction method and system based on a multi-layer meta-learning model. Background Art

[0002] With the rapid development of mobile communication technologies and applications such as the Internet of Things (IoT), cloud computing, and virtual reality / augmented reality, the traffic in mobile networks has shown an exponential growth trend. To meet the growing user demands and reduce the operating expenditure (OpEx), cellular network traffic load prediction plays a crucial role in mobile networks because many mobile applications rely on real-time or near-real-time traffic analysis of the radio access network. Traffic prediction methods are mainly divided into two categories: one is the time series model based on historical traffic statistics (such as ARIMA, exponential smoothing, etc.), and its defect is that it cannot effectively capture the spatio-temporal correlation of traffic loads between multiple base stations and the impact of sudden events; the other is the prediction framework based on machine learning (such as LSTM, CNN, etc.). Although it can extract local spatio-temporal features, it has insufficient collaborative modeling ability for multi-dimensional data (such as user movement trajectories, application type distributions, network slice requirements) of cross-regional base station groups, and there are problems such as high computational complexity and poor real-time performance. In addition, existing solutions generally lack the ability to adaptively model the dynamic topology of the network (such as the switching of base station sleep / activation states), resulting in a significant decrease in prediction accuracy in dynamic scenarios. Therefore, if the future traffic loads of multiple base stations can be accurately predicted, sleep strategies can be implemented for the corresponding base stations according to the future network traffic loads to reduce energy consumption, and network resources can be optimized before improving the service experience of mobile users, which becomes a key technical challenge for improving the intelligence level of mobile networks. Summary of the Invention

[0003] In view of the above problems, the first object of the present invention is to provide a cellular traffic prediction method based on a multi-layer meta-learning model. By constructing a multi-layer meta-learning model, using multiple sub-models for parallel prediction, and fusing the prediction results, the overall accuracy of the model is improved. A correction mechanism based on prediction error is added to make the entire model have better robustness.

[0004] The second object of the present invention is to provide a cellular traffic prediction system based on a multi-layer meta-learning model.

[0005] To achieve the first object, the first technical solution of the present invention is: A cellular traffic prediction method based on a multi-layer meta-learning model, including: Step S01: Construct a multi-layer meta-learning model, where: the first layer is the main model, which is used to generate weights according to the frequency domain information of the cellular network traffic load; the second layer is the weight allocation model, which is used to map the corresponding sub-models according to the weights; the third layer is the sub-model, which is used to predict the traffic load in the next time period; Step S02: Collect the cellular network traffic load data, preprocess the cellular network traffic load data, and use it to train the main model and the sub-models to obtain a trained multi-layer meta-learning model; Step S03: Use the trained multi-layer meta-learning model to predict the cellular network traffic load, calculate the traffic load prediction value, and obtain the predicted load information.

[0006] Preferably, in step S01, the number of the sub-models is multiple, and the sub-models are composed of multiple GRU neural networks.

[0007] Preferably, in step S01, the main model includes a DNN network, and the DNN network is used to reduce the dimension of the frequency domain information of the cellular network traffic load to obtain low-dimensional samples, decode the low-dimensional samples to obtain decoded samples, calculate the reconstruction error between the decoded samples and the frequency domain information, and obtain output samples, where the output samples include the reconstruction error and the low-dimensional samples.

[0008] Preferably, in step S01, the main model further includes a GMM network, and the GMM network is used to perform density estimation on the output samples of the DNN network and obtain weights according to the density estimation, where the weights include the weights of the frequency domain information of the cellular network traffic load in the corresponding sub-models.

[0009] Using multiple sub-models for parallel prediction provides a basis for improving the overall accuracy of the subsequent model. Soft weights are used to fuse the prediction results of multiple prediction sub-models instead of only adopting the prediction result of a certain sub-model, which improves the model accuracy.

[0010] Preferably, in step S02, the training of the main model is carried out according to a constraint function, and the constraint function is: ; ; Wherein, is the encoder parameter, is the decoder parameter, is the reconstruction error, is the output sample, is the th output sample of the DNN network of the th cellular, is the th output of the DNN network of the th cellular, is a minimum value, is the mixing probability, is the mean, is the variance, exp() represents the natural exponential function, and are constant coefficients respectively, is the number of input samples, is the output the th vector, is the th output sample of the DNN network of the the th vector of the anomaly degree.

[0011] Preferably, in step S02, the training of the sub-model includes training using different types of cellular network traffic load data, and the cellular network traffic load data includes real-time video data and non-video data.

[0012] Preferably, in step S02, the preprocessing of the cellular network traffic load data includes: normalizing the collected cellular network traffic load data by using the min-max normalization method to obtain a traffic load vector, and processing the traffic load vector by using the fast Fourier transform method to generate the frequency domain information of the cellular network traffic load, and the frequency domain information is used to train the main model and the sub-model.

[0013] Preferably, it further includes step S04, judging the prediction error between the predicted cellular network traffic load at the previous moment and the real cellular network traffic load. When the prediction error is greater than or equal to the error threshold, retrain the main model, update the weights, and re-output the predicted load information according to the weights. A correction mechanism based on the prediction error is added to make the whole model have better robustness.

[0014] To achieve the second object, the second technical solution of the present invention is: a cellular traffic prediction system based on a multi-layer meta-learning model, including: Model construction module: used to construct a multi-layer meta-learning model, where: the first layer is the main model, which is used to generate weights according to the frequency domain information and load information of the cellular network traffic load; the second layer is the weight allocation model, which is used to map the corresponding sub-model according to the weights; the third layer is the sub-model, which is used to predict the traffic load in the next time period; Model training module: used to collect cellular network traffic load data, preprocess the cellular network traffic load data, and used to train the main model and the sub-model to obtain a trained multi-layer meta-learning model; Result output module: used to predict the cellular network traffic load by using the trained multi-layer meta-learning model, calculate the traffic load prediction value, and obtain and output the predicted load information.

[0015] Preferably, it further includes a prediction error and correction module: which is used to judge whether to output the predicted load information according to the prediction error between the predicted load information at the previous moment and the actual load; When the prediction error at the previous moment is greater than or equal to the error threshold, retrain the main model to update the weights and output the predicted load information again.

[0016] Beneficial effects of the above technical solution: The cellular traffic prediction method and system based on the multi-layer meta-learning model provided by the present invention achieve a significant improvement in the prediction accuracy and robustness of cellular network traffic load. The present invention constructs multiple sub-models to parallelly process features such as the frequency domain information of cellular network traffic load, time series traffic data, and actual load information, and effectively performs traffic load prediction. Compared with the limitation of a single model relying on local feature extraction, the present invention utilizes the complementarity between sub-models to provide a more comprehensive feature expression basis for the subsequent fusion stage, improving the prediction accuracy. Adopting a dynamic weight allocation strategy, according to the prediction confidence of each sub-model in different spatio-temporal scenarios, the output results of multiple models are weighted and fused to avoid the error accumulation caused by the deviation of a single model. Introducing a prediction error and correction mechanism makes the entire model have better robustness. By real-time monitoring the prediction deviation of each sub-model, the weights of the sub-models are adjusted to effectively suppress the impact of abnormal traffic fluctuations on the overall prediction. Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 Flowchart of the cellular traffic prediction method based on the multi-layer meta-learning model provided by an embodiment of the present invention; Figure 2 Structural diagram of the multi-layer meta-learning model provided by an embodiment of the present invention; Figure 3 Flowchart of the operation of the multi-layer meta-learning model provided by an embodiment of the present invention; Figure 4 Structural diagram of the main model in the multi-layer meta-learning model provided by an embodiment of the present invention; Figure 5 Relationship diagram between R2 and K values of the multi-layer meta-learning model provided by an embodiment of the present invention; Figure 6 Relationship diagram between MAE and K values of the multi-layer meta-learning model provided by an embodiment of the present invention. Detailed implementation manners

[0019] The following further describes in detail the implementation manners of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than an exhaustive list of all embodiments. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0020] Terms such as "first", "second", etc. (if any) in the specification and claims are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0021] It should be understood that the term "and / or" used herein is only a kind of association relationship describing associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after.

[0022] Embodiment 1

[0023] The cellular traffic prediction method based on a multi-layer meta-learning model provided by the present invention is as Figure 1 shown, and includes: Step S01: Construct a multi-layer meta-learning model, where: the first layer is the main model, which is used to generate weights according to the frequency domain information of the cellular network traffic load; the second layer is the weight allocation model, which is used to map the corresponding sub-models according to the weights; the third layer is the sub-model, which is used to predict the traffic load in the next time period; Step S02: Collect cellular network traffic load data, preprocess the cellular network traffic load data, and use it to train the main model and the sub-model to obtain a trained multi-layer meta-learning model; Step S03: Use the trained multi-layer meta-learning model to predict the cellular network traffic load, calculate the traffic load prediction value, and obtain the predicted load information. In Step S04, judge the prediction error between the predicted cellular network traffic load at the previous moment and the actual cellular network traffic load. When the prediction error is greater than or equal to the error threshold, retrain the main model, update the weights, and re-output the predicted load information according to the weights.

[0024] By constructing a multi - layer meta - learning model, using multiple sub - models for parallel prediction, and fusing the prediction results, the overall accuracy of the model is improved. A correction mechanism based on prediction error is added to make the whole model have better robustness.

[0025] I. Construction of a multi - layer meta - learning model The structure of the multi - layer meta - learning model is as Figure 2 shown. The multi - layer meta - learning model of the present invention is divided into three layers. The first - layer main model mainly generates weights through the frequency - domain information of the cellular network traffic load. Since the frequency - domain information does not contain time scalars, it can be processed by a DNN network, and the soft classification weights are generated through the GMM simulated by the neural network. The second layer maps the weights to different sub - models, and the third layer uses different types of sub - models to predict the traffic load in the next time period. The sub - models need to learn the cellular network traffic load data, and a multi - layer prediction model based on the GRU network is used as the learner. Finally, the predicted load information is subtracted from the real load information to obtain the prediction error. When the prediction error is too large, the weights of the sub - models will be re - weighted. Figure 2 In it, FFT is the Fast Fourier Transform, DNN is the Deep Neural Network, GMM is the Gaussian Mixture Model composed of multi - layer neural networks, the GRU module is the Gated Recurrent Unit module which is an improved recurrent neural network, Live Video is a type of real - time video dataset, and Non Video is a type of non - video dataset.

[0026] The working process based on the multi - layer meta - learning model is as Figure 3 shown. It includes the following steps: Traffic data pre - processing To accelerate the training process of each cellular network traffic load prediction model, the elements of the initial traffic load vector of each cellular are normalized to the range of [0, 1] through the min - max normalization method: (1) where, is the initial network traffic load of the th cellular, is the initial network traffic load vector the value of the largest element in, is the initial network traffic load vector the value of the smallest element in, is the t - th element in the initial network traffic load vector, is the network traffic load vector of the th cellular after normalization processing.

[0027] The method of using FFT (Fast Fourier Transform) is used to generate the frequency - domain information of different cellular network traffic loads. It includes: assuming for the cellular The normalized network traffic load vector constructs a discrete periodic signal.

[0028] (2) Among them, is the normalized network traffic load vector of the cellular network after periodic processing. To facilitate the analysis of the cellular network traffic load, the entire time span of the data set is divided into continuous equal time intervals. represents the th time interval. In this embodiment, preferably T = 168 (one week is 168 hours).

[0029] (3) Among them, and are imaginary units, is the coefficient of the frequency component, is the fast Fourier transform process.

[0030] After calculation, the five frequency components in this embodiment are respectively (when ), (when ), (when ), (when ), and (when ). The five frequency components correspond to sine signals with a cellular period of one week, one day, 12 hours, 8 hours, and 6 hours for the th cellular network, which are used to represent the frequency domain information of the normalized network traffic load vector of each cellular network and are used for subsequent model training and prediction.

[0031] Sub - model pre - training The multi - layer meta - learning model provided by the present invention includes multiple sub - models. The sub - models are composed of K GRU neural networks, and each sub - model is in a parallel relationship. Each sub - model is trained using different types of network traffic load data. As Figure 2 shown, sub - model 1 is trained for real - time video data; sub - model k is trained for non - video data.

[0032] The forward propagation process during the training of a single GRU neural network is as follows (4) (5) (6) (7) Among them, is the input information, is the Sigmoid function, is the Hadamard product, is activated by the tanh function, is the update gate (controlling how much information from the previous moment enters the current state), is the reset gate (controlling how much of the previous moment's state is written into the current candidate state in). , , correspond to different learning parameters in different gates respectively, is the input state at time t-1, is the output state at time t.

[0033] Main model pre-training The main model is trained with the frequency-domain information of the cellular network traffic load. The main model consists of a DNN network (deep neural) and a GMM network (Gaussian mixture model). The structure of the main model is as Figure 4 shown. Figure 4 in is the frequency-domain information vector of the -th cell.

[0034] The DNN network consists of a deep autoencoder. The frequency-domain information is reduced in dimension by the encoder to obtain a low-dimensional sample , and then decoded by the decoder to obtain the sample .

[0035] (8) (9) Among them, are the encoder parameters, is the process of encoding with the encoder with as the parameter, are the decoder parameters, is the process of decoding with the decoder with as the parameter.

[0036] The output of the DNN network contains two variables, including: is the low-dimensional sample, is the reconstruction error ( and the reconstruction error between them).

[0037] (10) (11) Among them, is a 2D feature, is the Euclidean distance, is the cosine similarity.

[0038] (12) (13) The GMM network (Gaussian mixture model) is a clustering algorithm that performs density estimation by predicting the mixture membership of each sample. is the output of a multi-layer neural network with as parameters, is a K-dimensional vector, and K is also equal to the number of sub-models.

[0039] Given N data samples, for all 1 ≤ k ≤ K, the parameters in GMM are as follows: (14) (15) (16) Among them, is the mixing probability, is the mean, is the variance, is the th input sample under the th Gaussian mixture model component's density estimation.

[0040] The constraint function during the entire main model training process is: (17) (18) Among them, is the reconstruction error, is the th honeycomb's DNN network's th output 's degree of abnormality, is a minimum value (applied in the training process of the Gaussian mixture model to prevent the values on the diagonal of the covariance matrix from becoming 0, which would lead to the matrix being non-invertible), and are two constant coefficients respectively.

[0041] Finally, the output of the GMM network is the fitting coefficient , which is also the output of the entire main model. Suppose there are K sub-models: (19) is the frequency domain information vector of the th honeycomb cell and the sub-model weights of sub-models 1, 2, …, k (1 ≤ k ≤ K).

[0042] Traffic prediction After the main model and sub-models are trained, the output value of sub-model k (1 ≤ k ≤ K) is used to predict the traffic load of the th honeycomb cell at time interval t + 1.

[0043] Assume there are K sub-models. First, the frequency domain information vector of the th honeycomb cell is input into the main model to obtain the sub-model weights . Then, the traffic information at time t is input into the K sub-models simultaneously for parallel calculation to obtain the predicted value of the traffic load of the th honeycomb cell at time t + 1 .

[0044] (20) Multiply by to obtain the predicted load information, i.e., the predicted value of the traffic load .

[0045] (21) Prediction error and correction Calculate the distance between the true load information and the predicted load information, and adjust the sub-model weights.

[0046] (22) where is the mean absolute error (prediction error), is the true traffic load value of the honeycomb network at time t + 1, and is the predicted value of the traffic load of the honeycomb network at time t + 1.

[0047] If the prediction error ≥ error threshold ( can be set according to actual needs), the main model will be retrained, the fitting coefficient will be updated, and then the traffic load of the honeycomb network will be predicted again to calculate the prediction error.

[0048] If the prediction error < error threshold , the prediction of cellular network traffic load is completed, the prediction results are output, and the cellular traffic prediction based on the multi-layer meta-learning model is completed.

[0049] The present invention also provides a cellular network traffic load prediction system based on a multi-layer meta-learning model, including: Model construction module: used to construct a multi-layer meta-learning model, where: the first layer is the main model, used to generate weights according to the frequency domain information and load information of the cellular network traffic load; the second layer is the weight allocation model, used to map the corresponding sub-models according to the weights; the third layer is the sub-model, used to predict the traffic load in the next time period; Model training module: used to collect cellular network traffic load data, preprocess the cellular network traffic load data, and used to train the main model and the sub-model to obtain a trained multi-layer meta-learning model; Result output module: used to perform cellular network traffic load prediction using the trained multi-layer meta-learning model, calculate the cellular network traffic load prediction value, and obtain and output the predicted load information.

[0050] It also includes a prediction error and correction module: used to determine whether to output the predicted load information according to the prediction error between the predicted load information at the previous moment and the real load; When the prediction error at the previous moment is greater than or equal to the error threshold, retrain the main model to update the weights and re-output the predicted load information.

[0051] Simulation results of the cellular traffic prediction method based on the multi-layer meta-learning model I. Experimental environment and parameter settings The operating system of the experimental environment is Windows 10 (64-bit), the CPU is Intel Core i7-8750H, the memory is 32GB, and the hard disk is a 932GB mechanical hard disk. The development language uses Python 3.6.0, and the machine learning environment is numpy 1.19.5, tensorflow 1.2.0, pandas 1.1.5, torch 1.7.1, and Scikit-learn 1.0.

[0052] The BDC mobile network dataset is used in the simulation. BDC is a mobile network dataset provided by the "Big Data Challenge" project of Telecom Italia. Approximately 3 million mobile traffic records generated from November 1, 2013 to January 1, 2014 in Milan City are collected in the dataset. Specifically, the urban area of Milan is divided into 10,000 grids, and each grid has the same size of 235m × 235m. Each traffic record contains the timestamp of the record, the grid ID, and the mobile traffic load (i.e., the data volume).

[0053] Table 1 Parameter Settings

[0054] II. Simulation To evaluate the prediction effect, two indicators, the mean absolute error (MAE) and the coefficient of determination (R2), were adopted. MAE reflects the closeness between the predicted value and the true value, and R2 measures the fitting degree between the predicted value and the true value.

[0055] ; Among them, is the mean square error (prediction error), is the true traffic load value at time t + 1, is the traffic load predicted value at time t + 1.

[0056] ; Among them, R2 is the coefficient of determination, is the true traffic load value at time t + 1, is the traffic load predicted value at time t + 1. is the mean value of the true traffic load values.

[0057] In the simulation, 16 tests were conducted based on the BDC mobile network dataset, and the test results are as shown in Figure 5 and Figure 6 , among which, Figure 5 the horizontal axis is the number K of sub-models, and the vertical axis is the initial average R2 value. Figure 6 The horizontal axis is the number K of sub-models, and the vertical axis is the initial average MAE value. It can be seen that the more the number of sub-models, the smaller the error of the model, and the higher the fitting degree between the predicted value and the true data, that is, the better the performance of the model.

[0058] The model provided by the present invention was compared with other models in the prior art. When the training samples were uniformly 840, the performance comparison is shown in Table 2 below.

[0059] Table 2 MAK and R2 Simulation Results

[0060] The cellular traffic prediction method and system based on a multi-layer meta-learning model provided by the present invention achieve a significant improvement in the prediction accuracy and robustness of cellular network traffic load. The present invention constructs multiple sub-models to process features such as the frequency-domain information, time-series traffic data, and real load information of the cellular network traffic load in parallel, and effectively performs traffic prediction. Compared with the limitation of a single model relying on local feature extraction, the present invention utilizes the complementarity between sub-models to provide a more comprehensive feature expression basis for the subsequent fusion stage, thereby improving the prediction accuracy. A dynamic weight allocation strategy is adopted to weight and fuse the output results of multiple models according to the prediction confidence of each sub-model in different spatio-temporal scenarios, avoiding the error accumulation caused by the deviation of a single model. A prediction error and correction mechanism is introduced to make the entire model have better robustness. By monitoring the prediction deviation of each sub-model in real time, the weights of the sub-models are adjusted to effectively suppress the impact of abnormal traffic fluctuations on the overall prediction.

[0061] Obviously, the above-mentioned embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limiting the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is impossible to list all the implementation manners here. Any obvious changes or variations derived from the technical solutions of the present invention still fall within the protection scope of the present invention.

Claims

1. A cellular traffic prediction method based on a multi-layer meta-learning model, characterized in that, Including: Step S01: Construct a multi-layer meta-learning model, where: the first layer is the main model, which is used to generate weights according to the frequency-domain information of the cellular network traffic load; the second layer is the weight allocation model, which is used to map corresponding sub-models according to the weights; the third layer is the sub-model, which is used to predict the traffic load in the next time period; Step S02: Collect cellular network traffic load data, preprocess the cellular network traffic load data, and use it to train the main model and the sub-model to obtain a trained multi-layer meta-learning model; Step S03: Use the trained multi-layer meta-learning model to predict the cellular network traffic load, calculate the traffic load prediction value, and obtain the predicted load information.

2. The cellular traffic prediction method based on a multi-layer meta-learning model according to claim 1, wherein In step S01, the number of the sub-models is multiple, and the sub-models are composed of multiple GRU neural networks.

3. The cellular traffic prediction method based on a multi-layer meta-learning model according to claim 1, wherein In step S01, the main model includes a DNN network, which is used to reduce the dimension of the frequency-domain information of the cellular network traffic load to obtain a low-dimensional sample, decode the low-dimensional sample to obtain a decoded sample, calculate the reconstruction error between the decoded sample and the frequency-domain information, and obtain an output sample, where the output sample includes the reconstruction error and the low-dimensional sample.

4. The cellular traffic prediction method based on a multi-layer meta-learning model according to claim 3, wherein In step S01, the main model further includes a GMM network, which is used to perform density estimation on the output sample of the DNN network and obtain weights according to the density estimation, where the weights include the weights of the frequency-domain information of the cellular network traffic load in the corresponding sub-models.

5. The cellular traffic prediction method based on a multi-layer meta-learning model according to claim 4, wherein In step S02, the training of the main model is performed according to a constraint function, and the constraint function is: ； ； Among them, is the encoder parameter, is the decoder parameter, is the reconstruction error, is the output sample, is the th output sample of the DNN network of the th honeycomb, is the th output of the DNN network of the th honeycomb, is a minimum value, is the mixing probability, is the mean, is the variance, exp() represents the natural exponential function, and are constant coefficients respectively, is the number of input samples, is the th vector of the output, is the th output sample of the DNN network of the th vector of the th honeycomb.

6. The method for predicting cellular traffic based on a multi-layer meta-learning model according to claim 1, characterized in that, In step S02, the training of the sub-model includes using different types of cellular network traffic load data for training, where the cellular network traffic load data includes real-time video data and non-video data.

7. The cellular traffic prediction method based on a multi-layer meta-learning model according to claim 1, characterized in that In step S02, the preprocessing of the cellular network traffic load data includes: normalizing the collected cellular network traffic load data by using the min-max normalization method to obtain a traffic load vector, processing the traffic load vector by using the fast Fourier transform method to generate the frequency-domain information of the cellular network traffic load, and the frequency-domain information is used to train the main model and the sub-model.

8. The cellular traffic prediction method based on a multi-layer meta-learning model according to claim 1, characterized in that, It further includes step S04, judging the prediction error between the predicted cellular network traffic load at the previous moment and the real cellular network traffic load. When the prediction error is greater than or equal to the error threshold, re-train the main model, update the weights, and re-output the predicted load information according to the weights.

9. A cellular traffic prediction system based on a multi-layer meta-learning model, characterized in that, Including: Model construction module: used to construct a multi-layer meta-learning model, where: the first layer is the main model, which is used to generate weights according to the frequency-domain information and load information of the cellular network traffic load; the second layer is the weight allocation model, which is used to map corresponding sub-models according to the weights; the third layer is the sub-model, which is used to predict the traffic load in the next time period; Model training module: used to collect cellular network traffic load data, preprocess the cellular network traffic load data, and use it to train the main model and the sub-model to obtain a trained multi-layer meta-learning model; Result output module: used for predicting the cellular network traffic load using the trained multi-layer meta-learning model, calculating the traffic load prediction value, and obtaining and outputting the predicted load information.

10. The cellular traffic prediction system based on the multi-layer meta-learning model according to claim 9, characterized in that, It further includes a prediction error and correction module: used for judging whether to output the predicted load information according to the prediction error between the predicted load information at the previous moment and the actual load; When the prediction error at the previous moment is greater than or equal to the error threshold, retrain the main model to update the weights and output the predicted load information again.

Images