Training method and device of prediction model, electronic equipment and storage medium

Abstract

The application provides a training method and device of a prediction model, electronic equipment and a storage medium, comprising: obtaining a network index data set; performing ensemble empirical mode decomposition processing on the network index data set based on a first sliding window to determine a training data set, a test data set and a true value of the test data set; the true value of the test data set is determined based on data in the first sliding window after movement; training an initial prediction model by using the training data set, and testing the trained initial prediction model by using the test data set to obtain a prediction value; moving the first sliding window, and repeating the above process to obtain a target prediction model and a prediction value of each test data set; and evaluating and testing the target prediction model based on the true value and the prediction value of each test data set to obtain an evaluation test result. The application can effectively improve the training effect of the prediction model, and the target prediction model obtained by training has high prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a training method and device of a prediction model, an electronic device and a storage medium. BACKGROUND

[0002] At present, network index data has the characteristics of complex structure, large numerical span, and many influencing factors, so it is necessary to use a prediction model based on a neural network to automatically extract data features to fit the network prediction model. The LSTM (Long short-term memory) model, a variant of the RNN (Recurrent Neural Network) model, can be well applied to the field of time series prediction. According to related literature, the bidirectional LSTM (also known as Bi-LSTM) model is superior to the LSTM model in terms of accuracy and other aspects. The signal decomposition technology, EEMD (Ensemble Empirical Mode Decomposition), can decompose the time series into several stationary sequences, which are used as training data sets and test data sets to train the bidirectional LSTM model. The combination of the bidirectional LSTM model and the EEMD algorithm has a good prediction effect in predicting 5G network time series.

[0003] However, the EEMD algorithm uses global decomposition, that is, it considers the data in the test data set when decomposing the network index data. Generally, the data in the test data set should be unknown, so the EEMD algorithm has an information leakage problem. If the input training set with the problem is used for training, it will affect the prediction accuracy of the bidirectional LSTM model. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a training method and device of a prediction model, an electronic device and a storage medium, which can effectively improve the training effect of the prediction model and make the target prediction model obtained by training have high prediction accuracy.

[0005] In a first aspect, an embodiment of the present application provides a training method of a prediction model, comprising:

[0006] obtain a network index dataset; perform ensemble empirical mode decomposition on the network index dataset based on a first sliding window to determine a training dataset, a test dataset and a true value of the test dataset corresponding to the first sliding window; wherein the training dataset and the test dataset are determined based on data within the first sliding window, and the true value of the test dataset is determined based on data within the first sliding window after movement; train a pre-established initial prediction model using the training dataset corresponding to the first sliding window, and test the trained initial prediction model using the test dataset corresponding to the first sliding window to obtain a predicted value of the test dataset; move the first sliding window, and continue to perform ensemble empirical mode decomposition on the network index dataset based on the moved first sliding window until the network index dataset is completely decomposed to obtain a target prediction model and a predicted value of each test dataset; and evaluate the target prediction model based on the true value and the predicted value of each test dataset to obtain an evaluation test result corresponding to the target prediction model.

[0007] In an embodiment, performing ensemble empirical mode decomposition on the network index dataset based on a first sliding window to determine a training dataset, a test dataset and a true value of the test dataset corresponding to the first sliding window comprises: performing ensemble empirical mode decomposition on the network index dataset based on the first sliding window to obtain a plurality of intrinsic mode components corresponding to the first sliding window; for each intrinsic mode component, determining a training data subset and a test data subset corresponding to the intrinsic mode component based on a second sliding window; taking the training data subset corresponding to each intrinsic mode component as the training dataset corresponding to the first sliding window; and taking the test data subset corresponding to each intrinsic mode component as the test dataset corresponding to the first sliding window; and determining the first S data within the moved first sliding window in the network index dataset as the true value of the test dataset; wherein S is a sliding step of the first sliding window.

[0008] In an embodiment, the training data subset and the test data subset corresponding to the intrinsic modal component are determined based on a second sliding window, including: setting a second sliding window for the intrinsic modal component according to a window length in a second sliding window parameter; regarding each data in the intrinsic modal component located in the second sliding window as a first data subset corresponding to the intrinsic modal component; moving the second sliding window according to a sliding step in the second sliding window parameter, and regarding the first S data in the intrinsic modal component located in the second sliding window after the moving as a second data subset corresponding to the intrinsic modal component; wherein the second data subset is a true value set corresponding to the first data subset; regarding each data in the intrinsic modal component located in the second sliding window after the moving as a next first data subset corresponding to the intrinsic modal component, and continuing to move the second sliding window according to the sliding step in the second sliding window parameter until the second sliding window moves to the end of the intrinsic modal component, to obtain a plurality of first data subsets and a plurality of second data subsets; taking the last first data subset as a test data subset corresponding to the intrinsic modal component, and obtaining a training data subset corresponding to the intrinsic modal component based on the first data subsets except the last first data subset and each second data subset.

[0009] In an embodiment, the initial prediction model is trained by using the training data set corresponding to the first sliding window, and the trained initial prediction model is tested by using the test data set corresponding to the first sliding window to obtain the prediction value of the test data set, including: training the initial prediction model established in advance by using the training data subset corresponding to the first intrinsic modal component; testing the trained initial prediction model by using the test data subset corresponding to the first intrinsic modal component to obtain the prediction value of the test data subset; continuing to train the initial prediction model by using the training data subset corresponding to the next intrinsic modal component, and continuing to test the trained initial prediction model by using the test data subset corresponding to the next intrinsic modal component until the prediction value of each test data subset is obtained; taking the sum of the prediction values of each test data subset as the prediction value of the test data set corresponding to the first sliding window.

[0010] In an embodiment, the target prediction model is evaluated and tested based on the true value and the prediction value of each test data set to obtain an evaluation and testing result corresponding to the target prediction model, including: calculating the mean absolute error, the root mean square error, the mean absolute percentage error and the fitting degree of the target prediction model based on the true value and the prediction value of each test data set to obtain the evaluation and testing result corresponding to the target prediction model.

[0011] In an implementation, before the ensemble empirical mode decomposition processing is performed on the network indicator dataset based on the first sliding window, the method further includes: determining a correlation degree corresponding to each data in the network indicator dataset, and removing the data from the network indicator dataset if the correlation degree corresponding to the data is less than a preset correlation threshold; determining missing data in the network indicator dataset, filling the missing data, or removing the missing data from the network indicator dataset; and determining abnormal data in the network indicator dataset, and removing the abnormal data from the network indicator dataset.

[0012] In an implementation, the initial prediction model is an improved bidirectional long short-term memory model, and a random inactivation layer is arranged between specified layer neurons of the improved bidirectional long short-term memory model.

[0013] In a second aspect, an embodiment of the present application further provides a training device of a prediction model, including: a data acquisition module configured to acquire a network indicator dataset; a first sliding module configured to perform ensemble empirical mode decomposition processing on the network indicator dataset based on a first sliding window to determine a training dataset, a test dataset, and a true value of the test dataset corresponding to the first sliding window; wherein the training dataset and the test dataset are determined based on data in the first sliding window, and the true value of the test dataset is determined based on data in the first sliding window after moving; a training and testing module configured to train a pre-established initial prediction model using the training dataset corresponding to the first sliding window, and test the trained initial prediction model using the test dataset corresponding to the first sliding window to obtain a prediction value of the test dataset; a second sliding module further configured to move the first sliding window, and continue to perform ensemble empirical mode decomposition processing on the network indicator dataset based on the first sliding window after moving until the network indicator dataset is completely decomposed to obtain a target prediction model and the prediction value of each test dataset; and an evaluation and testing module configured to evaluate and test the target prediction model based on the true value and the prediction value of each test dataset to obtain an evaluation and testing result corresponding to the target prediction model.

[0014] In a third aspect, an embodiment of the present application further provides an electronic device including a processor and a memory, the memory storing computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of the first aspect.

[0015] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium storing computer executable instructions, when the computer executable instructions are invoked and executed by a processor, the computer executable instructions cause the processor to implement the method according to any one of the first aspect.

[0016] The embodiments of the present application provide a prediction model training method and device, an electronic device and a storage medium. First, a network index dataset is obtained, and then the network index dataset is processed by using ensemble empirical mode decomposition based on a first sliding window to determine a training dataset, a test dataset and a true value of the test dataset corresponding to the first sliding window, wherein the training dataset and the test dataset are determined based on data in the first sliding window, and the true value of the test dataset is determined based on data in the first sliding window after movement. Then, the initial prediction model is trained by using the training dataset corresponding to the first sliding window, and the initial prediction model after training is tested by using the test dataset corresponding to the first sliding window to obtain a prediction value of the test dataset. Further, the first sliding window is moved, and the network index dataset is processed by using ensemble empirical mode decomposition based on the first sliding window after movement. The above process is repeated until the network index dataset is decomposed to obtain a target prediction model and a prediction value of each test dataset. Finally, the target prediction model is evaluated based on the true value and the prediction value of each test dataset to obtain an evaluation test result corresponding to the target prediction model. The above method uses the sliding window to decompose the network index dataset, and processes the network index dataset by using ensemble empirical mode decomposition in the first sliding window to determine the training dataset and the test dataset based on data in the first sliding window, and determine the true value of the test dataset based on data in the first sliding window after movement. Since the data in the first sliding window is decomposed in sequence in a segmented manner, the true value of the test training set is not leaked in the network index dataset decomposition process, so that the information leakage problem of the ensemble empirical mode decomposition process can be effectively improved, and the training effect of the prediction model can be effectively improved, so that the target prediction model obtained by training has high prediction accuracy.

[0017] Additional features and advantages of the application will be set forth in the description that follows, and in part will be apparent from the description, or can be learned by practice of the application. The objectives and other advantages of the application will be realized and attained by the structure particularly pointed out in the description and claims.

[0018] In order to make the above-mentioned objectives, characteristics and advantages of the present application more apparent and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are referred to for detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0020] Figure 1 A flowchart of a training method of a prediction model provided for an embodiment of the present application is shown in the figure.

[0021] Figure 2 A structural diagram of a Bi-LSTM model provided for an embodiment of the present application is shown in the figure.

[0022] Figure 3 A diagram of Dropout provided for an embodiment of the present application is shown in the figure.

[0023] Figure 4 A diagram of an improved BiLSTM model provided for an embodiment of the present application is shown in the figure.

[0024] Figure 5 A flowchart of an EEMD-BiLSTM model provided for an embodiment of the present application is shown in the figure.

[0025] Figure 6 A flowchart of a sliding window type decomposition provided for an embodiment of the present application is shown in the figure.

[0026] Figure 7 A flowchart of a sliding window type EEMD-BiLSTM algorithm provided for an embodiment of the present application is shown in the figure.

[0027] Figure 8 A structural diagram of a training device of a prediction model provided for an embodiment of the present application is shown in the figure.

[0028] Figure 9 A structural diagram of an electronic device provided for an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below in conjunction with embodiments. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. 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.

[0030] Currently, the related technologies provide an empirical mode decomposition (EMD): the EMD overcomes the problem of non-adaptability of the base function in the wavelet transform, and the empirical mode decomposition does not need to analyze and study a signal in advance, but directly decomposes a signal into stationary intrinsic mode functions (IMF) through a sequence decomposition, each IMF component contains local characteristic information of different time scales of the original signal, and the signal x(t) decomposition is expressed as follows:

[0031]

[0032] wherein is the i th IMF component obtained through decomposition, is a signal residual component obtained after the n IMF components are screened out through decomposition, and the signal residual component often represents a direct current component or a trend of the signal.

[0033] The related technologies also provide an ensemble empirical mode decomposition (EEMD): the EEMD utilizes the statistical characteristics of white noise with uniform frequency distribution. On the basis of the EMD, the signal is added with Gaussian white noise, so that the signal has continuity in different scales, the characteristics of the extreme points of the signal are changed, the anti-mixed decomposition is promoted, and the mode mixing phenomenon can be effectively avoided. Moreover, the EEMD can also perform significance test on the IMF components by means of white noise set disturbance, analyze the period distribution of the energy spectrum density of each IMF component to determine the attribute, and thus select the required IMF component for analysis.

[0034] The EEMD algorithm steps are as follows:

[0035] I. In n trials, a new time sequence is constructed by adding a white noise sequence to the original signal sequence, N is the number of integrated trials;

[0036] II. Based on the EMD algorithm, the time sequence Y_n(t) polluted by noise can be decomposed into a sum of intrinsic mode functions IMF and a residual r, that is:

[0037]

[0038] In the formula, m is the iteration number of the sifting algorithm, and n is the number of integrations.

[0039] III. Steps I and II are repeated for N times, and in each iteration trial, different white noise sequences u(t) are added to the original signal sequence;

[0040] IV. The intrinsic mode function obtained through average integration of the n trials is obtained, and the final intrinsic mode function is obtained​​ :

[0041] .

[0042] On the basis of the EEMD algorithm, the related technology proposes to combine the bidirectional LSTM model with the EEMD algorithm to predict the 5G network time sequence, but the EEMD algorithm has an information leakage problem. Specifically, the process of the EEMD-BiLSTM prediction algorithm is to first decompose the original data into several IMF components by EEMD, then divide the training data set and the test data set, input the IMF component data in the training data set into the BiLSTM model to train and adjust the parameters, obtain the optimal prediction model, and then input the IMF data in the test data set into the model to train and predict. But this decomposition method is a global decomposition, taking the global minimum and maximum, that is, when EMD decomposition, the data to be divided into the test data set is also considered, which is equivalent to knowing the data to be predicted in the future. Obviously, this is unreasonable, and in the ideal state, the data in the test data set should be unknown when the model is trained, which causes the information leakage problem, and further affects the prediction accuracy of the bidirectional LSTM model.

[0043] Based on this, the present application provides a prediction model training method, device, electronic equipment and storage medium, which can effectively improve the training effect of the prediction model and make the target prediction model obtained by training have high prediction accuracy.

[0044] In order to facilitate the understanding of the present embodiment, first, a prediction model training method disclosed by the present embodiment is introduced in detail, referring to the flowchart of the prediction model training method shown in Figure 1 The method mainly includes the following steps S102 to S110:

[0045] Step S102, acquiring a network index data set. The network index data set can be a 5G video stream dynamic tracking data set collected by a mobile operator, which is composed of network key performance indicators (KPI, Key Performance Indicator), including channel related indicators, signal-to-noise ratio, and download / upload rate, etc.

[0046] Step S104, performing integrated empirical mode decomposition processing on the network index data set based on a first sliding window to determine the training data set, the test data set and the true value of the test data set corresponding to the first sliding window. The first sliding window can also be called a large sliding window, and the second sliding window can also be called a small sliding window. The integrated empirical mode decomposition processing is abbreviated as EEMD processing. The training data set and the test data set are determined based on the data in the first sliding window. The true value of the test data set is determined based on the data in the first sliding window after moving.

[0047] In an embodiment, a first sliding window can be set on the network index dataset according to a first sliding window parameter, and EEMD processing can be performed on the data in the first sliding window to obtain a plurality of intrinsic mode components corresponding to the first sliding window; further, a second sliding window can be set on the intrinsic mode components according to a second sliding window parameter, so as to divide the intrinsic mode components into a plurality of first data subsets and second data subsets based on the second sliding window, and then the first data subsets and the second data subsets are used to construct training data subsets and test data subsets corresponding to the intrinsic mode components, the training data subset corresponding to each intrinsic mode component in the first sliding window is taken as a training dataset corresponding to the first sliding window, and the test data subset corresponding to each intrinsic mode component in the first sliding window is taken as a test dataset corresponding to the first sliding window, and the first S data in the first sliding window after movement are taken as the true value corresponding to the test dataset, and S is the sliding step of the first sliding window and the second sliding window.

[0048] In step S106, the initial prediction model established in advance is trained by using the training dataset corresponding to the first sliding window, and the trained initial prediction model is tested by using the test dataset corresponding to the first sliding window to obtain the prediction value of the test dataset. The initial prediction model is an improved bidirectional long short-term memory model (Bi-LSTM model), and a random inactivation layer (Dropout) is arranged between the specified layer neurons of the improved bidirectional long short-term memory model.

[0049] In an embodiment, when the training data subset and the test data subset under one intrinsic mode component are determined, the initial prediction model can be trained by using the training data subset, and the trained initial prediction model can be tested by using the test data subset to obtain the corresponding prediction value, and the initial prediction model can be continuously trained and predicted by using the training data subset and the test data subset of the next intrinsic mode component, until the initial prediction model is trained and predicted by using the training data subset and the test data subset corresponding to each intrinsic mode component in the first sliding window, and the prediction values obtained by each test are added, so as to obtain the prediction value of the test dataset corresponding to the first sliding window.

[0050] Step S108, the first sliding window is moved, and the integrated empirical mode decomposition processing based on the first sliding window after the movement is continued until the network index dataset is completely decomposed, so as to obtain the target prediction model and the prediction value of each test dataset. In an embodiment, when the prediction value of the test dataset corresponding to the first sliding window is determined, the first sliding window can be moved according to the sliding step, and the above step S104 is executed for the data in the first sliding window after the movement, so as to obtain the training dataset, the test dataset and the true value of the test dataset corresponding to the first sliding window after the movement. The step S106 is executed again to obtain the prediction value of the test dataset corresponding to the first sliding window after the movement. The process is repeated until the network index data is completely decomposed, that is, the target prediction model and the prediction value of each first sliding window corresponding to the test dataset are obtained.

[0051] Step S110, based on the true value and the prediction value of each test dataset, the target prediction model is evaluated to obtain the evaluation test result corresponding to the target prediction model. In an embodiment, the mean absolute error, the root mean square error, the mean absolute percentage error and the fitting degree of the target prediction model can be calculated through the difference between the true value and the prediction value of each test dataset, and one or more of the mean absolute error, the root mean square error, the mean absolute percentage error and the fitting degree are used to evaluate the training effect of the prediction model.

[0052] The training method of the prediction model provided by the embodiment of the application uses the sliding window to decompose the network index dataset, performs the integrated empirical mode decomposition processing on the network index dataset in the first sliding window, determines the training dataset and the test dataset based on the data in the first sliding window, and determines the true value corresponding to the test dataset based on the data in the first sliding window after the movement. Since the data in the first sliding window is sequentially decomposed in a segmented manner, the true value of the test training set is not leaked in the network index dataset decomposition process, so that the information leakage problem of the integrated empirical mode decomposition processing can be effectively improved, and the training effect of the prediction model can be effectively improved, so that the target prediction model obtained by training has high prediction accuracy.

[0053] In an embodiment, the initial prediction model is an improved bidirectional long short-term memory model (Bi-LSTM model), and a random dropout layer (Dropout) is arranged between the specified layer neurons of the improved bidirectional long short-term memory model. The specified layer neurons can be the forward network layer and the backward network layer, or can be the Dense layer.

[0054] In actual application, the LSTM network is an extension of the RNN network, and the LSTM model is composed of the input word at the moment t , cell state , temporary cell state , hidden state , forget gate , memory gate , output gate Composition, on the basis of RNN model, increase "gate" structure, used to control information to stay, can consider long time data. Figure 2 A structure diagram of a Bi-LSTM model is shown, and the Bi-LSTM model is a combination of two LSTM models, one training forward input data and the other training input data in reverse, which can better capture the bidirectional data features.

[0055] Considering that the Bi-LSTM model has overfitting and complex co-adaptive relationship between neurons, the Bi-LSTM model is improved in the embodiment of the application. Specifically, Dropout can make each layer of neurons randomly select certain nodes and delete the forward and backward network connections in each iteration, as shown in a Dropout diagram. Figure 3 Each iteration process has different node combinations, resulting in different outputs, so it can be considered as a multi-model integration, which can capture more randomness.

[0056] In an optional implementation, the embodiment of the application adds Dropout with a dropout probability value p=0.5 in the forward network layer and the backward network layer of the BiLSTM model, and adds Dropout with p=0.3 in the two Dense layers. As shown in a diagram of an improved BiLSTM model, Figure 4 Take a network index data set including a download rate sequence (DL_bitrate) as an example, and specifically: (1) first set a first sliding window W1 for the download rate sequence; (2) perform EEDM decomposition on the download rate sequence in the first sliding window W1 to obtain a plurality of IMF components; (3) for each IMF component, set a second sliding window W2 for the IMF component to construct a training data subset and a test data subset corresponding to the IMF component, and train and test the improved BiLSTM model using the obtained training data subset and test data subset, the improved BiLSTM model outputs a test vector, the test vector is input to a fully connected layer, and the fully connected layer outputs a predicted value of the test data subset, and the above (3) is repeated until the predicted value of each IMF component corresponding to the test data subset is output; (4) move the first sliding window W1 and repeat (2) to (3) until the download rate sequence is decomposed.

[0057] In addition, the quality of data involves many factors, including accuracy, integrity, consistency, timeliness, credibility and interpretability, which directly determines the prediction and generalization ability of the target prediction model. Data preprocessing is to process the original dirty data in a corresponding manner to obtain standard, clean and continuous data. Therefore, data preprocessing is very necessary for data mining and analysis. In an embodiment, after obtaining the network indicator dataset, the network indicator dataset can be preprocessed according to the following (1) to (3), specifically:

[0058] (1) Determine the correlation of each data in the network indicator dataset. If the correlation of the data is less than a preset correlation threshold, the data is removed from the network indicator dataset. In an embodiment, the network indicator dataset is a multivariate dataset, and the prediction data in the network indicator dataset can be referred to as a prediction target variable, such as the download rate sequence (DL_bitrate) described above as a prediction target variable. According to the Pearson correlation coefficient, the correlation R of other influence characteristic variables and the prediction target variable (DL_bitrate) can be calculated. Assuming that the preset correlation threshold is set to 0.3, the data with a correlation R < 0.3 in the network indicator dataset can be deleted.

[0059] (2) Determine the missing data from the network indicator dataset, fill the missing data, or remove the missing data from the network indicator dataset. In an embodiment, the data with less missing values can be filled with average values, and the data with more missing values can be deleted.

[0060] (3) Determine the abnormal data from the network indicator dataset, and remove the abnormal data from the network indicator dataset. In an example, each data can be compared with a preset range. If the data exceeds the preset range, the data can be determined as abnormal data, and the abnormal data can be deleted. In another example, data containing a specified field can be determined as abnormal data, for example, data containing "2147483647" can be deleted.

[0061] According to the above steps (1) to (3), the network indicator dataset can be preprocessed to obtain 25700 pieces of 5G network data containing 15 characteristic variables (i.e., the preprocessed network indicator dataset). Part of the field name explanation is shown in Table 1:

[0062] Table 1

[0063]

[0064] After the network index data is preprocessed, the preprocessed network index data can be subjected to EEMD processing. For ease of understanding, the EEMD-BiLSTM model is first explained, as shown in a kind of EEMD-BiLSTM model flow chart Figure 5 Since the network data is non-stationary nonlinear complex data, and the network data value range is large, if it is directly input into the BiLSTM model for training and prediction, it can be imagined that the effect is often unsatisfactory, therefore, the original data set is first decomposed into several stationary IMF components using the EEMD method, the training data set and a test data are divided after the IMF components are normalized, then the training data set is input into the bidirectional long short-term memory (BiLSTM) model for learning, and a learning model is obtained, then the test data is input into the learned model for prediction, the predicted values of each IMF component are summed to obtain the final predicted value Y, the root mean square error, mean absolute error and other evaluation criteria are calculated to judge the prediction effect of the model.

[0065] Considering that the above EEMD decomposition is a one-time global decomposition of all data, there is an information leakage problem in the field of time series prediction, to solve this problem, the embodiment of the application proposes a sliding window EEMD decomposition method. For ease of understanding, the steps of performing ensemble empirical mode decomposition processing on the network index data set based on the first sliding window to determine the training data set corresponding to the first sliding window, the test data set and the true value of the test data set are first explained, and the steps 1 to 4 are specifically referred to as follows:

[0066] Step 1, perform ensemble empirical mode decomposition processing on the network index data set based on the first sliding window to obtain a plurality of intrinsic mode components corresponding to the first sliding window. In an embodiment, a large sliding window can be set according to the window length in the first sliding window parameter, and the data in the large sliding window is subjected to EEDM decomposition to obtain a plurality of IMF components corresponding to the large sliding window. For example, assuming that the network index data set is time series data, the time series data length is T, the large sliding window length is set to W1, and the sliding step is S, the window data [0: W1] in the first large sliding window is subjected to EEMD decomposition to obtain a plurality of IMF components.

[0067] Step 2, for each intrinsic mode component, determine the training data subset and the test data subset corresponding to the intrinsic mode component based on the second sliding window. In an embodiment, steps 2.1 to 2.5 can be referred to as follows:

[0068] Step 2.1: Set the second sliding window for the intrinsic mode component according to the window length in the second sliding window parameters. For example, for the i-th IMF component, set a small sliding window with a window length of W2 and a sliding step size of S to construct the first data subset X and the second data subset Y of the initial prediction model. The first ((W1-W2) / S)-1 first data subsets X are the inputs in the training data subset, the second data subset Y is the true value in the training data subset, and the last... This is a subset of the test data.

[0069] Step 2.2: Each data point within the second sliding window of the intrinsic mode component is taken as the first data subset corresponding to that intrinsic mode component. Referring again to the example above, the data within the first small sliding window can be taken as the first data subset X.

[0070] Step 2.3: Move the second sliding window according to the sliding step size in the second sliding window parameters, and take the first S data points of the intrinsic mode component that are located in the moved second sliding window as the second data subset corresponding to the intrinsic mode component; wherein, the second data subset is the set of true values ​​corresponding to the first data subset. Please continue to refer to the above example, slide the small sliding window for S steps, and take the first S data points in the moved small sliding window as the second data subset Y, that is, [W2:W2+S] as the second data subset Y. Here, the length of the first data subset X is W2, the length of the second data subset Y is S, and the second data subset Y is the set of true values ​​of the first data subset X.

[0071] Step 2.4: Take each data point within the moved second sliding window of the intrinsic mode component as the next first data subset corresponding to that intrinsic mode component. Continue moving the second sliding window according to the sliding step size in the second sliding window parameters until the second sliding window reaches the end of the intrinsic mode component, resulting in multiple first data subsets and multiple second data subsets. Referring to the example above, take the data within the second sliding window as the next first data subset X, then continue moving the second sliding window to construct the next second data subset Y. Repeating the above process will construct... The first data subset X and A second data subset Y.

[0072] Step 2.5: The last first data subset is used as the test data subset corresponding to the intrinsic mode component. Based on the other first data subsets (excluding the last first data subset) and each second data subset, the training data subset corresponding to the intrinsic mode component is obtained. Please continue to refer to the example above, and... The first data subset X is used as input to the training data subset, and... the last one of the second data subsets Y as the true value of the test data subset, and the last one of the first data subsets X as the predicted value of the test data subset. as the test data subset.

[0073] Step 3, taking the training data subset corresponding to each intrinsic modal component as the training data set corresponding to the first sliding window, and taking the test data subset corresponding to each intrinsic modal component as the test data set corresponding to the first sliding window. In actual application, the repeated operation is performed for each IMF component, so that the training data subset and the test data subset corresponding to each IMF component are obtained.

[0074] Step 4, determining the first S data in the network index data set located in the first sliding window after the movement as the true value of the test data set. Please continue to refer to the above example, moving the first sliding window backward by S steps to generate new window data , wherein the first S data in the new window data are the true value of the test data set corresponding to the first first sliding window. The new window data is processed according to the above steps 2 to 3 to construct the first data subset X and the second data subset Y, and then the corresponding training data set and test data set are obtained, and the cycle is repeated until all T data are decomposed.

[0075] It should be noted that because the total data is T, the predicted value of the last first sliding window is unknown, there is no true value to compare with, and the evaluation standard value cannot be calculated, so the data of the last window is used for model training, that is, the last first data subset X of the last first sliding window is not input into the prediction model for prediction, and only the first first data subset X is input into the prediction model for training.

[0076] On the basis of the above embodiment, the embodiment of the application further provides an implementation manner of training the initially established initial prediction model by using the training data set corresponding to the first sliding window, and testing the trained initial prediction model by using the test data set corresponding to the first sliding window to obtain the predicted value of the test data set, see the following steps a to d:

[0077] Step a, training the initially established initial prediction model by using the training data subset corresponding to the first intrinsic modal component. In an implementation manner, the first S first data subsets X constructed based on the first intrinsic modal component are input into the initial prediction model, and the initial prediction model outputs the corresponding predicted value, and the ​The second data subset Y is taken as a true value in the training data subset, the predicted value and the true value are substituted into a preset loss function to obtain a corresponding loss value, and then the model parameters of the initial prediction model are updated based on the loss value.

[0078] Step b, using the test data subset corresponding to the first intrinsic modal component, the trained initial prediction model is tested to obtain the predicted value of the test data subset. In an embodiment, the first data subset X and the second data subset Y corresponding to the first intrinsic modal component are input into the trained initial prediction model, and the initial prediction model outputs the corresponding predicted value . .

[0079] Step c, continue to train the initial prediction model using the training data subset corresponding to the next intrinsic modal component, and continue to test the trained initial prediction model using the test data subset corresponding to the next intrinsic modal component, until the predicted value of each test data subset is obtained. In an embodiment, the initial prediction model is trained using the first data subset X and the second data subset Y of the previous IMF component of the next IMF component, and the initial prediction model is tested using the test data subset of the next IMF component, and the process is repeated until the test data subset of the last IMF component is input into the initial prediction model to obtain the corresponding predicted value.

[0080] Step d, the sum of the predicted values of each test data subset is taken as the predicted value of the test data set corresponding to the first sliding window. In an embodiment, the predicted values of each IMF component are added to obtain the predicted value of the test data set corresponding to the first sliding window .

[0081] In actual application, after the predicted value of the first first sliding window is determined , the first sliding window is moved, and the above steps 1 to 4 and steps a to d are repeated for the data in the moved first sliding window, until all network index data sets are decomposed, and finally the target prediction model is evaluated based on the true value and the predicted value of each test data set to obtain the evaluation test result corresponding to the target prediction model.

[0082] ​​The embodiment of the present application also provides an implementation of evaluating and testing a target prediction model, which can calculate the mean absolute error (MAE), the root mean square error (RMSE), the mean absolute percentage error (MAPE) and the fitting degree (R 2 ) of the target prediction model based on the true value and the predicted value of each test data set, so as to obtain the evaluation test result corresponding to the target prediction model.

[0083] For the convenience of understanding the steps 1 to 4 and the steps a to d, refer to a sliding window type decomposition flowchart shown in Figure 6 , which assumes that the length of the time series data is T, sets the length of the large sliding window as W1, and the step length as S. First, the EEMD decomposition is performed on the first large sliding window data to obtain a plurality of IMF components. Then, a small sliding window is set for the i-th IMF component, the length of the small sliding window is W2, and the step length is S, so as to construct the first data subset X and the second data subset Y input to the BiLSTM model. The first W2 data are constructed as the first data subset X, , and the remaining data are constructed as the second data subset Y. The first data subset X and the second data subset Y are input into the model for training, and the last first data subset is input into the trained model for prediction to obtain the predicted value . The operation is repeated for each IMF, and the predicted values of the IMFs are added to obtain , which is saved. Then, the large sliding window W1 is slid by S steps to generate new window data . The first data subset X and the second data subset Y are constructed according to the above operation, and the operation is repeated until the predicted value is obtained. The cycle is repeated until all T data are decomposed. It is noted that the data of the last large sliding window are all used for model training. Since the total number of data is T, the predicted value corresponding to the window is unknown, and there is no true value for comparison, so the evaluation standard value cannot be calculated. Therefore, the last first data subset X of the last large sliding window is not input into the prediction model for prediction, and only the first first data subsets X are input into the prediction model for training. Finally, the predicted value is obtained, which is compared with the true data to calculate the evaluation standard value, wherein the evaluation standard value is the mean absolute error, the root mean square error, the mean absolute percentage error and the fitting degree.

[0084] The application embodiment further provides an application example of the training method of the prediction model, referring to Figure 7 The application embodiment adopts 25700 pieces of Irish data sets, sets a sliding window with a window size of 20000 and a step of 50, puts the original data set into the EEMD decomposition method in a moving window manner, and decomposes 12 IMF components and one residual component from one window size of data at a time. Then, for each IMF component, a sliding window with a window size of 500 and a step of 50 is set, a first data subset X and a second data subset Y are constructed, for example, the first first data subset X is [1, 500], the second data subset Y is [501, 550], the second first data subset X is [50, 550], and the second data subset Y is [551, 600], and the input data set with a length of N is constructed in this way. The first N-1 first data subsets X are input into the BiLSTM model for training and learning, and the second data subset Y corresponding to the last first data subset X [19501, 20000] is unknown, so the last first data subset X is used for prediction, and the predicted second data subset Y [20001, 20050] is saved. The training and prediction operation of each IMF component is repeated, and the predicted values of the IMFs are added to obtain the predicted value of the original data value, which is also saved. Thus, the first EEMD decomposition and BiLSTM training and prediction are completed. Then, the large window of 20000 data is moved by 50, and the data in the window [50, 20050] is decomposed, and the training and prediction operation is repeated. The predicted value Y [20051, 20100] is added to the previous predicted data. The operation is repeated until all data are decomposed. Note that the data of the last large window is used for training because the next 50 data are unknown. Finally, the predicted value of the data [20001, 25700] is obtained, and the real value and the predicted value are compared to calculate the RMSE, MAE, MAPE, R 2 Four evaluation standard values. The algorithm flowchart is shown in Figure 5 Two sliding windows are added on the basis of Figure 7 The application embodiment based on the improved sliding window EEMD-BiLSTM model not only avoids the problem of information leakage, but also improves the prediction effect compared with the EMD-BiLSTM model and the EEMD-BiLSTM model. The prediction accuracy effect is improved by 5.53%, and the fitting degree of the model is increased from 0.981 to 0.994.

[0085] Table 2

[0086]

[0087] In summary, the embodiment of the application belongs to the use of improved EEMD-BiLSTM prediction model to predict 5G network index data, and the EMD method adopts global decomposition, so the data in the test set is considered at the time of decomposition, but for prediction, the data in the test set should be unknown, which is the huge and serious defect of EMD-information leakage problem, and the sliding window decomposition data is used to solve the problem, the decomposition data and the prediction data are decomposed in the window, and the prediction data is in the next window, so the decomposition contains the prediction data. After each decomposition, the IMFs are input into the BiLSTM model for training and testing, and then the window is moved to decompose the data in the next window. However, EMD is prone to modal aliasing when decomposing a large amount of data, so the EEMD method is used to avoid modal aliasing, and the effect of the decomposed IMF is improved, and then the Dropout is added to the BiLSTM model to prevent overfitting and reduce the complex co-adaptive relationship between neurons, highlight the role of effective features, generalize the model, and improve the prediction accuracy of the model.

[0088] The embodiment of the application adopts the above-mentioned sliding window EEMD decomposition method, avoids the global decomposition of EEMD, and avoids the problem of considering the test set data in advance, and solves the information leakage problem well; and the Dropout is added to the BiLSTM to prevent overfitting and reduce the complexity between neurons, which optimizes the sliding window EEMD-BiLSTM prediction model to a certain extent. The embodiment of the application has at least the following characteristics:

[0089] (1) The data set is preprocessed to improve the feasibility of the data;

[0090] (2) The EEMD decomposition method avoids the modal aliasing phenomenon, and the number of IMF components is reduced from 15 to 9;

[0091] (3) The sliding window EEMD is used to segmentally decompose the data set, and the predicted data is in the next window, so as to avoid considering the test data in the decomposition;

[0092] (4) The corresponding Dropout is added to each layer of the BiLSTM network to optimize the model.

[0093] The experimental results show that the EEMD information leakage problem is solved well, and the EEMD-BiLSTM prediction model is more rigorous and feasible. The EEMD global decomposition is added with a sliding window, and becomes segmented decomposition, and decomposition is only performed in one window at a time, and the EEMD decomposition of the current window does not contain prediction data, and the predicted data is in the next window, so that information leakage is avoided. The method obviously improves the information leakage problem existing in the EEMD global decomposition, and the sliding window type EEMD-BiLSTM model is more dynamic, feasible and interpretable.

[0094] For the training method of the prediction model provided in the foregoing embodiment, an embodiment of the present application provides a prediction model training device, which refers to a structure schematic diagram of a prediction model training device shown in Figure 8 The device mainly includes the following parts:

[0095] The data acquisition module 802 is configured to acquire a network index data set.

[0096] The first sliding module 804 is configured to perform ensemble empirical mode decomposition processing on the network index data set based on a first sliding window, to determine a training data set, a test data set and a true value of the test data set corresponding to the first sliding window. The training data set and the test data set are determined based on data in the first sliding window, and the true value of the test data set is determined based on data in the first sliding window after movement.

[0097] The training and testing module 806 is configured to train a pre-established initial prediction model by using the training data set corresponding to the first sliding window, and test the trained initial prediction model by using the test data set corresponding to the first sliding window, to obtain a prediction value of the test data set.

[0098] The second sliding module 808 is further configured to move the first sliding window, and continue to perform ensemble empirical mode decomposition processing on the network index data set based on the first sliding window after movement, until the network index data set is decomposed, to obtain a target prediction model and a prediction value of each test data set.

[0099] The evaluation and testing module 810 is configured to evaluate and test the target prediction model based on the true value and the prediction value of each test data set, to obtain an evaluation and testing result corresponding to the target prediction model.

[0100] The training device of the prediction model provided by the embodiment of the present application utilizes the sliding window decomposition network index data set, performs integrated empirical mode decomposition processing on the network index data set in the first sliding window, determines the training data set and the test data set based on the data in the first sliding window, and determines the true value corresponding to the test data set based on the data in the moved first sliding window. Since the data in the first sliding window is sequentially decomposed in a segmented manner, the true value of the test training set is not leaked in the network index data set decomposition process, thereby effectively improving the information leakage problem of the integrated empirical mode decomposition processing, and further effectively improving the training effect of the prediction model, so that the target prediction model obtained by training has high prediction accuracy.

[0101] In an implementation, the first sliding module 804 is further configured to perform integrated empirical mode decomposition processing on the network index data set based on the first sliding window to obtain a plurality of intrinsic mode components corresponding to the first sliding window; for each intrinsic mode component, determine a training data subset and a test data subset corresponding to the intrinsic mode component based on the second sliding window; take the training data subset corresponding to each intrinsic mode component as the training data set corresponding to the first sliding window; and take the test data subset corresponding to each intrinsic mode component as the test data set corresponding to the first sliding window; and determine the first S data in the network index data set located in the moved first sliding window as the true value of the test data set; wherein S is the sliding step length of the first sliding window.

[0102] In an implementation, the first sliding module 804 is further configured to set the second sliding window for the intrinsic mode component according to the window length in the second sliding window parameter; take each data in the intrinsic mode component located in the second sliding window as the first data subset corresponding to the intrinsic mode component; move the second sliding window according to the sliding step length in the second sliding window parameter, and take the first S data in the intrinsic mode component located in the moved second sliding window as the second data subset corresponding to the intrinsic mode component; wherein the second data subset is a true value set corresponding to the first data subset; take each data in the intrinsic mode component located in the moved second sliding window as the next first data subset corresponding to the intrinsic mode component, and continue to move the second sliding window according to the sliding step length in the second sliding window parameter until the second sliding window moves to the end of the intrinsic mode component to obtain a plurality of first data subsets and a plurality of second data subsets; take the last first data subset as the test data subset corresponding to the intrinsic mode component, and obtain the training data subset corresponding to the intrinsic mode component based on the other first data subsets except the last first data subset and each second data subset.

[0103] In an implementation, the second sliding module 808 is further configured to: train the pre-established initial prediction model by using the training data subset corresponding to the first intrinsic modal component; test the trained initial prediction model by using the test data subset corresponding to the first intrinsic modal component to obtain the predicted value of the test data subset; continue to train the initial prediction model by using the training data subset corresponding to the next intrinsic modal component, and continue to test the trained initial prediction model by using the test data subset corresponding to the next intrinsic modal component until the predicted value of each test data subset is obtained; and take the sum of the predicted values of each test data subset as the predicted value of the test data set corresponding to the first sliding window.

[0104] In an implementation, the evaluation test module 810 is further configured to: calculate the mean absolute error, the root mean square error, the mean absolute percentage error and the fitting degree of the target prediction model based on the true value and the predicted value of each test data set to obtain the evaluation test result corresponding to the target prediction model.

[0105] In an implementation, the pre-processing module is further configured to: determine the correlation degree corresponding to each data in the network index data set, and remove the data from the network index data set if the correlation degree corresponding to the data is less than a preset correlation threshold; determine the missing data from the network index data set, fill the missing data, or remove the missing data from the network index data set; and determine the abnormal data from the network index data set and remove the abnormal data from the network index data set.

[0106] In an implementation, the initial prediction model is an improved bidirectional long short-term memory model, and a random inactivation layer is arranged between the specified layer neurons of the improved bidirectional long short-term memory model.

[0107] The device provided by the embodiments of the present application has the same implementation principle and technical effects as the foregoing method embodiments. For brevity, the part not mentioned in the device embodiment part can be referred to the corresponding content in the foregoing method embodiments.

[0108] The electronic device provided by the embodiments of the present application includes a processor and a storage device. The storage device stores a computer program, and the computer program performs the method according to any one of the foregoing embodiments when executed by the processor.

[0109] Figure 9 The electronic device provided by the embodiments of the present application includes a processor and a storage device. The storage device stores a computer program, and the computer program performs the method according to any one of the foregoing embodiments when executed by the processor.

[0110] The memory 91 can include a high-speed random access memory (RAM), and can also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 93 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used.

[0111] The bus 92 can be an ISA bus, a PCI bus, an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 9 Only one bidirectional arrow is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0112] The memory 91 is used to store programs, and the processor 90 executes the programs after receiving execution instructions. The method executed by the device defined by the flow process disclosed in any of the embodiments of the present application can be applied to the processor 90 or implemented by the processor 90.

[0113] The processor 90 can be an integrated circuit chip with signal processing capability. In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware in the processor 90 or the instruction in the form of software. The processor 90 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block disclosed in the embodiment of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiment of the present application can be directly embodied as a hardware code processor for execution, or can be executed by a combination of hardware and software modules in the code processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the memory 91, and the processor 90 reads the information in the memory 91 and combines the hardware to complete the steps of the above method.

[0114] The computer program product of the readable storage medium provided by the embodiment of the present application comprises a computer readable storage medium storing program codes, and the instructions included in the program codes can be used to execute the method described in the foregoing method embodiment. The specific implementation can be referred to the foregoing method embodiment, and will not be described here.

[0115] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts 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 instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0116] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them. The protection scope of the present application is not limited to this. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical range disclosed by the present application, or make equivalent replacements to some technical features. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for training a prediction model, characterized in that, include: Obtain a network metrics dataset, which consists of key network performance metrics, including channel-related metrics, signal-to-noise ratio, and download / upload rates. The network metric dataset is subjected to integrated empirical mode decomposition based on a first sliding window to determine the training dataset, test dataset, and true values ​​of the test dataset corresponding to the first sliding window; wherein the training dataset and the test dataset are determined based on the data within the first sliding window, and the true values ​​of the test dataset are determined based on the data within the first sliding window after the window has been moved. The initial prediction model is trained using the training dataset corresponding to the first sliding window, and the trained initial prediction model is tested using the test dataset corresponding to the first sliding window to obtain the prediction value of the test dataset. Move the first sliding window and continue to perform integrated empirical mode decomposition on the network indicator dataset based on the moved first sliding window until the network indicator dataset is completely decomposed to obtain the target prediction model and the prediction value of each test dataset. Based on the true values ​​and predicted values ​​of each of the test datasets, the target prediction model is evaluated and tested to obtain the evaluation test results corresponding to the target prediction model; Based on a first sliding window, ensemble empirical mode decomposition is performed on the network metric dataset to determine the training dataset, test dataset, and the true values ​​of the test dataset corresponding to the first sliding window, including: Based on the first sliding window, the network index dataset is subjected to integrated empirical mode decomposition to obtain multiple intrinsic mode components corresponding to the first sliding window. For each intrinsic mode component, the training data subset and test data subset corresponding to that intrinsic mode component are determined based on the second sliding window; The training data subset corresponding to each intrinsic mode component is used as the training dataset corresponding to the first sliding window; and the test data subset corresponding to each intrinsic mode component is used as the test dataset corresponding to the first sliding window. The first S data points within the first sliding window after the network metric dataset are determined as the true values ​​of the test dataset; where S is the sliding step size of the first sliding window. The training data subset and test data subset corresponding to the intrinsic mode component are determined based on the second sliding window, including: Set the second sliding window for the intrinsic mode component according to the window length in the second sliding window parameters; Each data point within the second sliding window of the intrinsic mode component is taken as the first data subset corresponding to the intrinsic mode component; According to the sliding step size in the second sliding window parameters, the second sliding window is moved, and the first S data points of the intrinsic mode component located in the moved second sliding window are taken as the second data subset corresponding to the intrinsic mode component; wherein, the second data subset is the set of true values ​​corresponding to the first data subset; Each data point in the intrinsic mode component that is located within the second sliding window after the movement is taken as the next first data subset corresponding to the intrinsic mode component. The second sliding window is then moved according to the sliding step size in the second sliding window parameters until the second sliding window moves to the end of the intrinsic mode component, resulting in multiple first data subsets and multiple second data subsets. The last first data subset is used as the test data subset corresponding to the intrinsic mode component, and the training data subset corresponding to the intrinsic mode component is obtained based on the other first data subsets excluding the last first data subset and each second data subset.

2. The training method for the prediction model according to claim 1, characterized in that, The initial prediction model is trained using the training dataset corresponding to the first sliding window, and the trained initial prediction model is tested using the test dataset corresponding to the first sliding window to obtain the predicted values ​​of the test dataset, including: The initial prediction model is trained using the training data subset corresponding to the first intrinsic mode component. The initial prediction model after training is tested using the test data subset corresponding to the first intrinsic mode component to obtain the predicted value of the test data subset. The initial prediction model is trained using the training data subset corresponding to the next intrinsic mode component, and the trained initial prediction model is tested using the test data subset corresponding to the next intrinsic mode component, until the predicted value of each test data subset is obtained. The sum of the predicted values ​​of each subset of test data is used as the predicted value of the test dataset corresponding to the first sliding window.

3. The training method for the prediction model according to claim 2, characterized in that, Based on the true values ​​and predicted values ​​of each of the test datasets, the target prediction model is evaluated and tested to obtain the evaluation test results corresponding to the target prediction model, including: Based on the true values ​​and predicted values ​​of each test dataset, the mean absolute error, root mean square error, mean absolute percentage error, and goodness of fit of the target prediction model are calculated to obtain the evaluation test results corresponding to the target prediction model.

4. The training method for the prediction model according to claim 1, characterized in that, Before performing integrated empirical mode decomposition on the network indicator dataset based on the first sliding window, the process further includes: Determine the relevance of each data point in the network indicator dataset. If the relevance of the data point is less than a preset relevance threshold, remove the data from the network indicator dataset. In addition, missing data is identified from the network metric dataset, and the missing data is filled in or removed from the network metric dataset; In addition, abnormal data is identified from the network metrics dataset, and the abnormal data is removed from the network metrics dataset.

5. The training method for the prediction model according to claim 1, characterized in that, The initial prediction model is an improved bidirectional long short-term memory model, in which a random deactivation layer is set between neurons in a specified layer.

6. A training device for a prediction model, characterized in that, include: The data acquisition module is used to acquire a network indicator dataset, which consists of key network performance indicators, including channel-related indicators, signal-to-noise ratio, and download / upload rate. The first sliding module is used to perform integrated empirical mode decomposition processing on the network indicator dataset based on the first sliding window to determine the training dataset, the test dataset, and the true value of the test dataset corresponding to the first sliding window; wherein the training dataset and the test dataset are determined based on the data within the first sliding window, and the true value of the test dataset is determined based on the data within the first sliding window after the sliding window has been moved. The training and testing module is used to train the pre-established initial prediction model using the training dataset corresponding to the first sliding window, and to test the trained initial prediction model using the test dataset corresponding to the first sliding window, so as to obtain the prediction value of the test dataset. The second sliding module is also used to move the first sliding window and continue to perform integrated empirical mode decomposition processing on the network indicator dataset based on the moved first sliding window until the network indicator dataset is decomposed to obtain the target prediction model and the prediction value of each test dataset. An evaluation and testing module is used to evaluate and test the target prediction model based on the true value and the predicted value of each test dataset, and obtain the evaluation and testing results corresponding to the target prediction model. The first sliding module is specifically used for: Based on the first sliding window, the network index dataset is subjected to integrated empirical mode decomposition to obtain multiple intrinsic mode components corresponding to the first sliding window. For each intrinsic mode component, the training data subset and test data subset corresponding to that intrinsic mode component are determined based on the second sliding window; The training data subset corresponding to each intrinsic mode component is used as the training dataset corresponding to the first sliding window; and the test data subset corresponding to each intrinsic mode component is used as the test dataset corresponding to the first sliding window. The first S data points within the first sliding window after the network metric dataset are determined as the true values ​​of the test dataset; where S is the sliding step size of the first sliding window. The first sliding module is specifically used for: Set the second sliding window for the intrinsic mode component according to the window length in the second sliding window parameters; Each data point within the second sliding window of the intrinsic mode component is taken as the first data subset corresponding to the intrinsic mode component; According to the sliding step size in the second sliding window parameters, the second sliding window is moved, and the first S data points of the intrinsic mode component located in the moved second sliding window are taken as the second data subset corresponding to the intrinsic mode component; wherein, the second data subset is the set of true values ​​corresponding to the first data subset; Each data point in the intrinsic mode component that is located within the second sliding window after the movement is taken as the next first data subset corresponding to the intrinsic mode component. The second sliding window is then moved according to the sliding step size in the second sliding window parameters until the second sliding window moves to the end of the intrinsic mode component, resulting in multiple first data subsets and multiple second data subsets. The last first data subset is used as the test data subset corresponding to the intrinsic mode component, and the training data subset corresponding to the intrinsic mode component is obtained based on the other first data subsets excluding the last first data subset and each second data subset.

7. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method described in any one of claims 1 to 5.

Images