Satellite link traffic prediction method based on LSTM and autoencoder

By combining the prediction network of LSTM and automatic encoder, the problem of medium- and long correlation between satellite link traffic prediction is solved, the prediction accuracy is improved and the calculation complexity is reduced, and efficient management and control of the traffic of satellite communication network is achieved.

CN114742207BActive Publication Date: 2024-11-22BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210427294.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-21
Publication Date
2024-11-22
Estimated Expiration
2042-04-21

AI Technical Summary

Technical Problem

The prior art is difficult to accurately predict the long correlation of inter-satellite link traffic, and traditional algorithms perform poorly in describing network trends, fluctuations, and periodicity, resulting in low accuracy in predicting traffic in satellite communication networks.

Method used

Using a prediction network based on LSTM and automatic encoder, the prediction network is built to improve prediction accuracy and reduce prediction accuracy by preprocessing the original traffic data, sliding window segmentation and model training, and combining the feature extraction and prediction capabilities of LSTM and automatic encoder. Computational complexity.

Benefits of technology

It effectively improves the accuracy of inter-satellite link traffic prediction, reduces the complexity and training time of the prediction network, and meets the requirements for traffic prediction accuracy and effectiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114742207B_ABST
    Figure CN114742207B_ABST
Patent Text Reader

Abstract

The present invention provides a satellite link traffic prediction method based on LSTM and autoencoder, comprising the following steps: step S1, preprocessing the original traffic data set D to obtain the data set D p ; Step S2, for data set D p Apply a sliding window for segmentation to obtain a set #imgabs0# and divide the set #imgabs1# into a training set and a test set; step S3, construct an autoencoder model with an encoder part and a decoder part; step S4, use the training set in step S2 to pre-train the autoencoder model; step S5, construct an LSTM model, and connect it to the autoencoder pre-trained in step S4 to form a prediction network; step S6, use the training set in step S2 to train the prediction network. The present invention can improve prediction accuracy and reduce computational complexity and training time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a satellite link traffic prediction method based on LSTM and an autoencoder, belonging to the technical field of traffic prediction in wireless communications. Background Art

[0002] In recent years, communication networks, represented by the Internet, have become one of the most critical infrastructures in modern society, serving as the engine of social development. With the rapid development of the Internet of Things (IoT), the load on communication networks has expanded from simple human-to-human interactions to include communication between people and things, and even between things. Social development and the progress of the times have placed higher demands on next-generation communication networks, driving their further development. New technologies are constantly emerging in various fields of communication technology, both domestically and internationally, particularly in wireless communications. Communication networks have permeated every aspect of our lives, profoundly impacting the world today. It's even fair to say that without communication networks, the current world would be impossible.

[0003] However, with the increasing scale of communication networks and the heterogeneity between different networks, the complexity of each network is also increasing, gradually evolving from a simple structure into a massive, complex system. Satellite communication networks are not only more complex than ordinary communication networks, but also often require greater financial investment. Therefore, the operation management, behavior monitoring, equipment maintenance, and security of satellite communication networks are becoming increasingly important and urgent from social, economic, and technical perspectives. From a social perspective, as an infrastructure on which human society increasingly relies, it is necessary to understand its basic statistical characteristics. From an economic perspective, obtaining statistical data on satellite communication network traffic can help improve resource utilization and reduce satellite communication network construction investment. From a technical perspective, the deployment of network equipment and the design of network protocols all depend on the workload characteristics of the communication network, and network administrators need to adjust network systems based on these characteristics. Managing satellite communication networks requires monitoring, statistics, and forecasting of their traffic. Therefore, how to establish a model that can adapt to the nature of satellite communication network traffic, used for the generation and prediction of satellite communication network traffic data, improve the accuracy of short-term and long-term predictions, reduce prediction errors, thereby promoting the scientific management and effective control of satellite communication networks, and providing a scientific basis for the construction and development of satellite communication networks has become one of the topics we must study.

[0004] Since the advent of the internet, numerous researchers and experts have tirelessly dedicated their time and energy to this hot topic, continuously developing more advanced and broadly applicable prediction algorithms in an effort to further improve the efficiency and reduce the error of network traffic prediction. However, with the increasing complexity of networks, traditional algorithms such as Markov, Kalman filtering, and ARMA models, while well-established in mathematical theory and highly interpretable, are often only able to describe sequences with short-term correlations. They perform poorly when trying to capture long-term correlations such as overall network trends, fluctuations, and even periodicity. Therefore, accurate prediction of traffic between satellite inter-satellite links remains a pressing issue. Summary of the Invention

[0005] The present invention proposes a satellite link traffic prediction method based on LSTM and autoencoder, which can improve prediction accuracy and reduce computational complexity and training time.

[0006] The present invention is achieved through the following technical solutions:

[0007] The satellite link traffic prediction method based on LSTM and autoencoder includes the following steps:

[0008] Step S1: original traffic data set D = {d1, d2, ..., d n-1 ,d n}Preprocess and get the data set n represents the total number of data points in the original traffic dataset;

[0009] Step S2: Data set D p Apply sliding window to split and obtain a set of multiple sequences And collect Divide into training set and test set;

[0010] Step S3, constructing an automatic encoder model having an encoder part and a decoder part, the encoder part includes a first input layer, a first LSTM layer and an intermediate hidden layer, the first LSTM layer is used to encode the input data into a feature vector contained in the intermediate hidden layer, the decoder part includes a first vector repetition layer, a second LSTM layer, a first BP neural network layer inserted in time sharing and a first output layer, the first vector repetition layer is used to repeat the feature vector contained in the intermediate hidden layer multiple times and input it into the second LSTM layer, the second LSTM layer and the first BP neural network layer inserted in time sharing are used to decode the feature vector;

[0011] Step S4: pre-train the autoencoder model using the training set in step S2;

[0012] Step S5: construct an LSTM model and connect it to the autoencoder pre-trained in step S4 to form a prediction network;

[0013] Step S6: Use the training set in step S2 to train the prediction network.

[0014] Furthermore, in step S5, the LSTM model includes a second vector repetition layer, a second LSTM layer and a second BP neural network layer inserted in time sharing, and after the parameters of each layer are randomly initialized, they are connected to the encoder part of the autoencoder.

[0015] Furthermore, in step S1, the original traffic data set D = {d1, d2, ..., d n-1 ,d n The preprocessing specifically includes: first filling in the missing values ​​in the original traffic data set D, replacing outliers, deleting duplicate data, and then normalizing the data.

[0016] Furthermore, the step S2 specifically includes the following steps:

[0017] Step S21: Determine the hyperparameter L x 、L y and L total =L x +L y , L x L is the length of the sequence required for each prediction. y The length of the sequence required for each prediction is L. total ;

[0018] Step S22: Sliding window segmentation from time point t=L x Start, to time t = nL x End, each time a split is performed, the sliding opening moves forward one data, and nL is obtained after the split total +1 sequence, each of length L total , thus obtaining a set of multiple sequences

[0019] Step S23: Each sequence in the Lth x The time points are divided into x and length L y Two sequences, nL total +1 length L x The sequence constitutes the set E data , nL total +1 length L yThe sequence of constitutes the set Y;

[0020] Step S24: Set E data And set Y is divided in the same proportion, and the corresponding set E is obtained data _train and set E data _test, set Y_train and set Y_test, set E data _train and set Y_train form the training set, set E data _test and the set Y_test form the test set.

[0021] Furthermore, in step S4, the data and labels of the training set input for the autoencoder model pre-training are all set E data _train, the data and labels of the test set are all set E data _test, the loss function of pre-training is: Among them, E data _train re is the data sequence restored by the autoencoder.

[0022] Furthermore, in step S6, the data of the training set inputted into the prediction network training is E data _train, the label is Y_train, and the data of the test set is E data _test, the label is Y_test, and the training of the prediction network is achieved through back propagation algorithm and gradient descent.

[0023] Furthermore, the optimizers used in the training of the autoencoder model and the training of the prediction network are both Adam optimizers.

[0024] Furthermore, after the pre-training of the autoencoder model is completed, the decoder part is discarded and only the encoder part is retained, and the LSTM model is connected to the encoder part.

[0025] Furthermore, for missing values ​​in the original traffic dataset D, if the missing values ​​are located at the beginning or end of the dataset, the corresponding time points are directly deleted, and other missing values ​​are filled using linear interpolation.

[0026] The present invention has the following beneficial effects:

[0027] 1. The present invention uses a prediction network combining an LSTM model and an autoencoder to process the sequence, which can better summarize the data features, thereby better solving the long-range correlation problem in satellite inter-satellite link traffic prediction and effectively improving the accuracy of satellite inter-satellite link traffic prediction. In order to avoid the situation where the complexity and training time of the prediction network increase sharply with the increase of the input sequence length, the autoencoder is first used for pre-training to reduce the length of the input sequence, thereby greatly reducing the computational complexity and training time, so as to meet the needs of scenarios with high requirements for traffic prediction accuracy and effectiveness; considering that the goals of autoencoder pre-training and the ultimate goals of the prediction network are similar but not completely consistent, after the prediction network is constructed, the autoencoder and the LSMT model are systematically trained, so that the encoder part can be fine-tuned while the prediction network is trained, which is more conducive to approaching the global optimal solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be further described in detail below with reference to the accompanying drawings.

[0029] Figure 1 Flowchart of the present invention.

[0030] Figure 2 This is an explanatory diagram of the automatic encoder of the present invention.

[0031] Figure 3 This is a structural diagram of the prediction network of the present invention.

[0032] Figure 4 This is the prediction result diagram of the present invention.

[0033] Figure 5 This is the prediction result diagram using only the LSTM neural network. DETAILED DESCRIPTION

[0034] like Figure 1 As shown in FIG, the satellite link traffic prediction method based on LSTM and autoencoder includes the following steps:

[0035] Step S1: original traffic data set D = {d1, d2, ..., d n-1 ,d n}Preprocess and get the data set n represents the total number of data points in the original traffic dataset. In this embodiment, n = 19888. Specifically, the process includes filling missing values ​​in the original traffic dataset D, replacing outliers, deleting duplicate data, and then normalizing the data. For missing values ​​in the original traffic dataset D, if the missing value is located at the beginning or end of the dataset, the corresponding time point is directly deleted. For other missing values, linear interpolation is used to fill them. The formula for linear interpolation is: The processes of replacing outliers and deleting duplicate data are both conventional techniques. In this embodiment, the raw traffic data set is derived from ISP internet traffic data (in bits). This data was collected from aggregated traffic on the UK academic network backbone between 9:30 AM on November 19, 2004, and 11:11 AM on January 27, 2005. Data was collected and counted every five minutes during this period.

[0036] Then use Min-max normalization for normalization. The specific formula is: Among them, d is the data in the original traffic data set, d min is the minimum value in the original traffic data set, d max is the maximum value in the original traffic data set, d p This is the normalized result.

[0037] Step S2: Data set D p Apply sliding window to split and obtain a set of multiple sequences And collect Divide into training set and test set, specifically including the following steps:

[0038] Step S21: Determine the hyperparameter L x 、L y and L total =L x +L y , L x L is the length of the sequence required for each prediction. y is the length of the sequence required for each prediction. Assuming that the prediction is made at time point t, it is necessary to Input the system and the system will predict The value of L total parameters, so the length of the sliding window is L total In this embodiment, L x =120, L y =24,L total =144;

[0039] Step S22: Sliding window segmentation from time point t=L x Start, to time t = nL x End, each time a split is performed, the sliding opening moves forward one data, and nL is obtained after the split total +1 sequence, each of length L total , thus obtaining a set of multiple sequences

[0040] Step S23: Each sequence in the Lth x The time points are divided into x and length L y Two sequences, nL total +1 length L x The sequence constitutes the set E data , nL total +1 length L y The sequence of constitutes the set Y;

[0041] Step S24: Set E data And set Y is divided in the same proportion, and the corresponding set E is obtained data _train and set E data _test, set Y_train and set Y_test, set E data _train and set Y_train form the training set, set E data _test and set Y_test form a test set. In this embodiment, the set E data The first 70% of the sequences in the set Y are divided into the training set, and the last 30% of the sequences are divided into the test set.

[0042] Step S3: Construct an autoencoder model with an encoder part and a decoder part. The autoencoder model is used to compress the input data X into hidden features and then reconstruct the output through this feature.

[0043] The encoder part includes a first input layer, a first LSTM layer and an intermediate hidden layer. The first LSTM layer is used to encode the input data into the feature vector contained in the intermediate hidden layer. The decoder part includes a first vector repetition layer, a second LSTM layer, a first BP neural network layer inserted in time-sharing mode and a first output layer. The first vector repetition layer is used to repeat the feature vector contained in the intermediate hidden layer multiple times and input it into the second LSTM layer. The second LSTM layer and the first BP neural network layer inserted in time-sharing mode are used to decode the feature vector. Finally, the restored data sequence is obtained in the input layer, such as Figure 2 The figure shows an illustration of the automatic encoder in this embodiment, in which X n Represents input data, C n is the hidden feature learned by the autoencoder, It is the data restored by the autoencoder using hidden features.

[0044] Step S4: Use the training set in step S2 to pre-train the autoencoder model. Specifically, the data and labels of the training set input for the autoencoder model pre-training are all set Edata _train, the data and labels of the test set are all set E data _test, the loss function of pre-training is: Among them, E data _train re The data sequence restored by the autoencoder. The lower the value of the loss function, the higher the degree of restoration of the data sequence by the autoencoder, which means that the feature vector finds the implicit features in the data sequence better. The optimizer used for pre-training is the Adam optimizer.

[0045] In this step, the forward information propagation in the autoencoder network can be described by the following formula:

[0046] z (l) =W (l) *a (l-1) +b (l) ,

[0047] a (l) =f l (z (l) ),

[0048] From the above two formulas, we can get:

[0049] z (l) =W (l) *f l-1 (z (l-1) )+b (l) ,

[0050] a (l) =f l (W (l) *a (l-1) +b (l) ),

[0051] The forward information propagation of the network can be regarded as a composite function, with the vector x as the input a(0) of the first layer and the output a(L) of the Lth layer as the output of the entire function:

[0052]

[0053] Among them, L represents the number of layers of the neural network, f l () represents the activation function of the layer neurons, W l represents the weight matrix from layer l-1 to layer l, b (l) represents the bias from layer l-1 to layer l, z (l) represents the net input of the l-layer neurons, a (l) represents the output of neurons in layer l, W and b represent the connection weights and biases of all layers in the network;

[0054] After the autoencoder model is pre-trained, the decoder part is discarded and only the encoder part is retained.

[0055] Step S5: Construct an LSTM model. The LSTM model includes a second vector repetition layer, a second LSTM layer, and a second BP neural network layer with time-sharing insertion. The LSTM model and the decoder have certain similarities, but their goals are different. The decoder aims to restore the original data sequence, while the LSTM model aims to predict future data sequences, so there are differences in parameter settings. After the LSTM model randomly initializes the parameters of each layer, it is connected to the encoder part of the autoencoder to form a prediction network.

[0056] LSTM is a structure that evolved from RNNs. Its essence remains the same as RNNs: the network memorizes previous information and applies it to the calculation of the current output. Nodes in the hidden layers are connected, and the input to a hidden layer includes not only the output of the input layer but also the output of the previous hidden layer. However, unlike RNNs, RNNs can only handle short-term correlations to a certain extent, but are unable to handle long-term correlations. RNNs are unable to handle long-term correlations because, when the sequence is long, the gradient descent from the new input is difficult to propagate back to the previous sequence, a phenomenon known as the vanishing gradient problem. Furthermore, RNNs can also experience exploding gradients during operation. LSTMs, however, improve upon the RNN's fundamental structure to address long-term correlations.

[0057] From a macro perspective, the LSTM network structure is much more complex than that of the RNN. From a micro perspective, the LSTM introduces cell states and uses three gates: input gate, forget gate, and output gate to maintain and control information.

[0058] The output of the LSTM prediction model can be calculated using the following formula:

[0059] i(t)=σ(W ih h(t-1)+W ix x(t)+W ix c(t-1)+b i ),

[0060] f(t)=σ(W fh h(t-1)+W fx x(t)+W fx c(t-1)+b f ),

[0061] c(t)=f(t)⊙c(t-1)+i(t)⊙tanh(W ch h(t-1)+W cxx(t)+b c ),

[0062] o(t)=σ(W oh h(t-1)+W ox x(t)+W oc c(t)+b o ),

[0063] h(t)=o(t)⊙tanh(c(t))

[0064] In the above formula, f t For the forget gate, i t is the input gate, o t is the output gate, c t is the cell state, h t is the hidden layer state value, W and b are weights and biases. ⊙ represents the scalar product of two vectors, and σ(x) is the Sigmod function. The definition of tanh is

[0065] The core of LSTM to maintain and control information lies in the forget gate, input gate and output gate. First, the forget gate combines the state value h of the previous hidden layer t-1 and the current input x t , using the Sigmod function σ(x), it is determined that the old information will be deleted. Then the input gate and tanh determine the hidden layer activation value h from the previous moment. t-1 and the current input value x t What new information is saved in the cell? Next, combine the forget gate and input gate to discard and save information, and get the cell state c at the current moment t Finally, the output gate combines with tanh to determine h t-1 、x t 、c t Which information can be output as the hidden layer state h at this moment? t .

[0066] LSTM uses the above method to decide how to discard, keep and update information. From a microscopic perspective, because the final result h t It is derived from the interaction of multiple functions and the combination of summation operations not found in RNNs (except for bias addition), making it less susceptible to vanishing gradients during backpropagation. From a macro perspective, the cell states of each LSTM timestep are connected by a straight line that combines the information of the input gate and the forget gate. This means that by properly setting the input gate and forget gate, the LSTM can long-term remember the value of the cell state at a specific timestep.

[0067] Step S6: Use the training set in step S2 to train the prediction network. Specifically, the data of the training set input for the prediction network training is E data _train, the label is Y_train, and the data of the test set is E data _test, the label is Y_test. The training of the prediction network is achieved through back propagation algorithm and gradient descent.

[0068] The quality of a neural network's output is determined by the parameters W and b. The process of training a neural network is actually to continuously update the parameters to get the best results. Assuming that the loss function is L(y, y'), the essence of parameter learning is to calculate the gradient of the loss function.

[0069] Given a sample (x, y), input it into the network model, get the network output y, and the parameter weight W in the lth layer (l) and bias b (l) Calculating the partial derivatives yields:

[0070]

[0071]

[0072] in, is the objective function about the neuron z in the lth layer (l) The partial derivative of is called the error term, so we need to calculate three partial derivatives, which are You can get:

[0073]

[0074]

[0075]

[0076]

[0077] Among them, diag() is a diagonal matrix, ⊙ is the dot product operator of the vector, which means multiplying each element. We use δ l To define the error term of the neurons in the ,th layer, then:

[0078] δ l =diag(f′ l (z (l) ))·(W (l+1) ) T ·δ (l+1) =f′ l (z (l) )⊙((W (l+1) ) T·δ (l+1) ),

[0079] As can be seen from the above formula, the error term of layer l can be calculated from the error term of layer l+1. This is the back propagation of the error. The meaning of the back propagation algorithm is: the error term of a neuron in layer l is the weighted sum of the error terms of all neurons in layer l+1 connected to the neuron, and then multiplied by the gradient of the neuron's activation function. After calculating the three partial derivatives, the gradient of the parameters is recalculated:

[0080]

[0081]

[0082] The main gradient descent methods used in deep learning are as follows:

[0083] 1. SGD

[0084] SGD stands for stochastic gradient descent. The difference between stochastic gradient descent and gradient descent is that gradient descent uses the entire data set for each update, while stochastic gradient descent uses only a batch of data at a time to calculate the gradient, rather than calculating the gradient for the entire data set. Because deep learning requires large amounts of data, it's unrealistic to calculate the gradient for all data points at once, as this would result in excessive computational time. Furthermore, calculating all gradients each time also loses some randomness, making it more susceptible to local errors. Therefore, while stochastic gradient descent doesn't always progress toward the true minimum, it can easily escape local minima.

[0085] The other gradient descent methods below are all based on SGD.

[0086] 2. Momentum

[0087] While stochastic gradient descent is in progress, momentum based on the previous gradient is added, which means that the parameter update is not only based on the current gradient, but also on the previous gradient.

[0088] 3. Adagrad

[0089] Adaptive learning rate, expressed as:

[0090]

[0091] As the number of runs increases, the learning rate decreases, and is affected by the gradient calculated each time. For parameters with large gradients, the learning rate becomes relatively small. ε is a smoothing parameter, usually set to 1e-4-1e-8 to avoid the denominator being zero. g is the gradient, and η is the learning rate.

[0092] 4. RMSprop

[0093] cache t =α*cachee t-1 +(1-α)(g t ) 2

[0094]

[0095] Compared with Adagrad, RMSprop has an additional decay α. In this way, all the previous squared gradients are no longer summed. Instead, a decay rate is used to gradually reduce them. A sliding average method is used. The earlier the gradient is generated, the less impact it has on the adaptive learning rate. This can more effectively avoid the problem of excessive decrease in Adagrad learning rate and enable faster convergence.

[0096] 5. Adam

[0097] This is a comprehensive learning method, which can be seen as a learning method of RMSprop plus Momentum, achieving better results than RMSprop.

[0098] In the training of the prediction network, the Adam optimizer is used for gradient descent.

[0099] The output of the prediction network is denormalized to obtain the prediction result, which is then compared with the actual data sequence to calculate indicators such as MAPE.

[0100] There are three evaluation criteria for the model: mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). Their definitions are as follows:

[0101]

[0102]

[0103]

[0104] The lower the value of the indicator, the better the model performance. In this example, the indicator we attach the most importance to is MAPE, and the core of the evaluation is also MAPE. For example, if the MAPE of Model 1 is lower than that of Model 2, but the RMSE of Model 1 is higher than that of Model 2, then we still believe that the performance of Model 1 is better than that of Model 2.

[0105] In this embodiment, the structure diagram of the prediction network is as follows: Figure 2 The structural parameters of the prediction network are shown in Table 1.

[0106] Table 1

[0107]

[0108]

[0109] In this embodiment, tensorflow is used as the backend, keras is used as the frontend, the training environment is NVIDIA GeForce RTX 2080Ti, the training time is 30 minutes, and the network status is basically converged, as shown in the following example: Figure 4 As shown in the figure, the RMSE of the training results is 239.564 and the MAPE is 3.651%. Figure 4 In the figure, the solid line is the predicted result, and the dotted line is the actual data.

[0110] An LSTM neural network is constructed as a comparison of the prediction network of this embodiment. The network structure parameters of the LSTM neural network are shown in Table 2:

[0111] Table 2

[0112] Input layer Receive input data, the input time series length is 100, and the number of features is 1 LSTM layer Predict future output sequences, the output feature number is 128 Output layer Output predicted time series

[0113] We also use tensorflow as the backend and keras as the frontend. The training environment is NVIDIA GeForce RTX2080 Ti. The training time is 30 minutes. The network status is basically converged. The RMSE of the training results is 277.115, and the MAPE is 4.329%. The prediction results are as follows: Figure 5 shown. Figure 5 In the figure, the solid line is the predicted result, and the dotted line is the actual data.

[0114] It can be seen that the prediction network of this embodiment has a good prediction effect and a high prediction accuracy.

[0115] The above description is merely a preferred embodiment of the present invention and therefore cannot be used to limit the scope of the present invention. In other words, equivalent changes and modifications made according to the scope of the patent application and the contents of the specification should still fall within the scope of the patent of the present invention.

Claims

1. A satellite link traffic prediction method based on LSTM and autoencoder, characterized by: The steps include: Step S1: original traffic data set D = {d1, d2, ..., d n-1 ,d n }Preprocess and obtain the data set n represents the total number of data points in the original traffic data set; Step S2: Data set D p Apply sliding window to split and get a set of multiple sequences And will collect Divide into training set and test set; Step S3, constructing an automatic encoder model with an encoder part and a decoder part, the encoder part includes a first input layer, a first LSTM layer and an intermediate hidden layer, the first LSTM layer is used to encode the input data into a feature vector contained in the intermediate hidden layer, the decoder part includes a first vector repetition layer, a second LSTM layer, a first BP neural network layer inserted in time sharing and a first output layer, the first vector repetition layer is used to repeat the feature vector contained in the intermediate hidden layer multiple times and input it into the second LSTM layer, the second LSTM layer and the first BP neural network layer inserted in time sharing are used to decode the feature vector; Step S4, pre-training the autoencoder model using the training set in step S2; Step S5, constructing an LSTM model and connecting it to the autoencoder pre-trained in step S4 to form a prediction network; Step S6: Use the training set in step S2 to train the prediction network.

2. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 1, characterized in that: In step S5, the LSTM model includes a second vector repetition layer, a second LSTM layer and a second BP neural network layer inserted in time sharing, and after the parameters of each layer are randomly initialized, they are connected to the encoder part of the automatic encoder.

3. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 1, characterized in that: In step S1, the original traffic data set D={d1, d2, ..., d n-1 ,d n The preprocessing specifically includes: first filling in the missing values ​​in the original traffic data set D, replacing the outliers, deleting the duplicate data, and then normalizing the data.

4. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 1, 2 or 3, characterized in that: The step S2 specifically includes the following steps: Step S21, determine the hyperparameter L x , L y and L total =L x +L y , L x L is the length of the sequence required for each prediction. y is the length of the sequence required for each prediction, and the length of the sliding window is L total ; Step S22: Sliding window segmentation from time point t=L x Start, to time t = nL x End, each time a split is performed, the sliding opening moves forward one data, and nL is obtained after the split total +1 sequence, each of length L total , thus obtaining a set of multiple sequences Step S23: Each sequence in the Lth x time points to obtain a length of L x and length L y Two sequences of nL total +1 length L x The sequence of data , nL total +1 length L y The sequence of constitutes the set Y; Step S24: Set E data And set Y is divided in the same proportion, and the corresponding set E is obtained data _train and set E data _test, set Y_train and set Y_test, set E data _train and set Y_train form the training set, set E data _test and the set Y_test form the test set.

5. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 4, characterized in that: In step S4, the data and labels of the training set input for the autoencoder model pre-training are all set E data _train, the data and labels of the test set are set E data _test, the loss function of pre-training is: Among them, E data _train re It is the data sequence restored by the automatic encoder.

6. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 5, characterized in that: In step S6, the data of the training set input into the prediction network training is E data _train, the label is Y_train, and the data of the test set is E data _test, the label is Y_test, and the training of the prediction network is achieved through back propagation algorithm and gradient descent.

7. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 6, characterized in that: The optimizer used in the training of the autoencoder model and the training of the prediction network is the Adam optimizer.

8. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 5, characterized in that: After the pre-training of the autoencoder model is completed, the decoder part is discarded and only the encoder part is retained, and the LSTM model is connected to the encoder part.

9. The satellite link traffic prediction method based on LSTM and autoencoder according to claim 3, characterized in that: For missing values ​​in the original traffic data set D, if the missing values ​​are located at the head or tail of the data set, the corresponding time points are directly deleted, and other missing values ​​are filled using linear interpolation.