A track data prediction method based on mixed long short-term memory network

Abstract

The present invention provides a method for predicting track data based on a frequency-mixing long short-term memory network, comprising: step 1, constructing a set of track data training samples; step 2, classifying the track data, setting the predicted track data with the same frequency as the low frequency, and setting the updated frequency higher than the predicted data as the high frequency, dividing the track data training samples into two groups of high and low, and separating the training set and the test set; step 3, using the training set as the input of the frequency-mixing long short-term memory network, and training to obtain the optimal prediction model; step 4, outputting the predicted data for the next time period. The present invention uses the long short-term memory model method to predict the track data, applies the long short-term memory model method to fit the track data, and outputs the data at the next time period point. It is of great significance for issues related to track data prediction.

Description

Technical Field

[0001] The invention relates to a track data prediction method based on a mixing long short-term memory network. Background Art

[0002] With the upgrading and development of sensors, the ability and technology to obtain track data have been improved, and people can obtain track data with richer dimensions and better real-time performance. When analyzing track data, there is still the problem of uneven data frequency due to uneven sensor deployment and varying accuracy. The imbalance is reflected in two aspects: the different acquisition frequencies of track data and the lag of low-frequency data. In traditional track data prediction methods, the frequencies of various types of data need to be the same. When the data frequencies are inconsistent, when the frequency of the dependent variable data is low, there will be insufficient data volume and a lack of other frequencies to participate in the prediction, which cannot be fully applied to different data acquisition frequencies. When using low-frequency data for prediction, the real-time performance of the data cannot be guaranteed, and the track data prediction cannot adapt to changes in track data.

[0003] In addition to traditional trajectory data prediction methods, models that use deep learning to process trajectory data, such as long short-term memory networks (LSTMs), have also been proposed to improve data prediction accuracy. However, LSTMs require structured data input, and when there are misaligned time points, high-frequency information in the mixed data is lost. Furthermore, they do not fully address the problem of predictive models adapting to real-world data fluctuations.

[0004] Therefore, when conducting track data prediction research, a prediction method is needed that can improve the real-time performance and accuracy of the predicted data without losing high-frequency data information. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a track data prediction method based on a mixed long short-term memory network to solve the problems of lack of high-frequency data to participate in the prediction, inability to adapt to data changes in real time, and insufficient data prediction accuracy when conducting track data prediction research.

[0006] Technical solution:

[0007] To achieve the above objectives, the present invention establishes a track data prediction method based on a mixed long short-term memory network, compiles predicted data into track data, selects associated track data of the data, and constructs a track data training sample set; classifies the associated track data into high and low frequency groups according to frequency, performs time alignment and data enhancement on different groups, and separates training sets and test sets; uses the mixed training samples as input to the mixed long short-term memory network, obtains predicted data after the mixed training samples are trained through two hidden layers, adjusts hyperparameters so that the average absolute error of the predicted data generated based on the training set meets the set threshold condition, and repeatedly trains to obtain a prediction model; and outputs predicted data for the next time period based on the prediction model and the test set data.

[0008] The specific technical solution includes the following steps:

[0009] Step 1: Select the track data that needs to be predicted as the dependent variable, select other track data of different frequencies as independent variables, and construct a mixed track data training sample set;

[0010] Step 2: Divide the mixed track data training sample set into a training set and a test set;

[0011] Step 3: Use the training set as the input of the mixed long short-term memory network to train and obtain the optimal prediction model;

[0012] Step 4: Based on the optimal prediction model and test set, output the prediction data for the next time period.

[0013] Step 1 includes: the track data includes longitude, latitude, altitude, speed, heading angle and vertical speed, and the track data to be predicted is selected as the dependent variable Y, and the dependent variable track data is expressed as Y: Where Y0 is the initial time, Y i is the i-th moment, For the Y i The dependent variable value at the moment, the update frequency of the dependent variable Y is low frequency, set the update frequency to m, that is, update once every m time periods, then the i-th moment is represented by t i =t0+i*m; select n other track data of different frequencies related to Y as independent variables. The update frequency of the independent variable data is higher than or equal to the dependent variable Y, that is, the independent variable data is updated faster than the dependent variable. Set the update frequency of the independent variable data to k, k≤m, and record the independent variable data set as X: {X (1) ,X (2) ,…,X (n)}, the set X contains the track data of n independent variables, where the independent variable X (n) Indicated as X (n) : Where τ0 is the initial time, τ j is the jth moment, is the value of the nth independent variable at the jth moment; X is the associated track data training sample, Y is the predicted data track data sample, and Y and X constitute a mixed track data training sample set.

[0014] Step 2 includes the following steps:

[0015] Step 2-1: Classify the associated track data, set the predicted data with the same frequency as low frequency, and the updated frequency higher than the predicted data as high frequency, and divide the track data training samples into two groups: high and low. Set X to have l data with the frequency equal to the predicted data frequency m, and r data with other high-frequency updates, and the corresponding high-frequency frequency is k, which is expressed as k: {k1, k2, ..., k r |k,k2,…,k r <m},k r is the update frequency of the rth high-frequency data;

[0016] Step 2-2: Standardize each variable of the track data training sample set using the z-score, and convert the original data into a standard score;

[0017] Step 2-3: Align the data grouped with the same update frequency according to the time point and track data. If there is no data at a time point, use the sliding least squares method to fill the missing value of each variable in a backward filling manner;

[0018] Step 2-4: Divide the independent variable data into training set and test set according to the training set time point and test set time point.

[0019] Step 2-2 includes: according to the time variable X of n independent variables (n) : is the value of the nth independent variable at the jth moment, standardized by z-score:

[0020]

[0021] in is the independent variable data after the z-score changes, is the mean of the time variable of n independent variables, σ (n) is the standard deviation of the independent variable, then the training sample of the single associated track data after the change is Z (n) : is the standardized value of the nth independent variable at the jth moment.

[0022] Step 2-3 includes: setting the track data of n independent variables to Z(n) : is the normalized value of the nth independent variable at the jth moment, and the frequency is set to k (n) , the jth moment is expressed as τ j =τ0+j*k (n) , if there is a time point τ i If there is no data, the sliding least squares method is used to fill the gaps, using the s-order polynomial f τ (τ) fits τ0 and τ j A time point τ i The sample value of is τ i The fitted value of :

[0023]

[0024] in is the kth data in τ i The sampling coefficient at the moment, v (n) is the fitted basis function of the nth dependent variable, is the s-order fitting basis function. When fitting, the sliding least squares weight ω is introduced i,j :

[0025]

[0026]

[0027] Where σ is the convergence coefficient, and the solution is calculated between τ0 and τ j between Obtain the dependent variable track data after filling in the gaps.

[0028] Steps 2-4 include: taking the dependent variable time after filling the gap according to the time period t: {t0,…,t i}, t i is the i-th moment, take t i The first five time periods are the test set time, and the rest are the training set time, that is, t0≤t xl <t i-4 is the training set time interval, t xl is the training set time, t i-4 ≤t cs ≤t i is the test set time interval, t cs is the test set moment, and the corresponding track data for training dependent variable data is Y xl : t xl The dependent variable data at the time, the track data of the dependent variable data used for testing is Ycs : t cs The dependent variable data at the time; Y cs is the test set;

[0029] For the associated track data training sample Z (n) , the independent variable data track data used for training is where τ l is the lth moment, is the nth independent variable in τ l The data at the moment, the independent variable data track data used for testing is where τ j is the jth moment, is the nth independent variable in τ j Data at the moment.

[0030] Step 3 includes the following steps:

[0031] Step 3-1: Track data based on the independent variable data used for training The variable data are divided into the time points τ0, τ1, ..., τ l Align to form a mixed data sample matrix, and use the aligned mixed training sample matrix as the input of the mixed long short-term memory network. According to the grouping in step 2-1, set represents the input of low-frequency variables at time τ, represents the input of high-frequency variables at time τ+k / m, and The timing information is propagated separately to the next moment and Design a long short-term memory network including a forget gate, an input gate, and an output gate, where f t+k / m is the output result of the forget gate, i t+k / m is the input gate result, is the output result of the candidate long short-term memory network, P t+k / m Output result of long short-term memory network, o t+k / m is the output result of the output gate, h t+k / m is the output result after passing through a single long short-term memory network, σ is the fully connected layer parameter, and tanh is the activation function:

[0032]

[0033] Where W f is the dependent variable Forget gate weight, U f is the independent variable Forget gate weight, is the output result after the previous long-short time network structure, b f is the forget gate bias term;

[0034]

[0035] W i is the dependent variable Input gate weight, U i is the independent variable Input gate weight, b i is the forget gate bias term;

[0036]

[0037] is the dependent variable Candidate output weights, is the independent variable Candidate output weights, is the candidate output bias term;

[0038]

[0039] Output the candidate results at time t+k / m Output at the previous moment And the forget gate output result f t+k / m The product of is added to obtain the long short-term memory network output result P t+k / m ;

[0040]

[0041] W o is the dependent variable Output gate weight, U o is the independent variable Output gate weight, b o is the output gate bias term;

[0042] h t+k / m =o t+k / m *tanh(P t+k / m )

[0043] The output gate result o at time t+k / m t+k / m And the output result P of the long short-term memory network after the activation function tanh is processed t+k / m Multiply them to obtain the dependent variable h at time t+k / m after processing by the long short-term memory network t+k / m , after k moments, the high-frequency dependent variable The result after processing is

[0044] High-frequency dependent variables processed by the long short-term memory network With the untreated low-frequency dependent variable After being processed by the activation function, the weight W of the long short-term memory network is h Processing, plus the long short-term memory network bias term b h , and get the output

[0045]

[0046] Step 3-2: After passing through the long short-term memory network, pass the penalty function Calculate the predicted value y is the true value of the dependent variable, W h is the weight of the long short-term memory network; and the test set Y cs Compare and calculate the mean absolute error

[0047]

[0048] in, is the predicted value at the i-th moment, ycs (i) is the actual value at the i-th moment, the regularization term coefficient λ of the penalty function is set from λ = 0, and is adjusted by increasing by 0.01 so that the mean absolute error of the predicted data generated based on the training set meets the set threshold condition

[0049] Step 3-3, adjust the weights of the output gate, input gate, forget gate, and long short-term memory network, and repeatedly train the mixed long short-term memory network to obtain the optimal prediction model. The optimal condition is that the penalty function reaches the minimum. The solidified prediction model at this time is recorded as

[0050] In step 3-2, the penalty function

[0051] Step 4 includes: according to the solidified prediction model and the test set Y cs , output the forecast data for the next time period

[0052] Beneficial effects:

[0053] In the case where the statistical frequency of track data is inconsistent, the present invention proposes to standardize the high-frequency data and map the high-frequency data characteristics to the low-frequency data, thereby solving the difficult problem of requiring the dependent variable and the independent variable to have the same frequency when making track data predictions. Without losing the high-frequency data characteristics, the data richness and effectiveness of the low-frequency data prediction are increased, thereby improving the prediction accuracy of the track data. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more apparent.

[0055] Figure 1 This is a flow chart of the data prediction method based on the mixed long short-term memory network.

[0056] Figure 2 It is a mapping diagram of high-frequency data features to low-frequency data of the mixed long short-term memory network.

[0057] Figure 3 This is a schematic diagram of a mixed long short-term memory network. DETAILED DESCRIPTION

[0058] Figure 1 This is a flow chart of the present invention's method for predicting data based on a mixed long short-term memory network. The contents of the specific modules include:

[0059] Step 1: Compile the prediction data into track data, select the associated track data of the data, and construct a track data training sample set. Set the data to be predicted, i.e. the dependent variable, as Y, and the dependent variable track data as Y: The update frequency of Y is low. Set the update frequency to m, that is, update once every m time periods, then t i =t0+i*m. Select n data related to Y as independent variable data. The update frequency of these data should be higher than or equal to the dependent variable Y. These independent variable data are recorded as X: {X (1) ,X (2) ,…,X (n)}, the track data variables of n dependent variables are expressed as where {τ0,…,τ j} are the dependent variable time points respectively. X is the associated track data training sample, Y is the predicted data track data sample, and Y and X constitute the track data training sample set.

[0060] Step 2: Classify the associated track data into high and low frequency groups according to frequency. Align the time points of the data in different groups according to frequency. Perform data enhancement on the low-frequency data using the sliding least squares method. Reserve the corresponding test time period according to the number of track data, and separate the training set and test set. The specific steps are as follows:

[0061] Step 2-1, classify the associated track data, set the predicted data with the same frequency as low frequency, and the update frequency higher than the predicted data as high frequency, and divide the track data training samples into two groups: high and low. Set X to have l data with the same frequency as the predicted data frequency m, and r data with other high frequency updates, and the corresponding high frequency frequency is k, that is, k: {k1, k2, ..., kr |k,k2,…,k r <m}。

[0062] Step 2-2: Standardize each independent variable of each track data using z-score. (n) : Then use z-score for standardization, that is is the variable data after the z-score changes, is the mean value of the independent variable track data, σ (n) is the standard deviation of the independent variable, then the training sample Z of the single associated track data after the change (n) :

[0063] Step 2-3: Align the data of the same group according to the time point and track data. If there is no data at a certain time point, use the sliding least squares method to fill the missing value of each variable in a backward filling manner. Set the track data of n dependent variables to Z (n) : Theoretically, there is data at every time point, and the frequency is set to k (n) , that is, τ j =τ0+j*k (n) , if there is a time point τ i If there is no data, the sliding least squares method is used to fill the gaps, using the s-order polynomial f τ (τ) fits τ0 and τ j At some point in time τ i Assume that is τ i The fitted value of .

[0064]

[0065] When fitting, weights are introduced:

[0066]

[0067]

[0068] Solve and calculate the j between The dependent variable track data after filling in the gaps can be obtained.

[0069] Step 2-4: Divide the independent variable data into training set and test set according to the training set time point and test set time point, and take the dependent variable according to the time period t: {t0,…,t i}, take t iThe first five time periods are the test set time, and the rest are the training set time, that is, t0≤t xl <t i-4 is the training set time interval, t i-4 ≤t cs ≤t i is the test set time interval. Then the corresponding track data for the dependent variable data used for training is Y xl : The trajectory data of the dependent variable data used for testing is Y cs : Similarly, for the associated track data training sample Z (n) For training, the independent variable data track data is The independent variable data track data used for testing is

[0070] Step 3: Use the mixed training sample as the input of the mixed long short-term memory network, pass through two layers of hidden layer neurons, adjust the number of neurons and penalty function of the second hidden layer neurons, so that the mean absolute error of the predicted data generated based on the training set meets the set threshold condition, and repeatedly train to obtain the optimal prediction model. Mixed long short-term memory network data processing architecture Figure 2 , the specific steps are as follows:

[0071] Step 3-1: The training set independent variable data organized according to steps 2-4 The variable data are divided into the time points τ0, τ1, ..., τ l Align to form a mixed data sample matrix, and use the mixed training samples as the input of the mixed long short-term memory network. The high-frequency data features of the mixed long short-term memory network are mapped to the low-frequency data. Figure 3 .

[0072] According to step 2-1 grouping, set represents the input of low-frequency variables at time τ, Represents the input of high-frequency variables at time τ+k / m. In this network structure, and The timing information is propagated separately to the next moment and According to the long short-term memory network,

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] h t+k / m =o t+k / m *tanh(C t+k / m )

[0079] According to the above output

[0080]

[0081] Step 3-2: After passing through the long short-term memory network, pass the penalty function Calculate the predicted value y is the true value of the dependent variable, W h is the weight of the long short-term memory network; and the test set Y cs Compare and calculate the mean absolute error

[0082]

[0083] The regularization coefficient λ of the penalty function is set from λ = 0 and is adjusted by increasing by 0.01 so that the mean absolute error of the predicted data generated based on the training set meets the set threshold condition.

[0084] Step 3-3: repeatedly train the mixed long short-term memory network to obtain the optimal prediction model.

[0085] Step 4: Based on the solidified prediction model and test set data and y cs , obtain the forecast data for the next time period

[0086] Example

[0087] The track data includes longitude, latitude, altitude, speed, heading angle and vertical speed, and speed is used as the track prediction variable.

[0088] Step 1: Compile the longitude, latitude, altitude, speed, heading angle and vertical speed data, and compile 1000 data in the same time period in chronological order.

[0089] Step 2-1: Use speed as the track prediction variable and classify the track data of the remaining independent variables. Heading angle and vertical speed are sensor data on the aircraft, with the same update frequency of ten seconds, and are classified into the high-frequency group. Longitude, latitude, altitude, and speed have the same update frequency of minutes and are classified into the low-frequency group.

[0090] Step 2-2: Perform z-score changes on the track data to achieve standardization;

[0091] Step 2-3: Use the sliding least squares method to fill in the missing values ​​of each track data in a backward manner.

[0092] Step 2-4: Select the 1st to 800th track data as the training set and the 801st to 1000th track data as the test set.

[0093] Step 3-1: Align the variable data according to the set of time points to form a mixed data sample matrix, and use the mixed training samples as the input of the mixed long short-term memory network;

[0094] Step 3-2: Adjust the model's hyperparameters. The regularization coefficient of the loss function is set to λ = 0.04. The RSME obtained on the test set is 0.414, which meets the set threshold of 0.5.

[0095] Step 3-3: repeatedly train the mixed long short-term memory network to obtain the optimal prediction model.

[0096] Step 4: Based on the solidified prediction model and test set data, obtain the speed data at the next time point.

[0097] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium is capable of storing a computer program that, when executed by the data processing unit, executes the invention of a method for track data prediction based on a mixed long short-term memory network provided by the present invention, as well as some or all of the steps in each embodiment. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0098] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, in essence or in other words, the part that contributes to the prior art, can be embodied in the form of a computer program, i.e., a software product. The computer program software product can be stored in a storage medium and includes several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, a MUU, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0099] The present invention provides a method for track data prediction based on a mixed long short-term memory network. There are numerous methods and approaches for implementing this technical solution. The above is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Any components not specified in this embodiment may be implemented using existing technologies.

Claims

1. A track data prediction method based on a mixed long short-term memory network, characterized in that: The steps include: Step 1: Select the track data that needs to be predicted as the dependent variable, select other track data of different frequencies as independent variables, and construct a mixed track data training sample set; Step 2: Divide the mixed track data training sample set into a training set and a test set; Step 3: Use the training set as the input of the mixed long short-term memory network to train and obtain the optimal prediction model; Step 4: Output the forecast data for the next time period based on the optimal forecast model and test set; Step 1 includes: the track data includes longitude, latitude, altitude, speed, heading angle and vertical speed, and the track data to be predicted is selected as the dependent variable , the dependent variable track data is expressed as ,in is the initial moment, For the A moment, For the The value of the dependent variable at the moment, the dependent variable The update frequency is low, set the update frequency to ,Right now The time period is updated once, then The moment is expressed as ; Select and Related Other track data of different frequencies are used as independent variables, and the update frequency of independent variable data is higher than or equal to that of dependent variable , set the independent variable data update frequency to , , let the independent variable data set be ,gather Include The track data of independent variables, where the independent variables Expressed as ,in is the initial moment, For the A moment, For the The independent variable in The value at each moment; X is the associated track data training sample, Y is the predicted track data sample, and Y and X constitute a mixed track data training sample set; Step 2 includes the following steps: Step 2-1: Classify the associated track data, set the predicted data with the same frequency as low frequency, and the update frequency higher than the predicted data as high frequency, and divide the track data training samples into high and low groups: Set X, the frequency is equal to the predicted data frequency for Other frequently updated data are , corresponding to the high frequency , expressed as , For the The update frequency of high-frequency data; Step 2-2: Standardize each variable of the track data training sample set using the z-score, and convert the original data into a standard score; Step 2-3: Align the data grouped with the same update frequency according to the time point and track data. If there is no data at a time point, use the sliding least squares method to fill the missing value of each variable in a backward filling manner; Step 2-4: Divide the independent variable data into training set and test set according to the training set time point and test set time point.

2. The method according to claim 1, characterized in that Step 2-2 includes: time variable of the independent variable , For the The independent variable in The values ​​at each moment are standardized using z-score: ， in is the independent variable data after the z-score changes, for The mean of the time variable of the independent variable, is the standard deviation of the independent variable, then the training sample of the changed single associated track data is , For the The independent variable in The normalized value at that moment.

3. The method according to claim 2, characterized in that Steps 2-3 include: Setting The track data of the independent variables is , For the The independent variable in The value after normalization at each moment, setting the frequency to , No. The moment is expressed as , if there is a time point If there is no data, the sliding least squares method is used to fill the gaps. Order polynomial Fitting and A point in time The sample value of for The fitted value of : ， in For the The data in The sampling coefficient at time t, For the The fitted basis functions of the dependent variables, for Order fitting basis function, when fitting, introduce sliding least squares weight : ， ， in is the convergence coefficient, and the solution is calculated in and between , and obtain the dependent variable track data after filling in the gaps.

4. The method according to claim 3, characterized in that Steps 2-4 include: taking the dependent variable time after filling the gap and calculating the time period as , For the A moment, take The first five time periods are the test set time, and the rest are the training set time, that is, is the training set time interval, is the training set time, is the test set time interval, is the test set moment, and the corresponding track data of the dependent variable data used for training is , for The dependent variable data at the moment, the track data of the dependent variable data used for testing is , for The dependent variable data at the moment; is the test set; For the associated track data training samples , the independent variable data track data used for training is ,in For the A moment, For the The independent variable in The data at the moment, the independent variable data track data used for testing is ,in For the A moment, For the The independent variable in Data at the moment.

5. The method according to claim 4, characterized in that Step 3 includes the following steps: Step 3-1: Track data based on the independent variable data used for training , the data of each variable are collected according to the time point Align to form a mixed data sample matrix, and use the aligned mixed training sample matrix as the input of the mixed long short-term memory network. According to the grouping in step 2-1, set Indicates the low-frequency variable at time Input, Indicates that high-frequency variables are at time Input, and The timing information is propagated separately to the next moment and , design a long short-term memory network including forget gate, input gate and output gate, where Output result of the forget gate, is the input gate result, Output the result for the candidate long short-term memory network, Output result for long short-term memory network, Output result for the output gate, To output the result after passing through a single long short-term memory network, is the fully connected layer parameter, is the activation function: ， in is the dependent variable Forget gate weight, is the independent variable Forget gate weight, Output result after the previous long-short time network structure, is the forget gate bias term; ， is the dependent variable Input gate weight, is the independent variable Input gate weight, is the forget gate bias term; ， is the dependent variable Candidate output weights, is the independent variable Candidate output weights, is the candidate output bias term; ， Will Candidate output results at time Output at the previous moment And the forget gate output result Add the product of to get the long short-term memory network output result ; ， is the dependent variable Output gate weight, is the independent variable Output gate weight, is the output gate bias term; ， Will Output gate result at time With the activation function The processed long short-term memory network output results Multiply and get the long short-term memory network processed Time dependent variable , after k moments, the high-frequency dependent variable The result after processing is ; High-frequency dependent variables processed by the long short-term memory network With the untreated low-frequency dependent variable After processing through the activation function, through the long short-term memory network weight Processing, plus the long short-term memory network bias term , and get the output : ； Step 3-2: After passing through the long short-term memory network, pass the penalty function Calculate the predicted value , is the true value of the dependent variable, is the weight of the long short-term memory network; and the test set Compare and calculate the mean absolute error : ， in, For the The predicted value at a moment, For the The actual value at the moment, the regularization coefficient of the penalty function from Start setting, and adjust by increasing by 0.01 at a time so that the mean absolute error of the predicted data generated based on the training set meets the set threshold condition ; Step 3-3, adjust the weights of the output gate, input gate, forget gate, and long short-term memory network, and repeatedly train the mixed long short-term memory network to obtain the optimal prediction model. The optimal condition is that the penalty function reaches the minimum. The solidified prediction model at this time is recorded as .

6. The method according to claim 5, characterized in that In step 3-2, the penalty function .

7. The method according to claim 6, characterized in that Step 4 includes: according to the solidified prediction model and test set , output the forecast data for the next time period .

Images