CN116341717A - Wind speed prediction method based on error compensation - Google Patents

Abstract

The invention discloses a wind speed prediction method based on error compensation, which is characterized in that an ARMA model is established by utilizing historical data, the order determined by the ARMA model is used for dividing a data set, then a SVR model is trained by utilizing the divided data and wind speed prediction is carried out, and an error training set is obtained by subtracting a real wind speed value from a predicted wind speed value of the training set. And (3) the obtained error training set is re-utilized by the method, an ARMA model is established for the error sequence, and the bias autocorrelation coefficient is utilized as a basis to divide the error training set. After the error sequence is processed, the ELM is used for predicting the error, and the obtained error prediction result and the wind speed prediction result are added to obtain a wind speed prediction value after error compensation. The verification experiment proves that compared with the traditional direct prediction method, the error compensation prediction wind speed prediction method adopted by the invention has better fitting effect in the period of higher frequency and quicker wind speed change, and the accuracy of wind speed prediction is improved as a whole.

Description

Wind speed prediction method based on error compensation

Technical Field

The invention belongs to the technical field of wind power prediction, and can be used for predicting univariate wind speed time series data, in particular to an ARMA and SVR wind speed prediction method based on ELM error compensation.

Background

Human demand for energy has increased more and more since the 21 st century. However, combustion of fossil fuels such as coal, oil, and natural gas still occupies the main body of the world's energy structure. However, the combustion of fossil fuels entails the emission of greenhouse gases, harmful gases, which cause serious environmental problems, and, due to their non-renewable nature, the reserves of fossil energy on earth are becoming smaller and smaller, eventually depleted.

Scientists began to study emerging energy sources in order to address the energy crisis. At present, common new energy technologies are: wind power generation, hydroelectric power generation, photovoltaic power generation, tidal energy, nuclear energy, geothermal energy and the like, wherein the wind power generation technology is one of the new technologies with the fastest growth of installed capacity in recent years, and is more than photovoltaic power generation, and the wind power generation technology is a second most renewable energy source next to hydropower in the world energy market.

However, the wind often has instability and fluctuation, which causes difficulty in wind power utilization, large fluctuation of power generation power, difficulty in grid connection, and even damage to a fan when the wind speed changes rapidly, large-area off-grid caused by the damage to the fan, and influences the power generation quality, thereby endangering personnel safety. In order to stabilize the generated power and improve the generated quality, it is necessary to predict the wind speed.

Wind speed prediction is very important in the field of wind power generation. Early wind speed predictions typically use linear models, such as autoregressive moving average (ARMA), to find relationships between historical values, and after 90 s, artificial intelligence techniques are increasingly emerging, and many nonlinear methods are used to predict wind speed, such as Artificial Neural Networks (ANNs), support Vector Machines (SVMs), extreme Learning Machines (ELMs), long-short term memory networks (LSTM), convolutional Neural Networks (CNNs), etc., which typically build predictive models by taking historical wind speeds and other features affecting wind speed as inputs. Although these methods can achieve good predictive effects as a whole, in some time periods where the frequency is relatively high and the speed changes drastically, the predictive effects may not be ideal.

In order to further improve the prediction precision of the model, the application provides a prediction method based on ELM error compensation, wherein the ARMA modeling method is used for dividing historical wind speed data to serve as input to establish SVR model to predict future wind speed, then the error of the historical wind speed predicted value is used for establishing ELM model to predict future error value, and then the error value is added with the wind speed predicted value to obtain the wind speed predicted value after error compensation. Finally, through verification experiments, the method is proved to be capable of improving the accuracy of wind speed prediction on the whole.

Disclosure of Invention

In general, the more features it is for an object to be predicted, the more accurate the prediction effect. In wind speed prediction, wind speed is the object of prediction, the value of which is affected by factors such as ambient temperature, atmospheric pressure, ambient humidity, air density and the like, and if more data including these features is collected, the prediction is more accurate, but in actual measurement, many features are ignored and only the object to be predicted is left. For wind speed predictions, many wind farms collect only historical wind speeds and few other factors, so how to correctly select the number of input historical wind speeds for the predictive model becomes critical. Meanwhile, the conventional methods, such as ARMA, often have a problem that the wind speed prediction accuracy is lowered as the prediction time is prolonged. To this end, the invention provides a wind speed prediction method based on error compensation, comprising the steps of: .

(1) Obtaining historical wind speed data of a wind power plant, carrying out stability test on the sequence, modeling an ARMA model if the stability test is passed, and converting the unstable sequence into a stable sequence by adopting a difference method if the stability test is not passed;

(2) Using the stable sequence obtained in the step (1) for ARMA modeling, adopting an AIC criterion, and solving a partial autocorrelation coefficient p and an autocorrelation coefficient q which enable the AIC value to be minimum by using an iterative algorithm;

wherein, ARMA model is:

wherein alpha is _i Is an autoregressive coefficient, beta _j Is a moving average coefficient, both of which are predetermined coefficients and are non-zero, ε _t As an error term at time t, x _t Represents the value of time t, epsilon _t-j Is the error term at the time t-j, x _t-i A value at time t-i;

(3) Dividing the data set by using the p value obtained in the step (2) as a basis;

(4) SVR modeling is carried out, a Gaussian kernel function is selected by adopting a manual trial-and-error method, the value of a penalty factor is determined, then a model is trained, and the accuracy of model prediction is checked by using a test set;

(5) Predicting the training set by using the trained SVR model to obtain a wind speed predicted value of the training set, and subtracting the predicted wind speed value from the real wind speed value of the training set to obtain an error training set;

(6) The error training set is processed by adopting the same processing method as the wind speed training set, namely, the steps (1) and (2) are repeated on the error sequence, the ARMA modeling method is used, the stability of the error sequence is firstly judged, if the error sequence is unstable, differential processing is needed, an ARMA model is built, and an AIC criterion is utilized and an iterative algorithm is used for solving a partial autocorrelation coefficient p 'and an autocorrelation coefficient q' which enable the AIC value of the ARMA model of the error sequence to be minimum;

(7) Dividing the error data set by using the partial autocorrelation coefficient p' in the step (6);

the dividing process of the step (3) and the step (7) is the same, namely, from the first value to the p 'value of the sequence, a group of data is formed at intervals of p' units, then, from the second value of the sequence, p 'units are formed, and p' +1 values are formed as a second group of data, and the like, until the data set is divided. For each set of data, the first p '-1 value is taken as input and the p' th value is taken as output, i.e. the p 'th value is predicted using the first p' -1 value.

(8) And (3) training an ELM model by using the error training set in the step (7), and carrying out error prediction by using the model to obtain an error prediction value. And finally, adding the error predicted value and the wind speed predicted value in the step (4) to obtain the wind speed predicted value after error compensation.

Output y of ELM _i The method comprises the following steps:

wherein x is _i Is an input to ELM; n represents the input dimension, namely the number of neurons of the input layer; k is the number of neurons in the hidden layer; omega _j And b _j Weights and biases between the jth hidden layer neuron to the input, respectively; beta _j Weights between the jth hidden layer neuron and the output; g (·) is an activation function, a Sigmoid function or an RBF function can be selected, and the Sigmoid function is used in the invention;

(9) And finally, evaluating the accuracy of model prediction by using the root mean square error.

Root Mean Square Error (RMSE) is defined as:

wherein the method comprises the steps of

Representing the predicted value, x _i Representing the true value, N represents the total number of samples.

Preferably, in the step (1), the plateau sequence means:

wherein E (X) _t ) Representing sequence X _t Var (X) _t ) Representing variance, r _t,t-i And r _0,i Autocorrelation coefficients, μ, σ, respectively representing values at times t and t-i and times 0 and i ² These two numbers are constants.

Preferably, in the step (1), the stationary sequence means that the mean and variance of the sequence do not change with time, and the autocorrelation coefficients at the same time intervals are the same, that is:

wherein E (X) _t ) Representing sequence X _t Var (X) _t ) Representing variance, r _t,t-i And r _0,i Autocorrelation coefficients, μ, σ, respectively representing values at times t and t-i and times 0 and i ² Is a constant;

the difference method refers to:

y _t ＝Y _t -Y _t-1

wherein y is _t Represents the value after the difference at time t, Y _t And Y _t-1 The values at time t and time t-1 of the original data are shown, respectively. Preferably, in the step (2), the AIC criterion means:

AIC＝2k-2ln(L)

where k is the number of unknown parameters in the model and L is the maximum likelihood function value of the model.

Preferably, in the step (2), the ARMA model is:

wherein alpha is _i Is an autoregressive coefficient, beta _j Is a moving average coefficient, both of which are predetermined coefficients and are non-zero, ε _t As an error term at time t, x _t Represents the value of time t, epsilon _t-j Is the error term at the time t-j, x _t-i Is the value at time t-i.

Preferably, in the step (3), the autocorrelation function means:

wherein ρ is _t,t-k X represents _t And X _t-k Is used as a function of the autocorrelation function of (c),

representing sequence X _t Is a mean value of (c).

Preferably, in the step (3), the manner of dividing the training set and the test set is as follows: the first three quarters of the divided data are used as training sets, and the last quarter is used as testing set.

Preferably, in the step (4), the original model of SVR is:

wherein s.t. represents constraint conditions, C is penalty factor, m is sample number, and ζ _i And

is a relaxation variable, ε is a tolerance deviation, y _i Is the output, w and b are hyperplane parameters that need to be optimized.

Preferably, in the step (4), the gaussian kernel function is:

where x 'represents the origin of the kernel function, x represents the sampling point, |x-x' || ² Representing the euclidean distance between the sampling point and the origin, σ controls the correlation between equally spaced sampling points of the gaussian kernel.

Preferably, in the step (4), the Root Mean Square Error (RMSE) is defined as:

wherein the method comprises the steps of

Representing the predicted value, x _i Representing the true value, N represents the total number of samples.

Compared with the prior art, the invention has the following advantages:

1. the invention provides a wind speed prediction method based on error compensation, which is an ARMA and SVR combination method based on ELM error compensation, and can divide a univariate wind speed time sequence into a plurality of sections of sequences for model training. Compared with the data partitioning method by the trial-and-error method, the SVR model training set is optimized, the selection of the input quantity is more reasonable, the error compensation prediction is carried out by combining the ELM, the prediction error in the period with larger wind speed fluctuation is made up, and the good prediction effect is finally obtained.

2. According to the SVR model, different amounts of intercepted data are selected as input, so that compared with a trial-and-error method, the SVR model has a better prediction effect.

3. According to the invention, by adopting the ELM as an error compensation method, the historical error is predicted and added with the wind speed predicted value, so that the prediction effect in a period with higher frequency and faster wind speed change is improved.

Drawings

The invention is further described below with reference to the accompanying drawings.

FIG. 1 is wind speed data of a certain wind farm in China;

FIG. 2 is a sequence after the difference of the original wind speeds of the training set, wherein the sequence is a stable sequence;

FIG. 3 is a graph of autocorrelation and partial autocorrelation functions after differentiating the original data;

FIG. 4 is a flow chart of ARMA and SVR wind speed prediction method design based on ELM error compensation;

FIG. 5 shows the predicted effect of the trained model when p is 4;

FIG. 6 shows the predicted effect of the trained model when p is 2;

FIG. 7 shows the predicted effect of the trained model when p is 6;

FIG. 8 is error training set data;

FIG. 9 is error training set data after differencing;

FIG. 10 is a graph of autocorrelation and partial autocorrelation functions of an error training set;

FIG. 11 is an ELM error prediction result;

FIG. 12 is a final wind speed prediction based on ELM with error compensation.

Detailed Description

The invention will be described in further detail with reference to the drawings and the specific examples.

For wind speed prediction, the wind speed is generally affected by factors such as ambient temperature, atmospheric pressure, ambient humidity, air density and the like, and these characteristics can be used as the object of machine learning, but in the actually collected data, in consideration of cost or other factors, some wind farms generally cannot detect these factors, only the object to be predicted, namely the wind speed, is collected, and the finally obtained data becomes a univariate time series only containing the historical wind speed. For this data, it cannot be directly used as input to some methods such as SVR, so that the original wind speed sequence must be processed in order to predict wind speed using some more advanced machine learning method. The processed sequence is characterized by being only historical values of a plurality of wind speeds, so that the prediction effect is possibly poor in some places with larger frequency fluctuation in the prediction process, and an error compensation method can be adopted to compensate the part of errors.

The invention aims to treat a group of univariate wind speed sequences into sequences which can be input as SVR models through an ARMA modeling method, then predict future wind speeds by using the SVR models, process historical errors through ELM to obtain predicted values of errors, and finally add the predicted future wind speeds with the predicted errors to obtain future wind speeds after error compensation.

According to the method, an ARMA model is established by utilizing historical data, the order determined by the ARMA model is used for dividing a data set, then the SVR model is trained by utilizing the divided data, wind speed prediction is carried out, and the real wind speed value of the training set is subtracted from the predicted wind speed value to obtain an error training set. And (3) the obtained error training set is re-utilized by the method, an ARMA model is established for the error sequence, and the bias autocorrelation coefficient is utilized as a basis to divide the error training set. After the error sequence is processed, the ELM is used for predicting the error, and the obtained error prediction result and the wind speed prediction result are added to obtain a wind speed prediction value after error compensation.

Before dividing a data set, carrying out fixed order on the original data set by using an ARMA modeling method, dividing the data set by using the obtained value of the partial autocorrelation coefficient p to obtain a time sequence with stronger multi-section correlation, dividing the divided data into a training set and a test set, training the training set for an SVR model, then checking the prediction effect of the trained model by using the test set, finally subtracting a true value from a predicted value to obtain an error sequence, dividing the data by using the partial autocorrelation coefficient of the ARMA model for the same processing of the error sequence and the original wind speed data, carrying out error prediction by using ELM, and adding the error prediction result and the wind speed prediction result to obtain the wind speed predicted value after error compensation.

As shown in fig. 4, the invention adopts the ARMA and SVR combined prediction method with ELM error compensation, and specifically comprises the following steps:

step 1, collecting wind speed data of a wind power plant, carrying out ADF stability test, carrying out ARMA modeling if the ADF stability test passes the test, and carrying out differential processing if the ARMA modeling does not pass the test. The wind speed data of the wind farm is shown in fig. 1, and is obviously an unstable sequence, because the mean value and the variance of the wind farm are continuously changed, so that differential processing is performed, and in fig. 2, the sequence is a stable sequence, and ARMA modeling can be performed.

The plateau sequence refers to:

wherein E (X) _t ) Representing sequence X _t Var (X) _t ) Representing variance, r _t,t-i And r _0,i Autocorrelation coefficients, μ, σ, respectively representing values at times t and t-i and times 0 and i ² These two numbers are constants.

The difference method refers to:

y _t ＝Y _t -Y _t-1

wherein y is _t Represents the value after the difference at time t, Y _t And Y _t-1 The values at time t and time t-1 of the original data are shown, respectively.

And 2, using the stable sequence obtained in the step 1 for ARMA modeling, adopting an AIC criterion, and solving a partial autocorrelation coefficient p and an autocorrelation coefficient q which enable the AIC value to be minimum by using an iterative algorithm. In the invention, the partial autocorrelation coefficient p of the model is mainly utilized, because p reflects the correlation between the current value and the historical value, the value of p represents that the correlation degree of the current value and all the historical values of the hysteresis p number is large, and the p values are utilized to form a group of data, so that reference can be provided for dividing the data, the data waste can not be caused when the number of the divided data is selected, and the machine performance loss can not be caused because the data is selected too much.

The ARMA model in this embodiment is:

wherein alpha is _i Is an autoregressive coefficient, beta _j Is a moving average coefficient, both of which are predetermined coefficients and are non-zero, ε _t As an error term at time t, x _t Represents the value of time t, epsilon _t-j Is the error term at the time t-j, x _t-i Is the value at time t-i. From the above, it can be seen that the current value x _t With the first p values

Next, the p and q values that minimize the AIC values are found using AIC criteria, in combination with the iterative idea. The AIC criterion gives the best estimation of ARMA model order and parameters at the same time, and is suitable for the problem of less sample data. Wherein the expression of the AIC criterion is:

AIC＝2k-2ln(L)

where k is the number of unknown parameters in the model and L is the maximum likelihood function value of the model.

By using the above steps, the obtained p value can be used as a basis for dividing the data set.

And 3, utilizing the p interception data obtained in the step 2, wherein p obtained in the embodiment is 4. The specific intercepting method comprises the following steps: from the data at the first moment to the data at the p-th moment, each p values are used as a group of new data, the previous p-1 number is selected as input data, and the p-th number is used as output; and then, according to the same rule, carrying out the same dividing operation on the data from the second moment to the p+1st moment, and repeating the steps until the data set is completely divided.

And 4, SVR modeling is carried out, a manual trial and error method is adopted to determine ideal punishment factors and kernel function parameters, then the divided data set is utilized for training, the accuracy of the model is checked by using the test set, and the prediction effect is shown in figure 5. The original model of SVR is as follows:

wherein s.t. represents constraint conditions, C is penalty factor, m is sample number, and ζ _i And

is a relaxation variable, ε is a tolerance deviation, y _i Is the output, w and b are hyperplane parameters that need to be optimized, +.>

Is a kernel function, important in SVR, and the kernel function of this embodiment is implemented by using a gaussian kernel function (radial basis function), and its expression is:

where x 'represents the origin of the kernel function, x represents the sampling point, |x-x' || ² Representing the euclidean distance between the sampling point and the origin, σ controls the correlation between equally spaced sampling points of the gaussian kernel.

Step 5, in order to prove that the partitioning basis selected by the method is relatively good, a trial-and-error method is utilized, different p values are selected as the partitioning basis to retrain the model, the prediction result is compared with the result of the step 4, the accuracy of the prediction is judged by using root mean square error, and referring to fig. 6, p=2 and 7,p =6.

The comparison shows that when p=4, namely the p value solved in the step 2, the prediction effect is best, at the moment, the root mean square error is only 0.91226, when p=2, the root mean square error is 0.91539, when p=6, the root mean square error is 0.91718, and the root mean square error is higher than the model when p obtains the optimal value, therefore, by comparison, the data intercepted according to the p value determined by the ARMA model is proved to be used as input, the trained SVR model has better prediction effect, and the prediction precision is improved.

And 6, testing the training set by using the model trained in the step 4, wherein the aim of the step is to obtain a predicted value of the training set by using the predicted model, and subtracting the predicted wind speed value of the training set from the actual wind speed value of the training set, so that an error training set can be obtained to establish an error prediction model.

And 7, for the error training set obtained in the step 6, adopting the same processing method as that of the air speed training set, namely repeating the steps 1 and 2 for the error sequence, and firstly judging the stability of the error sequence by using an ARMA modeling method, and if the stability is unstable, carrying out differential processing. As shown in fig. 8, it is apparent that the error sequence is an unstable time sequence, and it is necessary to perform differential processing to make it stable. The differential error sequence is shown in fig. 9, where the sequence is a plateau sequence.

Step 8, building an ARMA model, and solving a partial autocorrelation coefficient p 'and an autocorrelation coefficient q' which minimize the AIC value of the error sequence ARMA model by using an AIC criterion and an iterative algorithm;

step 9, dividing an error data set by using the partial autocorrelation coefficient p' obtained in the step 8; this step is the same as the partitioning process of step 3, i.e. starting from the first value to the p-th value of the sequence, every p units as a set of data, starting from the second value of the sequence, then further partitioning by p units, to the p+1-th value as a second set of data, and so on, until the data set is partitioned. For each set of data, the first p-1 values are taken as inputs and the p-th value is taken as output, i.e., the p-th value is predicted using the first p-1 values.

And step 10, establishing a prediction model for the divided error training set obtained in the step 8 by using ELM, and predicting future errors. The output of ELM is:

wherein x is _i Is an input to ELM; n represents the input dimension, namely the number of neurons of the input layer; k is the number of neurons in the hidden layer; omega _j And b _j Weights and biases between the jth hidden layer neuron to the input, respectively; beta _j To the output for the jth hidden layer neuronThe weight of the space; g (·) is the activation function. Here the activation function selects the Sigmoid function, whose expression is:

the comparison of error using ELM prediction with true error is shown in fig. 11. It can be seen that the prediction error fits the true error well as a whole, only burrs occur at individual points, but the final prediction result is not greatly affected in view of the small unit of error.

And finally, adding the obtained error predicted value and the wind speed predicted value to obtain a final wind speed predicted value, wherein the result is shown in fig. 12, the root mean square error of a predicted model is 0.89609, which is lower than the root mean square error 0.91226 of fig. 5, and the wind speed predicted effect after error compensation prediction is better than that of the direct predicted graph 5.

The verification experiment proves that compared with the traditional direct prediction method, the error compensation prediction wind speed prediction method adopted by the invention has better fitting effect in the period of higher frequency and quicker wind speed change, and the accuracy of wind speed prediction is improved as a whole.

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. A wind speed prediction method based on error compensation, comprising the steps of:

(1) Obtaining historical wind speed data of a wind power plant, carrying out stability test on the sequence, modeling an ARMA model if the stability test is passed, and converting the unstable sequence into a stable sequence by adopting a difference method if the stability test is not passed;

(2) Using the stable sequence obtained in the step (1) for ARMA modeling, adopting an AIC criterion, and solving a partial autocorrelation coefficient p and an autocorrelation coefficient q which enable the AIC value to be minimum by using an iterative algorithm;

wherein, ARMA model is:

wherein alpha is _i Is an autoregressive coefficient, beta _j Is a moving average coefficient, both of which are predetermined coefficients and are non-zero, ε _t As an error term at time t, x _t Represents the value of time t, epsilon _t-j Is the error term at the time t-j, x _t-i A value at time t-i;

(3) Dividing the data set by using the p value obtained in the step (2) as a basis;

(4) SVR modeling is carried out, a Gaussian kernel function is selected by adopting a manual trial-and-error method, the value of a penalty factor is determined, then a model is trained, and the accuracy of model prediction is checked by using a test set;

(5) Predicting the training set by using the trained SVR model to obtain a wind speed predicted value of the training set, and subtracting the predicted wind speed value from the real wind speed value of the training set to obtain an error training set;

(6) The same processing method as that of the wind speed training set is adopted for the error training set, namely, the steps (1) and (2) are repeated for the error sequence, the ARMA modeling method is adopted, the stability of the error sequence is firstly judged, if the error sequence is unstable, differential processing is needed, an ARMA model is built, and the AIC criterion is utilized and an iterative algorithm is used for solving the partial autocorrelation coefficient p which enables the AIC value of the ARMA model of the error sequence to be minimum ^′ And autocorrelation coefficient q ^′ ；

(7) Using the partial autocorrelation coefficient p in step (6) ^′ Dividing the error data set;

wherein the division of step (3) and step (7) is the same, i.e. starting from the first value of the sequence to the p-th ^′ Every other p ^′ In units of a set of data, followed by a further p-interval starting from the second value of the sequence ^′ Units of p ^′ The +1 values are the second set of data, and so on until the data set is partitioned. Each group of data, the first p ^′ -1 value as input, p ^′ The value being taken as output, i.e. p before use ^′ -1 value, predict p ^′ A value.

(8) And (3) training an ELM model by using the error training set in the step (7), and carrying out error prediction by using the model to obtain an error prediction value. And finally, adding the error predicted value and the wind speed predicted value in the step (4) to obtain the wind speed predicted value after error compensation.

Output y of ELM _i The method comprises the following steps:

wherein x is _i Is an input to ELM; n represents the input dimension, namely the number of neurons of the input layer; k is the number of neurons in the hidden layer; omega _j And b _j Weights and biases between the jth hidden layer neuron to the input, respectively; beta _j Weights between the jth hidden layer neuron and the output; g (·) is an activation function, a Sigmoid function or an RBF function can be selected, and the Sigmoid function is used in the invention;

(9) And finally, evaluating the accuracy of model prediction by using the root mean square error.

Root Mean Square Error (RMSE) is defined as:

wherein the method comprises the steps of

Representing the predicted value, x _i Representing the true value, N represents the total number of samples.

Preferably, in the step (1), the plateau sequence means:

wherein E (X) _t ) Representing sequence X _t Var (X) _t ) Representing variance, r _t,t-i And r _0,i Autocorrelation coefficients, μ, σ, respectively representing values at times t and t-i and times 0 and i ² These two numbers are constants.

2. The method of claim 1, wherein in the step (1), the stationary sequences means that the mean and variance of the sequences do not change with time, and the autocorrelation coefficients at the same time intervals are the same, namely:

wherein E (X) _t ) Representing sequence X _t Var (X) _t ) Representing variance, r _t,t-i And r _0,i Autocorrelation coefficients, μ, σ, respectively representing values at times t and t-i and times 0 and i ² Is a constant;

the difference method refers to:

y _t ＝Y _t -Y _t-1

wherein y is _t Represents the value after the difference at time t, Y _t And Y _t-1 The values at time t and time t-1 of the original data are shown, respectively.

3. The method of claim 1, wherein in the step (2), the AIC criterion is:

AIC＝2k-2ln(L)

where k is the number of unknown parameters in the model and L is the maximum likelihood function value of the model.

4. The method of claim 1, wherein in the step (2), the ARMA model is:

wherein alpha is _i Is an autoregressive coefficient, beta _j Is a moving average coefficient, both of which are predetermined coefficients and are non-zero, ε _t As an error term at time t, x _t Represents the value of time t, epsilon _t-j Is the error term at the time t-j, x _t-i Is the value at time t-i.

5. The method of claim 1, wherein in the step (3), the autocorrelation function means:

wherein ρ is _t,t-k X represents _t And X _t-k Is used as a function of the autocorrelation function of (c),

representing sequence X _t Is a mean value of (c).

6. The method of claim 1, wherein in the step (3), the training set and the test set are divided in such a way that: the first three quarters of the divided data are used as training sets, and the last quarter is used as testing set.

7. The method of claim 1, wherein in the step (4), the original model of SVR is:

wherein s.t. represents constraint conditions, C is penalty factor, m is sample number, and ζ _i And

is a relaxation variable, ε is a tolerance deviation, y _i Is the output, w and b are hyperplane parameters that need to be optimized.

8. The method of claim 1, wherein in the step (4), the gaussian kernel function is:

where x 'represents the origin of the kernel function, x represents the sampling point, |x-x' || ² Representing the euclidean distance between the sampling point and the origin,

sigma controls the correlation between equally spaced sampling points of the gaussian kernel.

9. The method of claim 1, wherein in the step (4), the Root Mean Square Error (RMSE) is defined as:

wherein the method comprises the steps of

Representing the predicted value, x _i Representing the true value, N represents the total number of samples.

Images