CN110309603A - A kind of short-term wind speed forecasting method and system based on wind speed characteristics - Google Patents

Abstract

The present invention discloses a kind of short-term wind speed forecasting method and system based on wind speed characteristics.The prediction technique includes: to obtain original wind series；Singular value decomposition is carried out to the wind series, obtains denoising sequence and noise remainder；Obtain the optimal modal number that variation mode decomposition is carried out to the denoising sequence；Variation mode decomposition is carried out to the denoising sequence, obtains each vector sequence；Information relevant to output wind speed height in the environmental factor is obtained using improved Mean Impact Value method, and establishes shot and long term memory network model in conjunction with the vector sequence or remainder；Optimize network model by improved the method for optimizing extremums, obtains prediction model；Each vector sequence and the noise sequence are predicted by the prediction model, obtain each component and remainder prediction result；Add up each component and remainder prediction result, obtains forecasting wind speed result；Wind speed forecasting method or system of the invention significantly improves prediction model reliability, can get high-precision forecast result.

Description

Short-term wind speed prediction method and system based on wind speed characteristics

Technical Field

The invention relates to the field of wind speed prediction, in particular to a short-term wind speed prediction method and system based on wind speed characteristics.

Background

In recent years, global energy situation is becoming more severe and energy demand is increasing. It is reported by the International Energy Agency (IEA) that the global energy demand has increased by 2.3% in 2018, which is the fastest increasing year in the past decade. However, as non-renewable resources are continuously consumed, environmental issues are continuously increased, and in order to keep balance between supply and demand and improve global environment, energy systems should be shifted to more sustainable ones. The core of energy system transformation is the development of renewable energy, and wind energy is one of the renewable energy, and the development speed should be accelerated.

At the present stage, global wind energy resources are rich, the cost of wind power generation is greatly reduced, and the wind power generation is rapidly developed. The Global Wind Energy Council (GWEC)' 2018 global wind power development report indicates that the installed capacity of the global wind energy industry in 2018 is 51.3GW, the total installed capacity is 519GW, and the installed capacity of China is the first in the world. The difficulty of real-time scheduling of a power grid is aggravated by accessing large-scale wind power into a power system, and the challenges are brought to the safety and stability of the power grid. The prediction of the wind power is one of the main ways to solve the problems, so that the prediction research of the wind power is very important for all countries in the world. And because the accurate wind speed prediction is favorable for solving the problems of wind power output power control, power grid safe and economic dispatching and the like, the wind speed prediction has very important research significance and value.

Disclosure of Invention

The invention aims to provide a short-term wind speed prediction method and system based on wind speed characteristics, which are used for obtaining a high-precision prediction result and improving the reliability of a prediction model.

In order to achieve the purpose, the invention provides the following scheme:

a method of short term wind speed prediction based on wind speed characteristics, the method comprising:

acquiring an original wind speed sequence; carrying out singular value decomposition on the wind speed sequence to obtain a denoising sequence and a noise residual term;

obtaining an optimal modal number for carrying out variational modal decomposition on the denoising sequence according to a modal number optimization method;

decomposing the denoising sequence according to a variational modal decomposition method of the optimal modal number to obtain a plurality of component sequences;

acquiring other environmental factors influencing wind speed prediction, acquiring information with high correlation with output wind speed in the environmental factors by using an improved average influence value method, and establishing a long-term and short-term memory network model by combining the component sequence or the remainder;

optimizing the network model through an improved extreme value optimization algorithm to obtain an optimized prediction model;

predicting the component sequences through the prediction model to obtain component prediction results;

predicting the noise remainder through the prediction model to obtain a remainder prediction result;

and accumulating the component and remainder prediction results to obtain a wind speed prediction result.

Optionally, the acquiring a denoising sequence and a noise residual specifically includes:

acquiring original wind speed data;

constructing a Hankel matrix for the wind speed sequence, acquiring singular values of the Hankel matrix through a singular value decomposition method, and sequencing the singular values from large to small; acquiring a difference spectrum of the singular values;

and determining the number of effective singular values according to the maximum catastrophe point in the singular value difference spectrum, reconstructing the wind speed signal, and acquiring a denoising signal and a noise remainder.

Optionally, the obtaining an optimal modal number for performing variational modal decomposition on the denoising sequence includes:

constructing a constraint variation model under the current modal number for the denoising sequence;

introducing a secondary penalty factor and a Lagrange multiplication operator, and changing the constraint problem into a non-constraint problem;

the self-adaptive decomposition of the signals is realized through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result under the condition of the mode number; if not, adjusting the center frequency and the bandwidth of the modal component until the Fourier transform of the modal component meets a convergence condition, and acquiring a decomposition result of the signal;

calculating the energy sum of each component under the current modal number;

and circulating the steps, and calculating the energy difference under the condition of the adjacent modal numbers, wherein the corresponding modal number is the optimal modal number of the acquired variational modal decomposition when the energy difference is obviously increased.

Optionally, performing variational modal decomposition for determining a modal number on the denoising sequence to obtain a plurality of component sequences, specifically including:

constructing a constraint variation model under the condition of an optimal modal number for the denoising sequence;

introducing a secondary penalty factor and a Lagrange multiplication operator, and changing the constraint problem into a non-constraint problem;

the self-adaptive decomposition of the signals is realized through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result of the signal; if not, adjusting the center frequency and the bandwidth of the modal component until a convergence condition is met, and acquiring a plurality of component sequences of the signal.

Optionally, the obtaining of other environmental factors affecting wind speed prediction, obtaining information with high correlation with output wind speed from the environmental factors by using an improved average influence value method, and establishing a long-term and short-term memory network model by combining the component sequences or the remainder specifically includes:

acquiring other environmental factors such as wind direction and temperature which influence wind speed prediction, and establishing a radial basis function neural network model for the environmental factors;

adding and subtracting 10% from each environmental factor to obtain a new sample; carrying out radial basis function neural network training on the sample, and obtaining an average influence value between the factors and a model output variable;

acquiring information with high correlation with output wind speed in the environmental factors according to the average influence value; normalizing the selected information to obtain a dimensionless data set;

acquiring a correlation coefficient matrix and an accumulated variance contribution rate of the data set, and selecting all principal components with the accumulated variance contribution rate larger than 85%;

and the principal component is combined with the component sequence or the remainder to establish a long-term and short-term memory network model.

Optionally, the obtaining information with high correlation with the output wind speed in the environmental factors by using an improved average influence value method, and predicting the information with the component sequence or the remainder sequence specifically include:

optimizing the network model through an improved extreme value optimization algorithm to obtain an optimized prediction model;

and determining a long-short term memory network model according to the optimal initialization parameters to obtain an optimized long-short term memory network prediction model.

Optionally, the predicting module predicts the component sequences to obtain component prediction results; predicting the noise remainder through the prediction model to obtain a remainder prediction result; accumulating the component and remainder prediction results to obtain a wind speed prediction result, which specifically comprises:

predicting each modal component through the prediction model to obtain a plurality of component prediction results;

predicting the remainder sequence through the prediction model to obtain a remainder prediction result;

and linearly superposing the component prediction results and the remainder prediction results to obtain a wind speed prediction result.

The invention also provides a short-term wind speed prediction system based on wind speed characteristics, comprising:

the wind speed data acquisition and separation module is used for acquiring an original wind speed sequence; carrying out singular value decomposition on the wind speed sequence to obtain a denoising sequence and a noise residual term;

the optimal modal number acquisition module is used for acquiring the optimal modal number for carrying out variational modal decomposition on the denoising sequence according to a modal number optimization method;

the variational modal decomposition module is used for decomposing the denoising sequence according to a variational modal decomposition method of the optimal modal number to obtain a plurality of component sequences;

the environment information screening module is used for acquiring other environment factors influencing wind speed prediction, acquiring information with high correlation with output wind speed in the environment factors by utilizing an improved average influence value method, and establishing a long-term and short-term memory network model by combining the component sequences or the remainder;

the parameter optimization module is used for optimizing parameters of the network model by using an improved extreme value optimization algorithm to obtain an optimized prediction model;

the prediction module is used for predicting the component sequences through the prediction model to obtain component prediction results; predicting the noise remainder through the prediction model to obtain a remainder prediction result; and accumulating each component prediction result and the remainder prediction result to obtain a wind speed prediction result.

Optionally, the wind speed data acquiring and separating module specifically includes:

the wind speed data acquisition unit is used for acquiring original wind speed data;

the singular value difference spectrum acquisition unit is used for constructing a Hankel matrix for the wind speed sequence, acquiring the singular value of the Hankel matrix through a singular value decomposition method and sequencing the singular value from large to small; acquiring a difference spectrum of the singular values;

and the signal reconstruction unit is used for determining the number of effective singular values according to the maximum catastrophe point in the singular value difference spectrum, reconstructing the wind speed signal and acquiring a denoising signal and a noise remainder.

Optionally, the optimal mode number obtaining module specifically includes:

the constraint variational model acquisition unit is used for constructing a constraint variational model under the current modal number for the denoising sequence;

the problem conversion unit is used for introducing a secondary penalty factor and a Lagrange multiplication operator and changing the constraint problem into an unconstrained problem;

the modal component acquisition unit is used for realizing the self-adaptive decomposition of the signal through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result under the condition of the mode number; if not, adjusting the center frequency and the bandwidth of the modal component until the Fourier transform of the modal component meets a convergence condition, and acquiring a decomposition result of the signal.

The energy acquisition unit is used for calculating the energy sum of each component under the current modal number;

and the energy difference acquisition unit is used for circulating the steps and calculating the energy difference under the condition of adjacent modal numbers, wherein the corresponding modal number is the optimal modal number of the acquired variational modal decomposition when the energy difference is obviously increased.

Optionally, the variational modal decomposition module specifically includes:

the constraint variational model construction unit is used for constructing a constraint variational model under the condition of the optimal modal number for the denoising sequence;

the problem conversion unit is used for introducing a secondary penalty factor and a Lagrange multiplication operator and changing the constraint problem into an unconstrained problem;

the modal component acquisition unit is used for realizing the self-adaptive decomposition of the signal through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result of the signal; if not, adjusting the center frequency and the bandwidth of the modal component until a convergence condition is met, and acquiring a plurality of component sequences of the signal.

Optionally, the environmental factor screening module specifically includes:

the system comprises an environmental factor acquisition unit, a radial basis function neural network model and a data processing unit, wherein the environmental factor acquisition unit is used for acquiring other environmental factors such as wind direction and temperature which influence wind speed prediction and establishing the radial basis function neural network model for the environmental factors;

the average influence value acquisition unit is used for adding and subtracting 10% to each environmental factor to acquire a new sample; carrying out radial basis function neural network training on the sample, and obtaining an average influence value between the factors and a model output variable;

the normalization unit is used for acquiring information with high correlation with the output wind speed in the environmental factors according to the average influence value; normalizing the useful information to obtain a dimensionless data set;

the principal component obtaining unit is used for obtaining a correlation coefficient matrix and an accumulated variance contribution rate of the data set and selecting all principal components with the accumulated variance contribution rate larger than 85%;

and the network model establishing unit is used for establishing a long-term and short-term memory network model by combining the principal component with the component sequence or the remainder.

Optionally, the parameter optimization module specifically includes:

a basic prediction model obtaining unit, configured to obtain a long-term and short-term memory network model;

the parameter optimization unit is used for optimizing the long-term and short-term memory network model through an improved extreme value optimization algorithm to obtain optimal initialization model parameters;

and the wind speed prediction model determining unit is used for determining the long-short term memory network model according to the optimal initialization parameters to obtain the optimized long-short term memory network wind speed prediction model.

Optionally, the prediction module specifically includes:

the component prediction unit is used for predicting each modal component through the prediction model to obtain a plurality of component prediction results;

the residual item prediction unit is used for predicting the residual item sequence through the prediction model to obtain a residual item prediction result;

and the superposition unit is used for linearly superposing the component prediction results and the remainder prediction results to obtain a wind speed prediction result.

Compared with the prior art, the invention has the following technical effects:

the wind speed prediction method and the system are a short-term wind speed prediction process based on wind speed characteristics. Firstly, a new mixed modal decomposition method is adopted to decompose a non-stationary time sequence into a plurality of relatively stationary modal components and a remainder sequence, noise information is separated from a non-stationary original wind speed sequence, and a variational modal decomposition method optimized by modal numbers is utilized to decompose a denoising sequence to obtain a plurality of components with obvious periodic characteristics. Meanwhile, environmental factors influencing wind speed are introduced to participate in prediction, an improved average influence value method is provided for screening variables participating in model input, model complexity is reduced, and prediction accuracy is improved. Finally, the invention selects the long-term and short-term memory network capable of memorizing the effective information for a long time as a prediction model, and optimizes the parameters of the long-term and short-term memory network by using the improved extreme value optimization algorithm, thereby accelerating the model training speed and improving the prediction precision. And predicting the component sequence and the remainder sequence by using the model, and accumulating the prediction result to obtain the final predicted wind speed. According to the method, the wind speed fluctuation and uncertainty are considered, the short-term wind speed prediction precision of the long and short-term memory network model is effectively improved, and meanwhile the reliability of the prediction model is improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

FIG. 1 is a flow chart of a short term wind speed prediction method based on wind speed characteristics according to embodiment 1 of the present invention;

FIG. 2 is a flowchart of a prediction method according to embodiment 2 of the present invention;

fig. 3 is a block diagram of a prediction system according to embodiment 3 of the present invention;

FIG. 4 is a graph of the distribution of the original wind speed data of the Sotavento wind power plant of the invention;

FIG. 5 is a graph showing the original wind speed data distribution of the Changma wind farm of the present invention;

FIG. 6 is a schematic diagram of a decomposition result of a denoising sequence of a Sotavento wind power plant;

FIG. 7 is a schematic diagram of a decomposition result of a denoising sequence of a Changma wind power plant;

FIG. 8 is a comparison box diagram of prediction errors before and after optimization of the Sotavento wind power plant long and short term memory network model parameters;

FIG. 9 is a comparison box diagram of prediction errors before and after optimization of the Changma wind power plant long and short term memory network model parameters;

FIG. 10 is a graph of predicted results of three models after the Sotavento wind farm of the invention;

FIG. 11 is a graph of the predicted results of the three models of the Changma wind farm.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

FIG. 1 is a flow chart of a wind speed prediction method according to the present invention. As shown in fig. 1, the short-term wind speed prediction method based on wind speed characteristics includes the steps of:

step 11: the method comprises the steps of obtaining an original wind speed sequence, carrying out singular value decomposition on the wind speed sequence, and obtaining a denoising sequence and a noise residual term.

At present, singular value decomposition is mainly used for noise reduction and periodic component extraction of vibration signals. In the invention, the noise information in the wind speed sequence is extracted by using the singular value decomposition difference spectrum method to obtain a noise reduction signal and a noise remainder. The method comprises the following specific steps:

step 111: and acquiring original wind speed data, and constructing a Hankel matrix for the wind speed sequence.

Step 112: and acquiring singular values of the Hankel matrix through a singular value decomposition method, and sequencing the singular values from large to small.

Step 113: and acquiring a difference spectrum of the singular values.

Step 114: and determining the number of effective singular values according to the maximum catastrophe point in the singular value difference spectrum, reconstructing the wind speed signal, and acquiring a denoising signal and a noise remainder.

Step 12: and obtaining the optimal modal number for carrying out variational modal decomposition on the denoising sequence according to a modal number optimization method.

In 2014, a Variational Modal Decomposition (VMD) method proposed by dragomirtski et al overcomes a series of defects such as modal aliasing and endpoint effect of Empirical Mode Decomposition (EMD), and has high operation efficiency and good noise robustness. The method seeks an optimal solution of a variation model in a complete non-recursive mode, and realizes effective separation of frequency domain decomposition of signals and each component according to the central frequency and bandwidth of each decomposition component. The research shows that the mode number has a large influence on the decomposition result, and the accuracy of the analysis result is influenced by too much or insufficient mode number. In order to make up for the defects of subjective selection of influencing parameters, the invention provides a modal number optimization method to overcome the defects caused by random selection of modal numbers. The components obtained by VMD decomposition are orthogonal, so that the sum of the energies of the components is equal to the energy of the original signal if the decomposition is proper from an energy point of view. When the mode number exceeds a certain proper value, imaginary components are generated, and the linear sum of the component energies is larger than the energy sum of the original decomposed components. The method comprises the following specific steps:

step 121: the denoising sequence constructs a constraint variational model under the current modal number: first, for each modal componentAndare respectively provided withRepresents u_k(t) an instantaneous amplitude and an instantaneous phase,a non-decreasing function, t is sampling time), and an analytic signal is obtained through Hilbert transformation; then, estimating the center frequency of each analytic signal, and transforming the frequency spectrum of each analytic signal to a fundamental frequency band; and finally, estimating the bandwidth of each modal component by using a Gaussian smooth index of the frequency shift signal, and constructing a constraint variation model shown in the formula (1).

In the formula,as component signal u_k(t) the center frequency of the frequency,representing partial derivatives of t over a function, j²-1, x (t) is the de-noising sequence.

Step 122: introducing a secondary penalty factor alpha and a Lagrange multiplication operator lambda (t), changing the constraint problem into an unconstrained problem, wherein the expanded Lagrange expression is as follows:

step 123: and obtaining an expression (3) through processes such as Fourier equidistant transformation and the like, and realizing the self-adaptive decomposition of the signal.

Wherein,are respectivelyThe fourier transform of (d). The center frequency is updated by the following equation:

step 124: and (4) iteratively updating until convergence meets the following conditions:

step 125: and calculating the energy sum of each component under the current modal number.

Step 126: and circulating the steps, calculating the energy difference under the condition of the adjacent modal numbers, and obtaining the optimal modal number of the fractal modal decomposition. The denoised signal energy and the energy difference are calculated by the following formula

In the formula, E_kFor the number of modes k, the sum of the energies of all components, E_k-1Is the sum of the energy values of the components after the last decomposition, eta_k，k-1Representing the energy difference, x_j(i) The j component sequence under the current mode number is shown, and n is the number of sampling points.

According to the formula (6), eta_k，k-1The smaller the value of (a), the less the signal may be decomposed; eta_k，k-1The larger the value of (A), the more prominent the over-resolution phenomenon of the VMD is. With increasing parameter k, over-decomposition occurs, η_k，k-1There will be a significant increase in the number of modes for VMD decomposition that can be represented by k-1 at the inflection point.

Step 13: carrying out variation modal decomposition of modal number determination on the denoising sequence to obtain a plurality of component sequences, which specifically comprises the following steps:

step 131: constructing a constraint variation model under the condition of an optimal modal number for the denoising sequence;

step 132: introducing a secondary penalty factor and a Lagrange multiplication operator, and changing the constraint problem into a non-constraint problem;

step 133: the self-adaptive decomposition of the signals is realized through processes such as Fourier equidistant transformation and the like;

step 134: performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result of the signal; if not, adjusting the center frequency and the bandwidth of the modal component until a convergence condition is met, and acquiring a plurality of component sequences of the signal.

Step 14: and acquiring information with high correlation with the output wind speed in the environmental factors by using an improved average influence value method, and predicting the information together with the component sequence or the remainder sequence.

And acquiring other environmental factors such as wind direction and temperature which influence the wind speed prediction, and selecting information with high correlation with the output wind speed from the environmental factors by utilizing an improved average influence value method. The existing mean-of-influence (MIV) method mainly adopts a Back Propagation (BP) neural network to calculate the correlation between the independent variable and the output variable, but because the BP neural network has high randomness and unstable result, a Radial Basis Function (RBF) neural network is adopted to replace the BP neural network. The method reduces the input dimensionality of a prediction model, improves the prediction performance, and is named as an improved average influence value (IMIV) method. The method comprises the following specific steps:

step 141: for the original data L (L) of environmental factors₁，l₂，…，l_p) Establishing a radial basis function neural network model, wherein p is the number of introduction factors;

step 142: for each factor l of the raw data_iPlus or minus 10% to obtain two new samples

Step 143: for newCarrying out radial basis function neural network training to respectively obtain training results y_i，z_iAnd calculate y_i，z_iThe difference value of (a) represents the variation value of the influence of the variation of the characteristic on the output result, and is denoted as a_i(see formula 8) calculating the example average to obtain the MIV value, which is denoted as b_i；

Step 144: repeating the step 141 and 143, calculating the MIV values of all the independent variables, and selecting the factor composition data set Y with a relatively large absolute value as { Y ═ Y-_ij1, 2, …, n.j 1, 2, …, m is data size, n is number of variables selected to participate in analysis.

Step 145: normalizing Y, the formula is shown in (9), and the result is expressed by X ═ X_ijY represents 1, 2, …, n.j, 1, 2, … m, and y represents_iIs the ith feature data set.

Step 146: and (4) calculating a correlation coefficient matrix R of X and the accumulated variance contribution rate eta, and selecting the number p of the main components with eta larger than or equal to 85%.

Where r is the data size, λ_iIs the ith eigenvalue of the correlation coefficient matrix R.

Step 15: and optimizing parameters of the long-short term memory network model (LSTM) through an improved extreme value optimization algorithm.

With the continuous development of deep learning technology, the concept of time series is applied in the structural design of the Recurrent Neural Network (RNN). Therefore, RNN has good time series analysis ability. The LSTM, as an improved RNN, inherits the RNN analysis capability on time sequence data, and makes up the deficiency of the RNN in long-term memory. The LSTM neural network model can effectively keep long-time memory, and certain results are obtained in the field of wind speed prediction. There are 3 gates in the memory unit of an LSTM neural network, which are the input gate, the forgetting gate, and the output gate, respectively. The LSTM update unit calculation process is as follows:

firstly, the information inputted at the t-th time passes through the input gate, and the value of the input layer of the memory unit is i_tCandidate value of hidden layer stateThen, the information passes through a forgetting gate, and the value of a forgetting layer of the memory unit is f_t: at this time, the hidden layer state update value is C_t: finally, the information is updated to o through the output gate, the value of the output layer and the output value of the final memory cell_t，h_tThe calculation formula is shown in a formula (11):

wherein x_tIs an input vector, h_tIs an output vector, α represents the gate activation function (typically a logical sigmoid function), is an element multiplication (Hadamard product) between two vectors, W represents the corresponding weight matrix, and b is the associated bias vector.

Therefore, a long-short term memory network model can be obtained. And then optimizing the long-term and short-term memory network model through an improved extreme value optimization algorithm to obtain optimal initialization model parameters. And determining a long-short term memory network model according to the optimal initialization parameters to obtain an optimized long-short term memory network prediction model.

The idea of the extremum optimization algorithm (EO) is derived from the theory of self-organizing criticality, which is characterized by non-equilibrium. Different from genetic algorithm, simulated annealing algorithm, ant colony algorithm, particle swarm optimization algorithm and the like, the algorithm cannot converge to a balanced state, and the generated fluctuation enables the algorithm to have better continuous searching and out-of-order optimal solution capability. The algorithm is easy to realize, small in calculation amount, good in algorithm effect and applicable to many projects. Due to the problem of long training time caused by free generation of LSTM initial parameters, the LSTM is optimized by combining the EO optimization idea, the training speed is accelerated, the LSTM prediction performance is improved, and the optimization algorithm is named as an improved mechanism optimization algorithm (IEO). The method comprises the following specific steps:

step 151: aiming at parameters needing to be trained by the LSTM, defining a proper population size, wherein a corresponding fitness function is f;

in the formula,x represents the predicted value and the actual value of the wind speed respectively, and n is the number of samples.

Step 152: initializing a parameter, setting the maximum number of iterations to T_maxRandomly generating an initial solution pop as 50;

step 153: calculating the fitness value f of the population, and sequencing the species pop according to the fitness value;

step 154: selection of species x with the smallest fitness value_mAnd x_m-1Generating random numbers rand according to a certain probability distribution_x1And rand_x2Replacing the original x_mAnd x_m-1；

Step 155: recalculating rand_x1And rand_x2The fitness value f;

if f (rand)_x1) > max (f), then x_max＝rand_x1，max(f)＝f(rand_x1) Updating population pop;

if f (rand)_x2) > max (f), then x_max＝rand_x2，max(f)＝f(rand_x2) Updating population pop;

step 156: if the stopping criterion is met, the algorithm stops and the result is output, otherwise, the step 153 is returned to for circular calculation.

Step 16: and predicting each modal component through the prediction model to obtain a plurality of component prediction results. And predicting the remainder sequence through the prediction model to obtain a remainder prediction result. And linearly superposing the component prediction results and the remainder prediction results to obtain a wind speed prediction result.

In order to verify that the method of the present invention has good prediction performance on the wind speed data of the actual wind farm, a data simulation experiment is performed by using the wind speed of the wind farm in china, and fig. 2 is a flowchart of the prediction method provided in this embodiment. As shown in fig. 2, the specific process includes:

step 21: obtaining an original wind speed sequence

In order to better verify the prediction capability of the method provided by the invention, two data sets are selected for short-term wind speed prediction respectively. The two data sets are respectively from Sotavento Spanish and Changma China in 2013 certain months, 250 sample data are respectively selected, wherein the first 200 data are used as a training set, and the last 50 data are used as a testing set. The two data sets in the text select different influencing factors and different sampling times, so that the wide applicability of the model in the text is verified. Sampling the wind speed of the Sotavento wind power plant once every 10 minutes, and selecting 7 factors of the wind speed, the wind direction, the air temperature, the air pressure, the specific volume, the specific humidity and the surface roughness to participate in prediction; the wind speed collection interval of the China Changma wind power plant is 5 minutes, and the environmental factors comprise wind direction, air temperature, motor rotating speed, pitch angle and daily generated energy accumulation. Wind speed is researched from the perspective of random volatility, and the variation characteristics of the wind speed of the two wind power plants are respectively shown in fig. 4 and fig. 5. As can be seen from the figure: the wind speed of the Sotavento wind power plant mostly fluctuates in the interval [6m/s, 16m/s ], has no obvious periodic characteristics and has strong random fluctuation; the wind speed of the China Changma wind power plant fluctuates at 6m/s and 14m/s, no obvious periodic characteristics exist, and the random fluctuation is strong, so that the research on the fluctuation characteristic of the wind speed is very important.

Step 22: mixed mode decomposition process

Step 221: in order to better extract the periodic characteristics in the original wind speed sequence, SVD denoising processing is carried out on the original wind speed, and a denoising sequence and a noise remainder are obtained.

Step 222: when k is 3-9, the energy and the energy difference of the denoising sequence are calculated by using a mode number selection method, and the result of the energy difference is shown in table 1. Thus, in both data sets, we have selected a VMD with a modal number of 7.

TABLE 1 energy Difference between different modal numbers

Step 223: in the Sotavento and chanma wind power plants, VMD decomposition is performed on the denoising sequence to obtain 7 modal components respectively as shown in fig. 6 and 7. From fig. 6 and 7, it can be seen that the periodic characteristics of each component are more obvious compared with the original wind speed sequence. At this point, the mixed mode decomposition process ends.

Step 224: to improve the prediction accuracy, an IMIV method is adopted to process multiple factors. First, correlation analysis is performed on multiple factors, and correlation between each factor and output is calculated, that is, an MIV value is calculated, and the results are shown in tables 2 and 3. Environmental factors which have small influence on wind speed prediction results of two wind power plants are filtered respectively, namely specific volumes and specific humidity of the 5 th and 6 th factors which have relatively small absolute values of the MIV of the Sotavento wind power plant are removed, and the 5 th factor of the Changma wind power plant is removed: the pitch angle.

TABLE 2 MIV values between factors and output sequences of Sotavento wind farm

TABLE 3 MIV values between various factors and output sequences of the Changma wind farm

Step 225: and principal component analysis is performed on other factors, so that the dimension of the model is reduced, redundant information is removed, and the calculation efficiency and the prediction precision are improved. And (3) respectively selecting all the principal components with the cumulative variance contribution rate of 85% by the two wind power plants as fixed input of the prediction model. The principal component analysis results are shown in table 4, and we can see that the characteristic values of the first two principal components of the Sotavento wind power plant and the chanma wind power plant are respectively: 1.00981, 0.99325, 1.03934 and 0.96260 all accord with principal component selection principles, and the cumulative variance contribution rates at the moment respectively reach 91.29% and 85.32%, so that the two wind farms select the first two principal components to participate in prediction.

Step 23: IEO optimization model prediction

Predicting future values of each component and the future values of the remainder respectively by utilizing an IEO optimized prediction model; and reconstructing an independent predicted value to obtain a final original wind speed prediction result.

Step 24: error index analysis

The prediction results are comprehensively evaluated by using the Mean Absolute Error (MAE), the absolute mean percentage error (MAPE), and the Root Mean Square Error (RMSE) as evaluation indexes (see formula 13). The smaller the application index value is, the closer the prediction result is to the actual wind speed value, and the higher the prediction performance of the model is.

Wherein V represents the sequence of actual wind speeds,representing the prediction sequence, n is the number of samples, and t is the sampling time.

TABLE 4 eigenvalue and cumulative variance contribution rates

Step 25: wind speed prediction results and error analysis

In the embodiment, LSTM, IEO optimized LSTM model, IEO-IMIV-LSTM and HMD based IEO-IMIV-LSTM model are compared to verify the IEO, the effectiveness of introducing multiple factors and performing IMIV processing, and the necessity and effectiveness of HMD decomposition on wind speed data;

step 251: compared with LSTM and IEO-LSTM models, the effect of IEO on optimizing LSTM is verified. We performed the analysis in terms of error magnitude and error fluctuation range:

(1) in Table 5, we present the MAE/MAPE/RMSE error results for the Sotavento and Chanma wind farms, from which it can be seen that the error index for the IEO-LSTM is less than the LSTM model in both farms compared to the original LSTM;

(2) the RMSE index is independently analyzed, and we see that in the prediction of a Sotavento wind power plant, the RMSE error index of an IEO-optimized LSTM model is reduced by 10.08%; similarly, the RMSE error index of the Changma wind power plant is reduced by 16.96 percent, so the IEO improves the prediction performance of the LSTM model;

(3) to more directly show the superiority of IEO, we plot error box plots between actual values and prediction errors for these two models in the prediction of the Sotavento and chanma wind farms, as shown in fig. 8 and 9. As is clear from FIG. 8, in the prediction of the Sotavento wind farm, the prediction error distributions of the IEO-LSTM are all closer to zero, and the error fluctuation range is reduced compared with the LSTM model. It can also be derived from fig. 9 that IEO optimizes the LSTM model in wind speed prediction for a chanma wind farm. This phenomenon indicates that, for the wind speed data of two wind farms, the IEO-optimized LSTM model can obtain higher accuracy in short-term wind speed prediction than the LSTM model, and has certain advantages in capturing the data fluctuation characteristics. In summary, we can conclude that IEO improves the predictive performance of LSTM neural networks.

Step 252: to verify the effects of IMIV and HMD, MATLAB was used to plot the wind speed predictions for IEO-LSTM, IEO-IMIV-LSTM, and the HMD-based IEO-IMIV-LSTM models, as shown in FIGS. 10 and 11. Wherein fig. 10 and 11 show the results of a Sotavento wind farm and a chanma wind farm, respectively. In each figure, the abscissa and ordinate represent the same meaning. The abscissa represents the prediction period and the ordinate represents the wind speed value. Wherein the band + curve represents the actual wind speed sequence, the band-line represents the IEO-LSTM model prediction result, the black curve represents the IEO-IMIV-LSTM model prediction result, and the band □ curve represents the HMD-based IEO-IMIV-LSTM model prediction result.

Step 2521: we compare the IEO-LSTM and IEO-IMIV-LSTM models, and verify the effect of introducing multiple factors and performing IMIV treatment.

(1) As can be seen from FIG. 10, in the Sotavento wind farm, the wind speed prediction curve of the IEO-IMIV-LSTM model is more trending closer to the original wind speed sequence; and the predicted result is numerically closer to the actual wind speed value than the IEO-LSTM model;

(2) as can be seen from fig. 11, in a chanma wind farm we can also derive: and multiple factors are introduced and processed, so that the prediction model is optimized, the stability of the prediction model is improved, and the prediction precision is improved.

By predicting the wind speeds of two different wind power plants, it can be seen that multiple factors are introduced into the wind speed prediction of the Sotavento and Changma wind power plants, and the prediction effect is improved in the aspects of numerical values and trends. At the same time, we calculate three error indicators for the two models, as shown in table 5. It can be seen that in the wind speed prediction of the Sotavento wind power plant, the error index of the IEO-IMIV-LSTM model is obviously reduced compared with the IEO-LSTM, and similarly, the same conclusion can be obtained in the Changma wind power plant as shown in Table 5. In conclusion, the effect of introducing multiple factors to IMIV treatment is further verified.

Step 2522: we compare the black curve in FIG. 10 with the curve with the band □, and can see that in the wind speed prediction of the Sotavento wind farm, the wind speed prediction result of the HMD-IEO-IMIV-LSTM model is closer to the actual wind speed in both value and trend compared with the IEO-IMIV-LSTM model. We can draw the same conclusion in the chanma wind farm comparing the black curve in fig. 11 with the curve with band □. Through the verification of the two data sets, the fact that HMD decomposition of the wind speed data can improve the prediction performance of the IEO-IMIV-LSTM neural network is fully demonstrated. Further analysis from the error perspective, in Sotavento and Chanma wind farms, the error index of HMD-IEO-IMIV-LSTM is smaller than that of the unprocessed IEO-IMIV-LSTM model of wind speed, which indicates that HMD decomposition of wind speed is essential for the IEO-IMIV-LSTM model provided herein. Thus, the validity and necessity of the HMID is verified.

TABLE 5 error index comparison results

Models 1-4 in Table 5 represent LSTM, IEO-IMIV-LSTM, HMD-IEO-IMIV-LSTM models, respectively.

FIG. 3 is a schematic structural diagram of a wind speed prediction system according to the present invention. As shown in fig. 3, the prediction system includes:

a wind speed data acquisition and separation module 31 for acquiring an original wind speed sequence; and carrying out singular value decomposition on the wind speed sequence to obtain a denoising sequence and a noise residual term.

The wind speed data acquiring and separating module 31 specifically includes:

the wind speed data acquisition unit is used for acquiring original wind speed data;

the singular value difference spectrum acquisition unit is used for constructing a Hankel matrix for the wind speed sequence, acquiring the singular value of the Hankel matrix through a singular value decomposition method and sequencing the singular value from large to small; acquiring a difference spectrum of the singular values;

and the signal reconstruction unit is used for determining the number of effective singular values according to the maximum catastrophe point in the singular value difference spectrum, reconstructing the wind speed signal and acquiring a denoising signal and a noise remainder.

And an optimal modal number obtaining module 32, configured to obtain an optimal modal number for performing variational modal decomposition on the denoising sequence according to a modal number optimization method.

The optimal mode number obtaining module 32 specifically includes:

the constraint variational model acquisition unit is used for constructing a constraint variational model under the current modal number for the denoising sequence;

the problem conversion unit is used for introducing a secondary penalty factor and a Lagrange multiplication operator and changing the constraint problem into an unconstrained problem;

the modal component acquisition unit is used for realizing the self-adaptive decomposition of the signal through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result under the condition of the mode number; if not, adjusting the center frequency and the bandwidth of the modal component until the Fourier transform of the modal component meets a convergence condition, and acquiring a decomposition result of the signal.

The energy acquisition unit is used for calculating the energy sum of each component under the current modal number;

and the energy difference acquisition unit is used for circulating the steps and calculating the energy difference under the condition of adjacent modal numbers, wherein the corresponding modal number is the optimal modal number of the acquired variational modal decomposition when the energy difference is obviously increased.

The variational modal decomposition module 33 is configured to decompose the denoising sequence according to a variational modal decomposition method of an optimal modal number to obtain a plurality of component sequences;

the variational modal decomposition module 33 specifically includes:

the constraint variational model construction unit is used for constructing a constraint variational model under the condition of the optimal modal number for the denoising sequence;

the problem conversion unit is used for introducing a secondary penalty factor and a Lagrange multiplication operator and changing the constraint problem into an unconstrained problem;

the modal component acquisition unit is used for realizing the self-adaptive decomposition of the signal through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result of the signal; if not, adjusting the center frequency and the bandwidth of the modal component until a convergence condition is met, and acquiring a plurality of component sequences of the signal.

And the environment information screening module 34 is configured to obtain other environment factors affecting wind speed prediction, obtain information with high correlation with an output wind speed from the environment factors by using an improved average influence value method, and perform wind speed prediction together with the wind speed sequence.

The environmental information screening module 34 specifically includes:

the system comprises an environmental factor acquisition unit, a radial basis function neural network model and a data processing unit, wherein the environmental factor acquisition unit is used for acquiring other environmental factors such as wind direction and temperature which influence wind speed prediction and establishing the radial basis function neural network model for the environmental factors;

the average influence value acquisition unit is used for adding and subtracting 10% to each environmental factor to acquire a new sample; carrying out radial basis function neural network training on the sample, and obtaining an average influence value between the factors and a model output variable;

the normalization unit is used for acquiring information with high correlation with the output wind speed in the environmental factors according to the average influence value; normalizing the useful information to obtain a dimensionless data set;

the principal component selection unit is used for acquiring a correlation coefficient matrix and an accumulated variance contribution rate of the data set and selecting all principal components with the accumulated variance contribution rate larger than 85%;

and the network model establishing unit is used for establishing a long-term and short-term memory network model by combining the principal component with the component sequence or the remainder.

And the parameter optimization module 35 is configured to obtain a long-term and short-term memory network model, optimize the network model through an improved extremum optimization algorithm, and obtain an optimized prediction model.

The parameter optimization module 35 specifically includes:

the parameter optimization unit is used for optimizing the long-term and short-term memory network model through an improved extreme value optimization algorithm to obtain optimal initialization model parameters;

and the wind speed prediction model determining unit is used for determining the long-term and short-term memory network model according to the optimal initialization parameters to obtain the optimized long-term and short-term memory network prediction model.

The prediction module 36 is configured to predict the component sequences through the prediction model to obtain component prediction results; predicting the noise remainder through the prediction model to obtain a remainder prediction result; and accumulating each component prediction result and the remainder prediction result to obtain a wind speed prediction result.

The prediction module 36 specifically includes:

the component prediction unit is used for predicting each modal component through the prediction model to obtain a plurality of component prediction results;

the residual item prediction unit is used for predicting the residual item sequence through the prediction model to obtain a residual item prediction result;

and the superposition unit is used for linearly superposing the component prediction results and the remainder prediction results to obtain a wind speed prediction result.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.

The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims (14)

1. A short-term wind speed prediction method based on wind speed characteristics, the prediction method comprising:

acquiring an original wind speed sequence; carrying out singular value decomposition on the wind speed sequence to obtain a denoising sequence and a noise residual term;

obtaining an optimal modal number for carrying out variational modal decomposition on the denoising sequence according to a modal number optimization method;

decomposing the denoising sequence according to a variational modal decomposition method of the optimal modal number to obtain a plurality of component sequences;

acquiring other environmental factors influencing wind speed prediction, acquiring information with high correlation with output wind speed in the environmental factors by using an improved average influence value method, and establishing a long-term and short-term memory network model by combining the component sequence or the remainder;

optimizing the network model through an improved extreme value optimization algorithm to obtain an optimized prediction model;

predicting the component sequences through the prediction model to obtain component prediction results;

predicting the noise remainder through the prediction model to obtain a remainder prediction result;

and accumulating the component and remainder prediction results to obtain a wind speed prediction result.

2. The short-term wind speed prediction method according to claim 1, wherein the obtaining of the denoising sequence and the noise residual specifically comprises:

acquiring original wind speed data;

constructing a Hankel matrix for the wind speed sequence, acquiring singular values of the Hankel matrix through a singular value decomposition method, and sequencing the singular values from large to small; acquiring a difference spectrum of the singular values;

and determining the number of effective singular values according to the maximum catastrophe point in the singular value difference spectrum, reconstructing the wind speed signal, and acquiring a denoising signal and a noise remainder.

3. The short-term wind speed prediction method according to claim 1, wherein the obtaining of the optimal number of modes for performing variational modal decomposition on the denoising sequence comprises:

constructing a constraint variation model under the current modal number for the denoising sequence;

introducing a secondary penalty factor and a Lagrange multiplication operator, and changing the constraint problem into a non-constraint problem;

the self-adaptive decomposition of the signals is realized through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result under the condition of the mode number; if not, adjusting the center frequency and the bandwidth of the modal component until the Fourier transform of the modal component meets a convergence condition, and acquiring a decomposition result of the signal.

Calculating the energy sum of each component under the current modal number;

and circulating the steps, and calculating the energy difference under the condition of the adjacent modal numbers, wherein the corresponding modal number is the optimal modal number of the acquired variational modal decomposition when the energy difference is obviously increased.

4. The short-term wind speed prediction method according to claim 1, wherein decomposing the de-noising sequence to obtain a plurality of component sequences according to a variational modal decomposition of an optimal modal number comprises:

constructing a constraint variation model under the condition of an optimal modal number for the denoising sequence;

introducing a secondary penalty factor and a Lagrange multiplication operator, and changing the constraint problem into a non-constraint problem;

the self-adaptive decomposition of the signals is realized through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result of the signal; if not, adjusting the center frequency and the bandwidth of the modal component until a convergence condition is met, and acquiring a plurality of component sequences of the signal.

5. The short-term wind speed prediction method according to claim 1, wherein other environmental factors affecting wind speed prediction are obtained, information with high correlation with output wind speed in the environmental factors is obtained by using an improved average influence value method, and a long-term and short-term memory network model is established by combining the component sequences or the remainder, specifically comprising:

acquiring other environmental factors such as wind direction and temperature which influence wind speed prediction, and establishing a radial basis function neural network model for the environmental factors;

adding and subtracting 10% from each environmental factor to obtain a new sample; carrying out radial basis function neural network training on the sample, and obtaining an average influence value between the factors and a model output variable;

acquiring information with high correlation with output wind speed in the environmental factors according to the average influence value; normalizing the information to obtain a dimensionless data set;

acquiring a correlation coefficient matrix and an accumulated variance contribution rate of the data set, and selecting all principal components with the accumulated variance contribution rate larger than 85%;

and the principal component is combined with the component sequence or the remainder to establish a long-term and short-term memory network model.

6. The short-term wind speed prediction method according to claim 1, wherein the optimizing the long-term and short-term memory network model through an improved extremum optimization algorithm to obtain an optimized prediction model specifically comprises:

optimizing the long-term and short-term memory network model through an improved extreme value optimization algorithm to obtain optimal initialization model parameters;

and determining a long-short term memory network model according to the optimal initialization parameters to obtain an optimized long-short term memory network prediction model.

7. The short-term wind speed prediction method according to claim 1, wherein the component prediction results are obtained by predicting a plurality of component sequences through the prediction model; predicting the noise remainder through the prediction model to obtain a remainder prediction result; accumulating the component and remainder prediction results to obtain a wind speed prediction result, which specifically comprises:

predicting each modal component through the prediction model to obtain a plurality of component prediction results;

predicting the remainder sequence through the prediction model to obtain a remainder prediction result;

and linearly superposing the component prediction results and the remainder prediction results to obtain a wind speed prediction result.

8. A short term wind speed prediction system, characterized in that the prediction system comprises:

the wind speed data acquisition and separation module is used for acquiring an original wind speed sequence; carrying out singular value decomposition on the wind speed sequence to obtain a denoising sequence and a noise residual term;

the optimal modal number acquisition module is used for acquiring the optimal modal number for carrying out variational modal decomposition on the denoising sequence according to a modal number optimization method;

the variational modal decomposition module is used for decomposing the denoising sequence according to a variational modal decomposition method of the optimal modal number to obtain a plurality of component sequences;

the environment information screening module is used for acquiring other environment factors influencing wind speed prediction, acquiring information with high correlation with output wind speed in the environment factors by utilizing an improved average influence value method, and establishing a long-term and short-term memory network model by combining the component sequences or the remainder;

the parameter optimization module is used for optimizing the network model through an improved extreme value optimization algorithm to obtain an optimized prediction model;

the prediction module is used for predicting the component sequences through the prediction model to obtain component prediction results; predicting the noise remainder through the prediction model to obtain a remainder prediction result; and accumulating each component prediction result and the remainder prediction result to obtain a wind speed prediction result.

9. The short term wind speed prediction system of claim 8, wherein the wind speed data acquisition and separation module specifically comprises:

the wind speed data acquisition unit is used for acquiring original wind speed data;

the singular value difference spectrum acquisition unit is used for constructing a Hankel matrix for the wind speed sequence, acquiring the singular value of the Hankel matrix through a singular value decomposition method and sequencing the singular value from large to small; acquiring a difference spectrum of the singular values;

and the signal reconstruction unit is used for determining the number of effective singular values according to the maximum catastrophe point in the singular value difference spectrum, reconstructing the wind speed signal and acquiring a denoising signal and a noise remainder.

10. The short-term wind speed prediction system according to claim 8, wherein the optimal mode number obtaining module specifically comprises:

the constraint variational model acquisition unit is used for constructing a constraint variational model under the current modal number for the denoising sequence;

the problem conversion unit is used for introducing a secondary penalty factor and a Lagrange multiplication operator and changing the constraint problem into an unconstrained problem;

the modal component acquisition unit is used for realizing the self-adaptive decomposition of the signal through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result under the condition of the mode number; if not, adjusting the center frequency and the bandwidth of the modal component until the Fourier transform of the modal component meets a convergence condition, and acquiring a decomposition result of the signal.

The energy acquisition unit is used for calculating the energy sum of each component under the current modal number;

and the energy difference acquisition unit is used for circulating the steps and calculating the energy difference under the condition of adjacent modal numbers, wherein the corresponding modal number is the optimal modal number of the acquired variational modal decomposition when the energy difference is obviously increased.

11. The short-term wind speed prediction system of claim 8, wherein the variational modal decomposition module specifically comprises:

the constraint variational model construction unit is used for constructing a constraint variational model under the condition of the optimal modal number for the denoising sequence;

the problem conversion unit is used for introducing a secondary penalty factor and a Lagrange multiplication operator and changing the constraint problem into an unconstrained problem;

the modal component acquisition unit is used for realizing the self-adaptive decomposition of the signal through processes such as Fourier equidistant transformation and the like; performing iterative updating, and judging whether the Fourier transform of the modal component meets a convergence condition; if so, acquiring a decomposition result of the signal; if not, adjusting the center frequency and the bandwidth of the modal component until a convergence condition is met, and acquiring a plurality of component sequences of the signal.

12. The short-term wind speed prediction system of claim 8, wherein the environmental information filtering module specifically comprises:

the system comprises an environmental factor acquisition unit, a radial basis function neural network model and a data processing unit, wherein the environmental factor acquisition unit is used for acquiring other environmental factors such as wind direction and temperature which influence wind speed prediction and establishing the radial basis function neural network model for the environmental factors;

the average influence value acquisition unit is used for adding and subtracting 10% to each environmental factor to acquire a new sample; carrying out radial basis function neural network training on the sample, and obtaining an average influence value between the factors and a model output variable;

the normalization unit is used for acquiring information with high correlation with the output wind speed in the environmental factors according to the average influence value; normalizing the useful information to obtain a dimensionless data set;

the principal component selection unit is used for acquiring a correlation coefficient matrix and an accumulated variance contribution rate of the data set and selecting all principal components with the accumulated variance contribution rate larger than 85%;

and the network model establishing unit is used for establishing a long-term and short-term memory network model by combining the principal component with the component sequence or the remainder.

13. The short term wind speed prediction system of claim 8, wherein the parameter optimization module specifically comprises:

the parameter optimization unit is used for optimizing the long-term and short-term memory network model through an improved extreme value optimization algorithm to obtain optimal initialization model parameters;

and the wind speed prediction model determining unit is used for determining the long-term and short-term memory network model according to the optimal initialization parameters to obtain the optimized long-term and short-term memory network prediction model.

14. The short term wind speed prediction system of claim 8, wherein the prediction module specifically comprises:

the component prediction unit is used for predicting each modal component through the prediction model to obtain a plurality of component prediction results;

the residual item prediction unit is used for predicting the residual item sequence through the prediction model to obtain a residual item prediction result;

and the superposition unit is used for linearly superposing the component prediction results and the remainder prediction results to obtain a wind speed prediction result.