CN116307139A - Wind power ultra-short-term prediction method for optimizing and improving extreme learning machine - Google Patents

Abstract

The invention discloses a wind power ultra-short-term prediction method for optimizing and improving an extreme learning machine, which comprises the following steps: step 1, analyzing wind power historical data, removing abnormal values based on a quartile method, complementing the missing values based on an interpolation method, determining to input the data as 16 sampling point data before the current moment, forming a data set by the normalized data, and dividing a training set and a testing set according to 8:2; step 2, performing kernel matrix weighted output on the limit learning machine by adopting poly and rbf kernel functions in an implicit layer before an output layer; step 3, optimizing a parameter set of the multi-core extreme learning machine by using a whale optimization algorithm to obtain adaptive parameters of a wind power ultra-short-term prediction model; step 4, measuringTest set input optimization and improved extreme learning machine establishes wind power ultra-short term prediction model, and outputs power value of wind power after half an hour

The wind power ultra-short-term prediction model of the optimal improved extreme learning machine can be established, and the prediction precision is effectively improved.

Description

Wind power ultra-short-term prediction method for optimizing and improving extreme learning machine

Technical Field

The invention relates to the technical field of new energy consumption, in particular to a wind power ultra-short-term prediction method for optimizing and improving an extreme learning machine.

Background

The wind power generation has the advantages of no pollution, reproducibility, easy acquisition and the like. But is greatly affected by the climate environment, and has strong randomness, volatility and instability. As wind power is largely connected to power systems, the structure of the power grid has changed, and the characteristics of "dual high" power systems are increasingly prominent. The novel power system under the high permeability of wind power is extremely challenged to safely and stably run, accurate wind power prediction is the key for solving the problem, and the prediction result is beneficial to wind power consumption, station operation and maintenance, scheduling decision, power market transaction and the like.

Many prediction models have been developed for ultra-short-term prediction of wind power, and the models can be divided into the following categories: physical methods and statistical learning methods. The physical method is used for establishing hydrodynamic and random differential equations according to the topography of the wind power plant to solve information such as weather and the like in the future, and is used for fitting wind power by combining a weather-power characteristic curve of a fan. The statistical learning method is used for carrying out feature analysis on wind power plant historical data, establishing a mapping relation between wind power features and prediction results, and common models comprise deep learning, a neural network, kalman filtering, statistical learning and the like, and the method is more used for ultra-short-term and short-term prediction tasks.

When wind power ultra-short-term prediction is performed, extremely high requirements are placed on the training time of the model, the running time of the prediction model is required to be smaller than the prediction time scale, and otherwise, the prediction meaning is lost. Extreme learning machines have received great attention due to their fast running speed, strong learning ability, and simple structure. However, the kernel function between the output layer and the hidden layer is unique, so that generalization capability and prediction precision are difficult to be achieved, and meanwhile, the robustness is poor. The multi-core extreme learning machine integrates a plurality of kernel functions into a network, and outputs the kernel functions in a weighted manner, so that the defect of the single-core extreme learning machine can be effectively overcome. Meanwhile, network parameters are optimized by using a whale optimization algorithm, so that optimal prediction is realized.

Disclosure of Invention

The invention aims to solve the technical problem of providing the wind power ultra-short-term prediction method for optimizing and improving the extreme learning machine, which can establish a wind power ultra-short-term prediction model of the optimal improved extreme learning machine and effectively improve the prediction precision.

In order to solve the technical problems, the invention provides a wind power ultra-short-term prediction method for optimizing and improving an extreme learning machine, which comprises the following steps:

step 1, analyzing wind power historical data, removing abnormal values based on a quartile method, complementing the missing values based on an interpolation method, determining to input the data as 16 sampling point data before the current moment, forming a data set by the normalized data, and dividing a training set and a testing set according to 8:2;

step 2, performing kernel matrix weighted output on the limit learning machine by adopting poly and rbf kernel functions in an implicit layer before an output layer;

step 3, optimizing a parameter set of the multi-core extreme learning machine by using a whale optimization algorithm to obtain adaptive parameters of a wind power ultra-short-term prediction model;

step 4, inputting the test set into an optimizing and improving extreme learning machine to establish a wind power ultra-short-term prediction model, and outputting the power value of the wind power after half an hour

Preferably, in step 1, eliminating abnormal values based on the quartile method specifically includes the following steps:

(11) All the historical power sequences are ordered from small to large, and the historical power is divided into a plurality of subsequences { P } according to 2MW ₁ ,P ₂ ,…,P _n }；

(12) And determining a reasonable power interval of each sub-sequence, and eliminating abnormal values outside the interval and abnormal values of working conditions.

[p _low ,p _up ]＝[Q ₁ -1.5ΔQ,Q ₃ +1.5ΔQ]

Wherein [ p ] _low ,p _up ]Threshold value representing reasonable output, Q ₁ Represents the 1 st quantile, Q ₃ Represents the 3 rd quantile, Δq=q ₃ -Q ₁ 。

Preferably, in step 1, the interpolation method is based on the deficiency value complement concretely as follows:

wherein p is _t The wind power with the missing value or the abnormal value is represented; Δt represents the step size, using 15min resolution data, the ultra-short term is a prediction within 4 hours, and 32 is the data within four hours before and after.

Preferably, in step 2, the performing a kernel matrix weighted output on the limit learning machine specifically includes:

wherein, C represents penalty parameter; t represents a target vector matrix; h represents a kernel matrix; k (·) represents a kernel function;

the calculation process of the multi-core function is as follows:

K _poly (x,x _i )＝(x,x _i +μ) ^λ

K(x,x _i )＝αK _rbf +(1-α)K _poly

wherein σ represents the kernel parameters of the rbf kernel; μ and λ represent nuclear parameters of the poly nucleus; alpha represents a weight coefficient.

Preferably, in step 3, optimizing a parameter set of the multi-core extreme learning machine by using a whale optimization algorithm to obtain adaptive parameters of a wind power ultra-short-term prediction model specifically comprises the following steps:

(31) Initializing whale optimization algorithm parameters, specifically setting population quantity 50;

(32) Determining optimization parameters of the multi-core extreme learning machine, specifically { C, alpha, sigma, mu, lambda };

(33) And calculating the fitness of the individual according to the following formula, and storing the current optimal individual and the current optimal position.

D＝|CL _t ^* -L _t |

L _t+1 ＝L _t ^* -AD

Wherein A and C represent constant coefficients; l (L) _t ^* Representing a current optimal position vector; l (L) _t Representing a position vector; d represents the current and optimal distance.

A＝2ar-a

C＝2r

Wherein r represents [0,1 ]]Random numbers of (a); a represents a control parameter; t is t _max Representing the maximum number of iterations;

(34) The population is optimized along the spiral path in the contracted circle:

wherein delta represents [0,1 ]]Random numbers of (a); q represents a random number of (0, 1); Δd= |l _t ^* -L _t I (I); b represents a constant for describing a spiral shape;

(35) If A is more than or equal to 1, the position is randomly selected and updated according to the current optimal update:

D＝|CL _t ^r -L _t |

L _t+1 ＝L _t ^r -AD

wherein L is ^r Representing a random position; a represents a control parameter; t is t _max Representing the maximum number of iterations;

(36) The maximum iteration condition is reached, and the optimal solution is the optimal parameters { C, alpha, sigma, mu, lambda }.

The beneficial effects of the invention are as follows: according to the invention, the wind power data are cleaned, abnormal values in the data are removed based on a quartile method and are supplemented based on an interpolation method, a plurality of kernel functions are utilized to optimize and improve the limit learning machine, meanwhile, a parameter set of the model is optimized based on a whale optimization algorithm, a wind power ultra-short-term prediction model of the optimal improved limit learning machine is established, and the prediction precision is effectively improved.

Drawings

FIG. 1 is a schematic flow chart of the method of the present invention.

FIG. 2 is a graph of the predicted outcome of the present invention in an example application.

Detailed Description

As shown in FIG. 1, the ultra-short-term prediction method for optimizing and improving the wind power of the extreme learning machine comprises the following steps:

step 1, analyzing wind power historical data, removing abnormal values based on a quartile method, complementing the missing values based on an interpolation method, determining to input the data as 16 sampling point data before the current moment, forming a data set by the normalized data, and dividing a training set and a testing set according to 8:2;

step 2, performing kernel matrix weighted output on the limit learning machine by adopting poly and rbf kernel functions in an implicit layer before an output layer;

step 3, optimizing a parameter set of the multi-core extreme learning machine by using a whale optimization algorithm to obtain adaptive parameters of a wind power ultra-short-term prediction model;

step 4, inputting the test set into an optimizing and improving extreme learning machine to establish a wind power ultra-short-term prediction model, and outputting the power value of the wind power after half an hour

The calculation process for eliminating abnormal values by the quartile method is as follows:

all the historical power sequences are ordered from small to large, and the historical power is divided into a plurality of subsequences { P } according to 2MW ₁ ,P ₂ ,…,P _n }。

And determining a reasonable power interval of each sub-sequence, and eliminating abnormal values outside the interval and abnormal values of working conditions.

[p _low ,p _up ]＝[Q ₁ -1.5ΔQ,Q ₃ +1.5ΔQ]

Wherein [ p ] _low ,p _up ]Threshold value representing reasonable output, Q ₁ Represents the 1 st quantile, Q ₃ Represents the 3 rd quantile, Δq=q ₃ -Q ₁ 。

The interpolation method is used for complementing the missing value or the abnormal value, and the calculation process is as follows:

wherein p is _t The wind power with the missing value or the abnormal value is represented; Δt represents the step size. The reason for selecting the front and rear 32 points is that the invention adopts data with 15min resolution, the ultra-short term is prediction within 4 hours, and the data within four hours is 32.

Training samples are constructed according to the historical power values of the first 16 sampling points at the current moment, and the training set vector is [ p ] _t-16 ,p _t-9 ,…,p _t-1 ,p _t+2 ]Sample set normalized according to 8:2 are divided into training and testing sets.

The optimization and improvement limit learning machine calculation process is as follows:

wherein, C represents penalty parameter; t represents a target vector matrix; h represents a kernel matrix; k (·) represents a kernel function.

The calculation process of the multi-core function is as follows:

K _poly (x,x _i )＝(x,x _i +μ) ^λ

K(x,x _i )＝αK _rbf +(1-α)K _poly

wherein σ represents the kernel parameters of the rbf kernel; μ and λ represent nuclear parameters of the poly nucleus; alpha represents a weight coefficient.

The parameter set of the multi-core extreme learning machine is optimized, and the parameter set comprises a core coefficient, a penalty coefficient, a core function weight coefficient and the like. The calculation process is as follows:

(31) Parameters of a whale optimization algorithm are initialized, and specifically population quantity 50 is set.

(32) And determining optimization parameters of the multi-core extreme learning machine, specifically { C, alpha, sigma, mu, lambda }.

(33) And calculating the fitness of the individual according to the following formula, and storing the current optimal individual and the current optimal position.

D＝|CL _t ^* -L _t |

L _t+1 ＝L _t ^* -AD

Wherein A and C represent constant coefficients; l (L) _t ^* Representing a current optimal position vector; l (L) _t Representing a position vector; d represents the current and optimal distance.

A＝2ar-a

C＝2r

Wherein r represents [0,1 ]]Random numbers of (a); a represents a control parameter; t is t _max Representing the maximum number of iterations.

(34) Population optimization along spiral path simultaneously in contracted circle

Wherein delta represents [0 ],1]Random numbers of (a); q represents a random number of (0, 1); Δd= |l _t ^* -L _t I (I); b represents a constant, describing a spiral shape.

(35) If A is more than or equal to 1, the position is randomly selected and updated according to the current optimal update

D＝|CL _t ^r -L _t |

L _t+1 ＝L _t ^r -AD

Wherein L is ^r Representing a random position; a represents a control parameter; t is t _max Representing the maximum number of iterations.

(36) And (5) reaching the maximum iteration condition, and obtaining an optimal solution, namely an optimal parameter.

So far, wind power ultra-short-term power prediction can be realized by utilizing the optimization and improvement extreme learning machine method.

Example 1:

to verify the effectiveness of the method of the invention, the following experiments were performed; and (3) performing example simulation by using real data of a wind farm with a installed capacity of 110MW in Jiangsu China, wherein the data resolution is 15min. The input data is power history data recorded 16 times before the predicted point is selected, a training set and a testing set are formed, and the training set is used for training and optimizing a prediction model according to the divided training set. And obtaining a prediction result through the test set data.

Deterministic prediction performance is typically model evaluated from two indicators: mean Absolute Percent Error (MAPE), root Mean Square Error (RMSE).

The mean absolute percentage error is defined as follows:

wherein, p represents the actual output value of wind power,

and C represents the installed capacity of the wind power plant, and n represents the number of samples.

The mean absolute percentage error is defined as follows:

table 1 optimization and improvement of extreme learning machine model prediction result evaluation index

TABLE 2 parameters after optimization of whale

The prediction results are shown in table 1, table 2 and fig. 2. As can be seen from FIG. 2, the prediction method of the optimizing and improving extreme learning machine has high prediction precision on wind power output power. In conclusion, the wind power prediction method can realize the prediction of wind power and can be used for practical engineering application.

Claims (5)

1. The wind power ultra-short-term prediction method for optimizing and improving the extreme learning machine is characterized by comprising the following steps of:

step 1, analyzing wind power historical data, removing abnormal values based on a quartile method, complementing the missing values based on an interpolation method, determining to input the data as 16 sampling point data before the current moment, forming a data set by the normalized data, and dividing a training set and a testing set according to 8:2;

step 2, performing kernel matrix weighted output on the limit learning machine by adopting poly and rbf kernel functions in an implicit layer before an output layer;

step 3, optimizing a parameter set of the multi-core extreme learning machine by using a whale optimization algorithm to obtain adaptive parameters of a wind power ultra-short-term prediction model;

step 4, inputting the test set into an optimizing and improving extreme learning machine to establish a wind power ultra-short-term prediction model, and outputting windPower value after half an hour of electric power

2. The ultra-short-term prediction method for optimizing and improving wind power of extreme learning machine according to claim 1, wherein in step 1, eliminating abnormal values based on a quartile method specifically comprises the following steps:

(11) All the historical power sequences are ordered from small to large, and the historical power is divided into a plurality of subsequences { P } according to 2MW ₁ ,P ₂ ,…,P _n }；

(12) And determining a reasonable power interval of each sub-sequence, and eliminating abnormal values outside the interval and abnormal values of working conditions.

[p _low ,p _up ]＝[Q ₁ -1.5ΔQ,Q ₃ +1.5ΔQ]

Wherein [ p ] _low ,p _up ]Threshold value representing reasonable output, Q ₁ Represents the 1 st quantile, Q ₃ Represents the 3 rd quantile, Δq=q ₃ -Q ₁ 。

3. The ultra-short-term prediction method for optimizing and improving wind power of extreme learning machine according to claim 1, wherein in step 1, the deficiency value is complemented based on interpolation method specifically comprises:

wherein p is _t The wind power with the missing value or the abnormal value is represented; Δt represents the step size, using 15min resolution data, the ultra-short term is a prediction within 4 hours, and 32 is the data within four hours before and after.

4. The ultra-short term prediction method for optimizing and improving wind power of extreme learning machine according to claim 1, wherein in step 2, the performing of the kernel matrix weighted output on the extreme learning machine is specifically:

wherein, C represents penalty parameter; t represents a target vector matrix; h represents a kernel matrix; k (·) represents a kernel function;

the calculation process of the multi-core function is as follows:

K _poly (x,x _i )＝(x,x _i +μ) ^λ

K(x,x _i )＝αK _rbf +(1-α)K _poly

wherein σ represents the kernel parameters of the rbf kernel; μ and λ represent nuclear parameters of the poly nucleus; alpha represents a weight coefficient.

5. The method for optimizing and improving wind power ultra-short-term prediction of an extreme learning machine according to claim 1, wherein in step 3, a parameter set of a multi-core extreme learning machine is optimized by using a whale optimization algorithm, and the method for obtaining adaptive parameters of a wind power ultra-short-term prediction model specifically comprises the following steps:

(31) Initializing whale optimization algorithm parameters, specifically setting population quantity 50;

(32) Determining optimization parameters of the multi-core extreme learning machine, specifically { C, alpha, sigma, mu, lambda };

(33) And calculating the fitness of the individual according to the following formula, and storing the current optimal individual and the current optimal position.

D＝|CL _t ^* -L _t |

L _t+1 ＝L _t ^* -AD

Wherein A and C represent constant coefficients; l (L) _t ^* Representing a current optimal position vector; l (L) _t Representing a position vector; d represents the current and optimal distance.

A＝2ar-a

C＝2r

Wherein r represents [0,1 ]]Random numbers of (a); a represents a control parameter; t is t _max Representing the maximum number of iterations;

(34) The population is optimized along the spiral path in the contracted circle:

wherein delta represents [0,1 ]]Random numbers of (a); q represents a random number of (0, 1); Δd= |l _t ^* -L _t I (I); b represents a constant for describing a spiral shape;

(35) If A is more than or equal to 1, the position is randomly selected and updated according to the current optimal update:

D＝|CL _t ^r -L _t |

L _t+1 ＝L _t ^r -AD

wherein L is ^r Representing a random position; a represents a control parameter; t is t _max Representing the maximum number of iterations;

(36) The maximum iteration condition is reached, and the optimal solution is the optimal parameters { C, alpha, sigma, mu, lambda }.

Images