CN116960960B - Short-term wind power prediction method for wind turbine - Google Patents

Abstract

The invention discloses a method for predicting short-term wind power of a wind turbine, which comprises the following steps: acquiring historical wind power data, processing the historical wind power data, inputting the processed data into a CNN network for feature extraction, and extracting feature data to train CNN-LSTM and CNN-GRU sub-networks; calculating forgetting factors of CNN-LSTM and CNN-GRU sub-networks; inputting current real-time wind power data into trained CNN-LSTM and CNN-GRU sub-networks to predict, obtaining a predicted value, calculating forgetting factors of the current stage through the trained BILSTM network, and correcting the predicted value by using the forgetting factors of the current stage; and carrying out weighted fusion on the sub-network by using a CRITIC self-adaptive weight evaluation system to obtain a final predicted value. The invention improves the prediction precision and the anti-interference capability of the network.

Description

Short-term wind power prediction method for wind turbine

Technical Field

The invention relates to the technical field of wind power generation prediction, in particular to a short-term wind power prediction method for a wind turbine.

Background

The rapid development of wind power generation relieves the problems of energy shortage, environmental pollution and the like, but simultaneously brings new challenges to related wind power enterprises and power systems. Because wind energy has the characteristics of intermittence, randomness and volatility, wind energy cannot be effectively controlled like traditional energy control, wind power generation power cannot be accurately predicted like electric load, reliable operation and stability of an electric power system are seriously affected, and development of wind power generation is limited. Therefore, the method for forecasting the short-term wind power of the wind power generator is researched and designed, so that activities such as timely making a scheduling plan, timely adjusting the standby capacity of the power system, keeping the power supply and the demand balance, reducing the running cost of the power system and the like are of great practical significance.

At present, short-term wind power prediction models can be divided into four types, including a physical model, a statistical model, a probability model and an artificial intelligence model. Wherein, the artificial intelligent model refers to wind power prediction from the deep learning perspective, i.e. learning a complex mapping between input and output from a large amount of historical data. Compared with other models, the artificial intelligent model can quickly capture complex nonlinearity and uncertainty of the wind power time sequence, has strong flexibility, can be suitable for various different scenes, has high prediction precision, and particularly can better track mutation data. In recent years, various artificial intelligence models have emerged, such as Multi-layer Perceptron (MLP), adaptive general regression neural networks (Adaptive General Recurrent Neural Networks, AGRNN), fast training neural networks (Fast Training Neural Networks, FTNN), recurrent neural networks (Recurrent Neural Networks, RNN), and the like. However, most of the existing wind power prediction methods based on deep learning use a single neural network, and the architecture is simple, but in the input of the network, only historical wind power is often considered, related influence factors are not considered, the predicted numerical value is directly used as a final estimated value of the network, the forgetting degree and deviation in the network prediction process are ignored, and finally, larger errors are often caused in the prediction method.

Disclosure of Invention

The invention aims to solve the problems that when the existing network is used for wind power prediction, the network is single in structure, and in the input of the network, only historical wind power is often considered, and related influence factors are not considered, so that the network prediction precision is low. Therefore, the invention provides a short-term wind power prediction method for a wind turbine, which better provides a reliable scheme for power dispatching of a wind power plant.

In order to achieve the above purpose, the invention provides a method for predicting short-term wind power of a wind turbine, which comprises the following steps:

acquiring historical wind power data, processing the historical wind power data, inputting the processed data into a CNN network for feature extraction, and extracting feature data to train CNN-LSTM and CNN-GRU sub-networks;

calculating forgetting factors of the CNN-LSTM and CNN-GRU sub-networks, training a BILSTM network based on the forgetting factors, and obtaining a trained BILSTM network;

inputting current real-time wind power data into the trained CNN-LSTM and CNN-GRU sub-networks to predict, obtaining a predicted value, calculating a forgetting factor of a current stage through the trained BILSTM network, and correcting the predicted value by using the forgetting factor of the current stage;

and carrying out weighted fusion on the sub-network by using a CRITIC self-adaptive weight evaluation system to obtain a final predicted value.

Preferably, the historical wind power data is:

x′(i，：)＝[p，v，d，f，t，h]；

wherein p is wind power, v is ambient wind speed, d is wind direction, f is air pressure, t is temperature, h is air humidity, and x' (i: is) is historical wind power data.

Preferably, the processing the historical wind power data includes:

performing sliding window extraction and normalization processing on the historical wind power data, designing a sliding window, determining the size and the step length of the sliding window, extracting data according to a preset rule, and obtaining a feature set X _t And target set Y _t And for the feature set X _t And target set Y _t And (5) carrying out normalization processing.

Preferably, the normalization method comprises the following steps:

wherein X 'is' _t Representing feature set X _t Transposed matrix of X ', X' _t (: j) represents X' _t In (c) column components, max (X' _t (: j)) and min (X' _t (: j)) are X 'respectively' _t Maximum and minimum values in (: j);

wherein Y' _t Representing the target set Y _t Transposed matrix of Y' _t (i) represents Y' _t In (c) row component, max (Y' _t (i,:)) and min (Y' _t (i,:)) are Y 'respectively' _t Maximum and minimum values in (i, i).

Preferably, the training CNN-LSTM and CNN-GRU subnetworks, comprising:

and capturing the dependency relationship in the data generated by the CNN network by utilizing the gating units in the CNN-LSTM and CNN-GRU sub-networks, obtaining two groups of wind power predicted values based on parallel processing of the CNN-LSTM and CNN-GRU sub-networks, and performing inverse normalization processing on the wind power predicted values to obtain the trained CNN-LSTM and CNN-GRU sub-networks.

Preferably, the method for performing the inverse normalization process is as follows:

wherein pt (i:) is the network output predicted value after inverse normalization, ptnorm (i:) is the network output predicted value before inverse normalization, max (Y' _t (i,:)) and min (Y' _t (i,:)) are Y 'respectively' _t Maximum and minimum values in (i, i).

7. The method for short-term wind power prediction for a wind turbine according to claim 1, wherein the method for calculating forgetting factors of the CNN-LSTM and CNN-GRU sub-networks is as follows:

wherein F (i, i) is a time-varying forgetting factor array, and ya (i, i) is an actual measured value of wind power.

Preferably, the method for correcting the predicted value by using the forgetting factor of the current stage is as follows:

wherein pyfup is a network predicted value corrected by the forgetting factor after time lapse, py is a network predicted value corrected by the forgetting factor after time lapse, fup is F more than or equal to 0, and Fdown is F less than 0.

Preferably, the obtaining the final predicted value includes:

and introducing a CRITIC self-adaptive weight evaluation system, respectively giving different weights to the sub-networks according to the distribution condition of the sub-network prediction data, and carrying out weighted fusion on the sub-networks to obtain a final wind power prediction value.

Preferably, the CRITIC adaptive weight evaluation system is:

C _j ＝S _j ×A _j

wherein Y represents the data set of each subnet predicted value, Y _mn Representing a single element of the mth row and nth column in the dataset Y _n Representing all data of the nth column in data set Y, Y _ij Elements representing the j-th column of the ith row of the normalized data set Y, S _j Represents the data volatility coefficient, m represents the number of rows of the data set Y, n represents the number of columns of the data set Y,representing the mean value of the j-th column data of the normalized data set Y, Y _ik Elements representing the ith row and the kth column of the normalized dataset Y, +.>Represents the mean value of the k column data of the data set Y after normalization, R _ij Representing a matrix of correlation coefficients, A _j Representing the coefficient of conflict, C _j Representing information quantity coefficient, W _j Representing the calculated weights.

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

(1) According to the invention, the CNN network is utilized to perform feature extraction on wind power data, so that the defect that the traditional single-network wind power data feature extraction is incomplete is overcome, and the comprehensive wind power data feature improves the prediction precision of the network; the forgetting property of the network in the wind power prediction process is considered, the time-varying forgetting factor is introduced, the forgetting degree of the network is measured, and the network predicted value is subjected to forgetting correction by using the forgetting degree. In addition, the subnetworks are fused based on the CRITIC weight evaluation system, namely, the prediction results of the subnetworks are weighted and summed to obtain a final prediction output value.

(2) The invention further improves the prediction precision of the network, can track the data with larger fluctuation, has high prediction precision, and can provide more accurate data support for the power personnel to formulate the relevant power dispatching plan.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application, illustrate and explain the application and are not to be construed as limiting the application. In the drawings:

FIG. 1 is a diagram of an LSTM network in accordance with an embodiment of the present invention;

FIG. 2 is a block diagram of a GRU network in accordance with an embodiment of the invention;

FIG. 3 is a diagram of a CNN-LSTM network in accordance with an embodiment of the present invention;

FIG. 4 is a diagram of a CNN-GRU network in accordance with an embodiment of the present invention;

FIG. 5 is a diagram of a network architecture for multi-network convergence in an embodiment of the invention;

FIG. 6 is a flowchart of a network algorithm implementation of multi-network convergence in an embodiment of the invention;

FIG. 7 is a schematic diagram of a sliding window data extraction process according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of a time-varying forgetting factor generation process according to an embodiment of the present invention;

fig. 9 is a workflow diagram of CRITIC adaptive weight evaluation architecture in an embodiment of the invention.

Detailed Description

It should be noted that, in the case of no conflict, the embodiments and features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings in conjunction with embodiments.

It should be noted that the steps illustrated in the flowcharts of the figures may be performed in a computer system such as a set of computer executable instructions, and that although a logical order is illustrated in the flowcharts, in some cases the steps illustrated or described may be performed in an order other than that illustrated herein.

The invention provides a method for predicting short-term wind power of a wind turbine, as shown in FIG. 6, comprising the following steps:

after carrying out sliding window extraction and normalization processing on the historical wind power data, inputting the historical wind power data into a CNN network for feature extraction; training a gating cycle unit (GRU, gated Recurrent Unit) and a Long Short-Term Memory network (LSTM) by using the extracted characteristic data; calculating forgetting factors of each sub-network, and training a Bi-directional Long-Short Term Memory (BILSTM) network; sending real-time data of the unit into a trained sub-network to predict wind power; calculating a forgetting factor of the current stage by the BILSTM network, and correcting a predicted value of the sub-network by using the forgetting factor; and (5) carrying out sub-network superposition by using a self-adaptive weight evaluation system (CRITIC, criteria Importance Through Intercriteria Correlation) weight system, thereby obtaining a final predicted value.

The method specifically comprises the following steps:

step 1: and acquiring historical related data required by wind power prediction.

The method comprises the steps of obtaining power information of a wind motor within a period of time by using a sensor, and recording wind speed, wind direction, air pressure, temperature and air humidity data x ', x' = [ x (i, j) ], i epsilon (1, l), j epsilon (1, 6) at the same moment by using the sensor; the once sampled output of the wind motor is denoted by x '(i, the expression of x' (i, i):

x′(i，：)＝[p，v，d，f，t，h]； (1)

wherein p is wind power, v is ambient wind speed, d is wind direction, f is air pressure, t is temperature, and h is air humidity.

Step 2: wind power data for training the subnetwork is selected and set.

A sliding window is designed, the sliding window size is set to 7×1, and the sliding step size is set to 1 (see fig. 7). Carrying out sliding window extraction on the data in the step 2, extracting a group of data at every 7 moments in the embodiment, and independently extracting the data at the eighth moment;

obtaining a feature set X _t And target set Y _t ，X _t The expression of (2) is:

X _t ＝[X′ _t1 ，Y′ _t2 ，Y′ _t3 ，Y′ _t4 ，Y′ _t5 ，Y′ _t6 ] (2)

wherein X 'is' _t1 ，Y′ _t2 ，Y′ _t3 ，Y′ _t4 ，Y′ _t5 ，Y′ _t6 Respectively the extracted wind power, wind speed, wind direction, air pressure, air temperature and air humidity data.

In order to ensure consistency and comparability among different characteristic data, the invention carries out normalization processing on the obtained characteristic set and the target set, and the specific processing mode is as follows:

wherein X 'is' _t Representing feature set X _t Transposed matrix of X ', X' _t (: j) represents X' _t In (c) column components, max (X' _t (: j)) and min (X' _t (: j)) are X 'respectively' _t Maximum and minimum values in (: j).

Wherein Y' _t Representing the target set Y _t Transposed matrix of Y' _t (i) represents Y' _t In (c) row component, max (Y' _t (i,:)) and min (Y' _t (i,:)) are Y 'respectively' _t Maximum and minimum values in (i, i).

Step 3: training CNN-LSTM and CNN-GRU subnetworks.

The data obtained in the step 2 is used for training the CNN-LSTM (shown in figure 3) and CNN-GRU (shown in figure 4) networks. The normalized data is input into the CNN layer for convolution, pooling and full connection processing, and depth features of the data are extracted and then transferred to the subnet layer. The subnet layer consists of LSTM (fig. 1) and GRU (fig. 2) that run in parallel and process data from the CNN layer. By utilizing gating units, LSTM and GRU can effectively capture long-term dependencies that exist in the data generated by the CNN layer. And obtaining two groups of wind power predicted values through parallel processing of the sub-network layers. In order to recover the dimension of the predicted wind power, the output value of the sub-network layer is inversely normalized, see formula (5).

Wherein pt (i:) is the network output predicted value after inverse normalization, ptnorm (i:) is the network output predicted value before inverse normalization, max (Y' _t (i,:)) and min (Y' _t (i,:)) are Y 'respectively' _t Maximum and minimum values in (i, i).

Step 4: and introducing and calculating a time-varying forgetting factor, and training the BILSTM network by using the time-varying forgetting factor. The time-varying forgetting factor is a concept introduced by the invention, and is used for measuring the forgetting degree of the network and correcting the forgetting degree of the network.

And (3) solving a time-varying forgetting factor of each sub-network in the sub-network layer by using a formula (6), so as to measure the forgetting degree of the network at each moment.

The obtained time-varying forgetting factors are input into a BILSTM network and are trained.

Wherein F (i, i) is a time-varying forgetting factor array, and ya (i, i) is an actual measured value of wind power.

The memory of human can be divided into two types according to whether redundant information exists or not and whether target information exists or not, and the method adopts a similar classification method to the time-varying forgetting factor in order to better express and correct the forgetting degree of the sub-network.

Specifically, the time-varying forgetting factors fall into two categories: a time-varying up forgetting factor and a time-varying down forgetting factor (equation 7). When the time-varying forgetting factor is positive, the wind power predicted value of the power grid is lower than the actual value; and when the forgetting factor is negative in time-varying, the wind power predicted value of the power grid is larger than the actual value.

Step 5: and utilizing the time-varying forgetting factor of the previous period and generating the time-varying forgetting factor of the current period by means of the BILSTM network, and carrying out forgetting correction on the predicted value of the sub-network.

Inputting current relevant wind power data into a network, repeating the step 2 to obtain data for wind power prediction, loading CNN-GRU and CNN-LSTM sub-networks trained in the step 3, performing parallel prediction of wind power by using the CNN-GRU and CNN-LSTM sub-networks, loading BILSTM networks trained in the step 4, generating a time-varying forgetting factor of the current period by using a time-varying forgetting factor of the previous period by using the BILSTM network (as shown in figure 8), and performing forgetting correction on a predicted value of the sub-network by using the time-varying forgetting factor according to the formula (8) (as shown in figure 5).

Wherein pyfup is a network predicted value corrected by forgetting over time, py is a network predicted value corrected by forgetting factor over time, fup is F.gtoreq.0, and Fdown is F.gtoreq.0.

Step 6: and (3) introducing a CRITIC self-adaptive weight evaluation system, endowing each sub-network with different weights according to the data characteristics of each sub-network by the CRITIC weight system, and carrying out weighted fusion on the sub-networks to obtain a final wind power predicted value.

And (3) introducing a CRITIC self-adaptive weight evaluation system (formula 9 and formula 10), respectively giving different weights to the subnetworks according to the distribution condition of the subnetwork prediction data, and then carrying out weighted fusion on the subnetworks (as shown in figure 9) to obtain the final wind power prediction value of the network.

The CRITIC weight evaluation system is an objective assignment method based on data volatility, and the idea is that the volatility (contrast strength) and the conflict (correlation) are two indexes. CRITIC weight systems are introduced primarily to optimize overall prediction accuracy.

Cj＝Sj×Aj

Wherein Y represents the data set of each subnet predicted value, Y _mn Representing a single element of the mth row and nth column in the dataset Y _n Representing all data of the nth column in data set Y, Y _ij Elements representing the j-th column of the ith row of the normalized data set Y, S _j Represents the data volatility coefficient, m represents the number of rows of the data set Y, n represents the number of columns of the data set Y,representing the mean value of the j-th column data of the normalized data set Y, Y _ik Elements representing the ith row and the kth column of the normalized dataset Y, +.>Represents the mean value of the k column data of the data set Y after normalization, R _ij Representing a matrix of correlation coefficients, A _j Representing the coefficient of conflict, C _j Representing information quantity coefficient, W _j Representing the calculated weights.

The network designed by the invention has strong robustness, can track data with larger volatility, has higher prediction precision compared with the traditional single network, and can better provide a reliable scheme for power dispatching of a wind power plant.

In order to overcome the defects of low reliability, low prediction precision and poor robustness of a single network prediction result, a plurality of networks are fused, and the prediction precision and the anti-interference capability of the networks are improved. Firstly, the CNN network is utilized to conduct feature extraction on wind power data, the defect that the traditional single-network wind power data feature extraction is incomplete is overcome, and the comprehensive wind power data feature improves the prediction accuracy of the network. According to the method, the forgetting property of the network in the wind power prediction process is considered, the time-varying forgetting factor is introduced, the forgetting degree of the network is measured, and the network predicted value is corrected by using the forgetting degree. In addition, the subnetworks are fused based on the CRITIC weight evaluation system, namely, the prediction results of the subnetworks are weighted and summed to obtain a final prediction output value. The above summary further improves the accuracy of the network prediction. The invention can track data with larger fluctuation, has high prediction precision, and can provide more accurate data support for electric staff to formulate related power dispatching plans.

The foregoing is merely a preferred embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions easily conceivable by those skilled in the art within the technical scope of the present application should be covered in the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims (8)

1. A method for short-term wind power prediction for a wind turbine, comprising:

acquiring historical wind power data, processing the historical wind power data, inputting the processed data into a CNN network for feature extraction, and extracting feature data to train CNN-LSTM and CNN-GRU sub-networks;

calculating forgetting factors of the CNN-LSTM and CNN-GRU sub-networks, training a BILSTM network based on the forgetting factors, and obtaining a trained BILSTM network;

inputting current real-time wind power data into the trained CNN-LSTM and CNN-GRU sub-networks to predict, obtaining a predicted value, calculating a forgetting factor of a current stage through the trained BILSTM network, and correcting the predicted value by using the forgetting factor of the current stage;

the method for calculating the forgetting factors of the CNN-LSTM and CNN-GRU subnetworks comprises the following steps:

f (i) is a time-varying forgetting factor array, ya (i) is an actual measured value of wind power, and pt (i) is an inverse normalized network output predicted value;

the method for correcting the predicted value by using the forgetting factor of the current stage comprises the following steps:

wherein pyfup is a network predicted value corrected by the forgetting factor after time lapse, py is a network predicted value corrected by the forgetting factor after time lapse, fup is F more than or equal to 0, and Fdown is F less than 0;

and carrying out weighted fusion on the sub-network by using a CRITIC self-adaptive weight evaluation system to obtain a final predicted value.

2. The method for short-term wind power prediction for a wind turbine of claim 1, wherein the historical wind power data is:

x′(i，：)＝[p，v，d，f，t，h]；

wherein p is wind power, v is ambient wind speed, d is wind direction, f is air pressure, t is temperature, h is air humidity, and x' (i: is) is historical wind power data.

3. The method for short-term wind power prediction for a wind turbine of claim 1, wherein processing the historical wind power data comprises:

performing the historical wind power dataSliding window extraction and normalization processing, designing a sliding window, determining the size and step length of the sliding window, and extracting data according to a preset rule to obtain a feature set X _t And target set Y _t And for the feature set X _t And target set Y _t And (5) carrying out normalization processing.

4. The method for short-term wind power prediction for a wind turbine according to claim 3, wherein the method for performing normalization processing is as follows:

wherein X 'is' _t Representing feature set X _t Transposed matrix of X ', X' _t (: j) represents X' _t In (c) column components, max (X' _t (: j)) and min (X' _t (: j)) are X 'respectively' _t Maximum and minimum values in (: j);

wherein Y' _t Representing the target set Y _t Transposed matrix of Y' _t (i) represents Y' _t In (c) row component, max (Y' _t (i,:)) and min (Y' _t (i,:)) are Y 'respectively' _t Maximum and minimum values in (i, i).

5. The method for short-term wind power prediction for wind turbines according to claim 1, wherein training the CNN-LSTM and CNN-GRU sub-networks comprises:

and capturing the dependency relationship in the data generated by the CNN network by utilizing the gating units in the CNN-LSTM and CNN-GRU sub-networks, obtaining two groups of wind power predicted values based on parallel processing of the CNN-LSTM and CNN-GRU sub-networks, and performing inverse normalization processing on the wind power predicted values to obtain the trained CNN-LSTM and CNN-GRU sub-networks.

6. The method for short-term wind power prediction for a wind turbine according to claim 5, wherein the method for performing inverse normalization processing is as follows:

wherein pt (i:) is the network output predicted value after inverse normalization, ptnorm (i:) is the network output predicted value before inverse normalization, max (Y' _t (i,:)) and min (Y' _t (i,:)) are Y 'respectively' _t Maximum and minimum values in (i, i).

7. The method for short-term wind power prediction for a wind turbine according to claim 1, wherein the obtaining a final predicted value includes:

and introducing a CRITIC self-adaptive weight evaluation system, respectively giving different weights to the sub-networks according to the distribution condition of the sub-network prediction data, and carrying out weighted fusion on the sub-networks to obtain a final wind power prediction value.

8. The method for short-term wind power prediction for a wind turbine according to claim 7, wherein the CRITIC adaptive weight evaluation system is:

C _j ＝S _j ×A _j

wherein Y represents the data set of each subnet predicted value, Y _mn Representing a single element of the mth row and nth column in the dataset Y _n Representing all data of the nth column in data set Y, Y _ij Elements representing the j-th column of the ith row of the normalized data set Y, S _j Represents the data volatility coefficient, m represents the number of rows of the data set Y, n represents the number of columns of the data set Y,representing the mean value of the j-th column data of the normalized data set Y, Y _ik Elements representing the ith row and the kth column of the normalized dataset Y, +.>Represents the mean value of the k column data of the data set Y after normalization, R _ij Representing a matrix of correlation coefficients, A _j Representing the coefficient of conflict, C _j Representing information quantity coefficient, W _j Representing the calculated weights.