Wind power data recovery method based on change point grouping-two-step quantile combined with DCNN
By combining variable point grouping method and two-step quantile algorithm with DCNN, abnormal data is identified and cleaned and missing data is reconstructed, the problem of abnormal data affecting wind power prediction in wind farms is solved, and data quality and recovery accuracy are improved.
Patent Information
- Application Number
- CN202210391251.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-14
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-04-14
AI Technical Summary
Abnormal data generated during the operation of the wind farm due to factors such as unit shutdown, load reduction, communication noise and equipment failure affect the accuracy of wind power prediction. The existing data processing methods are insufficient and the data quality cannot be accurately identified, resulting in inadequate data quality.
The method based on variable point grouping-two-step quantiles and deep convolutional neural network (DCNN) is used to clean and recover wind power data, and abnormal data is identified and deleted through variable point grouping method and bi-step quantile algorithm. DCNN is used to reconstruct missing data.
It significantly improves the data cleaning effect, ensures the quality of wind power output data, enhances the identification and processing ability of abnormal data, and improves the accuracy and reliability of wind power data recovery.
Smart Images

Figure CN114968991B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of wind farm electric meters, and in particular to a wind power data recovery method based on the combination of change point grouping-two-step quantile and DCNN. Background Art
[0002] As a new type of renewable clean energy, wind power energy is of great significance in the energy system of modern society. However, various uncontrolled output influencing factors in the wind power generation process make wind power have characteristics such as volatility, intermittency and randomness, which have an adverse impact on the stability and reliability of power system operation.
[0003] An important means to eliminate these adverse effects is to improve the predictability of wind power generation by cleaning and restoring the operating data output by the wind turbine meter. Therefore, accurately obtaining the wind speed and power data of the actual operation of the wind turbine can provide fundamental data support for the economic and safe operation and optimization control strategy of the wind farm. However, during the operation of the wind farm, a large amount of abnormal data will be generated due to factors such as unit shutdown, load reduction, communication noise and equipment failure. If these data are used directly without processing, it will affect the accuracy of wind power prediction. At present, there are still many deficiencies in the collection, management, analysis and mining methods of wind turbine operation data, and the quality difference of the collected data cannot be accurately identified, resulting in the lack of guarantee of data quality. Summary of the invention
[0004] In order to overcome the shortcomings of the prior art, the present invention provides a wind power output data recovery method based on the combination of change point grouping-two-step quantile algorithm and DCNN, which more accurately improves the data cleaning effect and ensures the quality of wind power output data. Wherein, DCNN is a deep convolutional neural network.
[0005] To achieve the above object, the technical solution adopted by the present invention is: a wind power data recovery method based on change point grouping-two-step quantile combined with DCNN, comprising:
[0006] Step S1: Analyze the factors affecting the controlled output of wind power, and classify the wind speed-power data forms according to different causes of abnormal data;
[0007] Step S2: collecting wind speed-power data samples in the wind speed interval of the wind turbine generator set, thereby forming a wind speed-power data sample set of the wind turbine generator set;
[0008] Step S3: Clean the obtained wind speed-power data samples using the change point grouping method to remove abnormal data accumulated at the bottom of the curve, abnormal data accumulated in the middle of the curve, and scattered abnormal data around some curves;
[0009] Step S4: using a two-step quantile algorithm to identify the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, and then deleting all the abnormal data to obtain a normal data set;
[0010] Step S5: Use a deep convolutional neural network (DCNN) to perform curve optimization, reconstruct the partially missing wind speed-power data points after cleaning, and obtain reconstructed wind speed-power data points without data missing.
[0011] Furthermore, the step S2: collecting wind speed-power data samples of a wind farm meter data to form a set, specifically includes:
[0012] Step S21: The data sample set within a certain wind speed interval is:
[0013] W={(v 1 ,p 1 ),(v 2 ,p 2 ),…,(v i ,p i ),…,(v n ,p n )}
[0014] In the formula, v i and p i They represent the wind speed and power of the ith data sample, respectively, and the data are arranged in descending order according to the power value, that is, let p i >p i+1 , i∈(1,n-1);
[0015] Step S22: Calculate the variance of each wind speed-power data point as S i :
[0016]
[0017] Among them, p j is the power of the jth data point, Represents the average value of the power from the 1st data point to the i-th data point;
[0018] Step S23: Calculate the variance change rate k(i):
[0019] k(i)=|S i -S i-1 |, i = 2, 3,…, n.
[0020] Furthermore, in step S3, the data is cleaned by using the change point grouping method to remove abnormal data of bottom accumulation type, abnormal data of middle accumulation type and abnormal data of scattered type around some curves, which specifically includes:
[0021] Step S31: Use the change point grouping method to clean the data, and when removing abnormal data, use the least square method to process the variance change rate k(i);
[0022] Let x q (i) is the variable i in the data sample order, with independent variable x 1 ,x 2 ,…,x r and a function of the dependent variable k, where q = 1,…, r, x q (i) is a completely known non-random function of i, k(i) is a random variable, which is divided into two sections, the front and back sections, each of which obeys a first-order linear model, and the regression coefficient mutates at i=j, that is:
[0023]
[0024] In the formula, β 1 (1) =[α (1) β 1 (1) … β r (1) ] T and β 2 (2) =[α (2) β 1 (2) … β r (2) ] T is the coefficient column vector, e is the natural logarithm, j is the regression change point, and since k(i) is continuous, there are constraints:
[0025] β 1 (1)T x(j)=β 2 (2)T x(j)
[0026] Where x(j) is a completely known non-random function at the regression change point;
[0027] Step S32: According to the principle of least squares method, the weighted objective function of the data sample model in step S21 is:
[0028]
[0029] In the formula, the weight of each item is i Inversely proportional to the error variance of sample k(i);
[0030] Step S33: Find the minimum value of the objective function under the constraint conditions to determine the estimated value of the change point j:
[0031]
[0032] After finding the change point j, the data after the change point becomes more and more discrete and is deleted as abnormal data. Therefore, the data is divided into normal data and abnormal data, namely:
[0033]
[0034] Where W n is the normal data set of wind speed range in the data sample; W 0 is the abnormal data set of the wind speed interval in the data sample, and the same processing is performed on other wind speed intervals, that is, the normal data set and abnormal data set of the wind speed-power curve of the entire wind speed interval of the wind turbine are obtained.
[0035] Furthermore, the step 4: using a two-step quantile algorithm to identify the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, and then deleting the abnormal data to obtain a normal data set, specifically includes:
[0036] Step S41: Arrange the wind speed-power data samples in the wind speed interval in ascending order of power and divide them into four equal parts. The three values P at the split point are 1 , P 2 and P 3 is the quantile, and the power ascending intermediate data set is recorded as X = {x 1 ,x 2 ,…,x n}, the quantile calculation method is as follows:
[0037] (1) Calculate the second quartile, i.e. the median P 2 for:
[0038]
[0039] (2) Calculate the first and third quartiles P 1 and P 3 :
[0040] When n=2k,k=0,1,2…,from P 2 Divide the data sample X into two parts but not including P 2 , calculate the median of these two parts respectively, and the two medians obtained are P 1 and P 3 ;
[0041] When n=4k+3,k=0,1,2…,there are:
[0042]
[0043] Among them, a and b are quantile parameters, with 0 <a<1,0<b<1,a+b=1;
[0044] Similarly, when n=4k+1, k=0, 1, 2…, we have:
[0045]
[0046] (3) Get the quantile range:
[0047] I PR =P 3 -P 1
[0048] Obtain inner limits for outliers in the sample data:
[0049] [F 1 ,F u ]=[P 1 -1.5I PR ,P 3 +1.5I PR ]
[0050] The data outside the inner limit in sample X are all the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, that is, the power abnormal data set is:
[0051] X 1 =[x i ,x i+1 ,…,x i+s |x <F 1 ∪x>F u ,0 <i<n,0<s<n,i,s∈R]
[0052] Step S42: Arrange the wind speed-power data samples in the wind speed interval in ascending order of wind speed and divide them into four equal parts. The three values Q at the dividing point are 1 , Q 2 and Q 3 The middle data set of wind speed in ascending order is recorded as Y = {y 1 ,y 2 ,…,y n}, the quantile calculation method is as follows:
[0053] (1) Calculate the second quartile, i.e. the median Q 2 for:
[0054]
[0055] (2) Calculate the first and third quartiles Q 1 and Q 3 .
[0056] When n=2k,k=0,1,2…,from Q 2 The data sample Y is divided into two parts (excluding Q 2 ), calculate the median of these two parts respectively, and the two medians obtained are Q 1 and Q 3 .
[0057] When n=4k+3,k=0,1,2…,there are:
[0058]
[0059] Among them, a and b are quantile parameters, with 0 <a<1,0<b<1,a+b=1。
[0060] Similarly, when n=4k+1, k=0, 1, 2…, we have:
[0061]
[0062] (3) Get the quantile range I QR :
[0063] I QR =Q 3 -Q 1
[0064] Obtain inner limits for outliers in the sample data:
[0065] [F 1 ′,F′ u ]=[Q 1 -1.5I QR ,Q 3 +1.5I QR ]
[0066] The data outside the inner limit in sample Y are all accumulation-type abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, that is, the wind speed abnormal data set is:
[0067] Y 1 =[y i ,y i+1 ,…,y i+s |y <F 1 ′∪y>F′ u ,0 <i<n,0<s<n,i,s∈R]
[0068] Delete the abnormal data to obtain the training sample data set.
[0069] V=W n -X 1 -Y 1 =[(v i -x i,p i -y i )|(v i -x i ,p i -y i )∈R,i∈R]
[0070] Among them, W n Normal data set for wind speed range, X 1 is the power anomaly data set, Y 1 Wind speed anomaly dataset;
[0071] Get the wind speed-power data of the wind turbine generator set's actual output V'=[(x′ i ,y′ i )|x′ i ,y′ i ∈R,i∈R] as the sample validation set for training, where x′ i The power data of wind speed-power data of the actual output of the wind turbine, y′ i The wind speed data is the wind speed-power data of the actual output of the wind turbine.
[0072] Furthermore, the step 5: using DCNN to perform data recovery, reconstructing the curve composed of wind speed-power data at some time points with missing data after cleaning, and obtaining the reconstructed wind speed-power data of the wind turbine at the time point without missing data is specifically implemented as follows;
[0073] Step S51: DCNN structure is as follows Figure 3 As shown, it consists of one input layer, three hidden layers and one output layer. First, the process of recovering the power data includes:
[0074] (1) The loss function is: Where T is the training period;
[0075] Initialize the minimum loss function, let Loss best The value is infinite, the training cycle counter t=1; go to step (2);
[0076] (2) Determine whether t is less than or equal to T: If so, set the patience value counter count to 1 and go to step (3); otherwise, go to step (8);
[0077] (3) Determine whether count is less than or equal to patience: If so, proceed to step (4); otherwise, Loss best =L(θ t ), go to step (7); where patience is a positive integer, which is the patience value used to monitor model training;
[0078] (4) Use the training sample set V to train the deep convolutional neural network. The network parameters obtained by training in the tth training cycle are θ t , the corresponding deep convolutional neural network Go to step (5);
[0079] (5) Use the validation sample set V' to Verify and get the optimized deep convolutional neural network for the tth training cycle Go to step (6);
[0080] (6) Calculation The loss function L(θ t ): If L(θ t ) is greater than Loss best , then count count is increased by one, and the process goes to step (3);
[0081] in, K represents the power data set x′ in the wind speed-power data of the actual output of the wind turbine i The total number of samples, k is the power data set x′ in the wind speed-power data of the actual output of the wind turbine i The sample index, X k The power data set x′ is the wind speed-power data of the actual output of the wind turbine i The kth sample in Y k is the kth label in the output label set, X k With Y k Correspondingly, 1≤t≤T;
[0082] (7) Increase t by 1 and return to step (2);
[0083] (8) Loss best The corresponding network parameters are used as the optimal network model parameters The corresponding optimized deep convolutional neural network is used as the optimal network model
[0084] (9) Recover the wind speed data according to steps (1) to (8).
[0085] Compared with the prior art, the present invention has the following advantages:
[0086] (1) The present invention uses a deep convolutional neural network (DCNN) method to restore wind power output data based on the change point grouping-two-step quantile algorithm for cleaning the historical wind speed and wind power data of the wind farm. The change point grouping-quartile method combines the change point grouping method with the two-step quantile algorithm, and uses two quartile methods to clean the wind speed and power data. Compared with the traditional quartile method, it is more accurate and improves the data cleaning effect;
[0087] (2) The change point grouping method can clean the bottom accumulation type abnormal data, the middle accumulation type abnormal data and some scattered abnormal data around the curve. The two-step quantile algorithm can clean the upper accumulation type abnormal data and the remaining scattered abnormal data around the curve. Therefore, the effective combination of the two can realize the reliable and effective cleaning of the four types of abnormal data, avoiding the problem that the traditional data cleaning method will fail when encountering a large amount of accumulation type abnormal data, and the cleaning effect is significant. BRIEF DESCRIPTION OF THE DRAWINGS
[0088] Figure 1 It is a flow chart of the wind power data recovery method based on the combination of change point grouping-two-step quantile algorithm and DCNN of the present invention;
[0089] Figure 2 is a wind speed-power scatter diagram in an embodiment of the present invention;
[0090] Figure 3 It is a DCNN structure diagram in an embodiment of the present invention;
[0091] Figure 4 The following is a flow chart of data recovery by DCNN in an embodiment of the present invention. DETAILED DESCRIPTION
[0092] The present invention provides a wind power data recovery method based on the combination of change point grouping-two-step quantile and DCNN, which does not rely on large data volume samples, improves the accuracy of wind turbine wind speed-power data recovery, and has a good recognition effect on various types of abnormal data. The data reconstruction effect is excellent, and provides an effective method for processing wind turbine wind speed-power data.
[0093] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below through specific implementations and in conjunction with the accompanying drawings.
[0094] like Figure 1 As shown, the wind power data recovery method based on change point grouping-two-step quantile combined with DCNN of the present invention includes the following steps:
[0095] Step S1: Analyze the factors affecting the controlled output of wind power and divide the data forms according to different causes of abnormal data;
[0096] Step S2: Collect wind speed-power data samples from wind farm electric meter data to form a set;
[0097] Step S3: Use the change point grouping method to clean the data, and remove the bottom accumulation type abnormal data, the middle accumulation type abnormal data, and the scattered abnormal data around some wind power curves;
[0098] Step S4: using a two-step quantile algorithm to identify the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, and then deleting the above abnormal data to obtain a normal data set;
[0099] Step S5: Use DCNN to perform curve optimization, reconstruct the wind speed-power data with missing data at some time points after cleaning, and obtain the reconstructed wind speed-power data of the wind turbine at the time points without missing data.
[0100] The above step S1: analyzing the factors affecting the controlled output of wind power, and dividing the data forms according to the different causes of abnormal data, specifically includes:
[0101] Step S11: The wind farm meter data has a large number of abnormal points that deviate from the normal working characteristics due to factors such as power curtailment, wind speed sensor failure, communication equipment failure, electromagnetic interference, wind turbine disconnection, extreme weather conditions, wind turbine blade contamination or damage, etc. The abnormal points of different causes are inconsistent in the data form shown in the wind speed-power scatter plot. According to the scatter point distribution form in the wind speed-power scatter plot, the abnormal data can be divided into four categories:
[0102] ① Abnormal data accumulation at the bottom of the curve. Abnormal data accumulation at the bottom of the curve is manifested as a horizontal data band in the wind power curve, with power fluctuating around zero, showing data accumulation. The causes of this type of data include unit failure, communication equipment failure or measurement terminal failure, unplanned wind turbine shutdown and maintenance, etc.
[0103] ② Abnormal data accumulated in the middle of the curve. Abnormal data accumulated in the middle of the curve is manifested in the wind power curve as one or more horizontal dense data bands outside the lower limit of the probability power curve, which will remain unchanged at a state lower than full power in a large wind speed range, showing data accumulation. The cause of this type of abnormal data is wind power curtailment or communication failure;
[0104] ③ Abnormal data accumulated on the upper part of the curve. Abnormal data accumulated on the upper part of the curve is manifested as one or more horizontal dense data bands located outside the upper limit of the probability curve. The power will remain unchanged during the low wind speed dispatch period, which is manifested as data accumulation. The cause of this type of abnormal data is communication error or wind speed sensor failure;
[0105] ④ Scattered abnormal data around the curve. Scattered abnormal data around the curve is manifested in the wind power curve as irregular scattered points with low density near the power curve, which are randomly scattered outside the boundary of the power curve. This type of abnormal data is caused by random factors such as signal propagation noise, sensor failure, and extreme weather conditions.
[0106] The above step S2: collecting wind speed-power data samples of wind farm electric meter data to form a set, specifically includes:
[0107] Step S21: record the data sample set within a certain wind speed interval as:
[0108] W={(v 1 ,p 1 ),(v 2 ,p 2 ),…,(v i ,p i ),…,(v n ,p n )}
[0109] In the formula, v i and p i They represent the wind speed and power of the ith data sample, respectively, and the data are arranged in descending order according to the power value, that is, let p i >p i+1 , i∈(1,n-1);
[0110] Step S22: Calculate the variance of each point as S i :
[0111]
[0112] Among them, p j is the power of the jth data point, Represents the average value of the power from the 1st data point to the i-th data point;
[0113] Step S23: Observe whether the variance change is obvious through the variance change rate k(i).
[0114] k(i)=|s i -s i-1 |,i=2,3,…,n
[0115] The above step S3: using the change point grouping method to clean the data, remove the abnormal data of accumulation type at the bottom of the curve, the abnormal data of accumulation type in the middle of the curve and the scattered abnormal data around some curves, specifically includes:
[0116] Step S31: When dealing with the change point problem, the least square method is used to identify the change point of the variance change rate k(i).
[0117] Let x q (i) is the variable i, with independent variable x 1 ,x 2 ,…,x r and a function of the dependent variable k, where q = 1,…, r. x q (i) is a completely known non-random function of i, k(i) is a random variable, which is divided into two sections, the front and back sections, each of which obeys a first-order linear model, and the regression coefficient mutates at i=j, that is:
[0118]
[0119] In the formula, the coefficient column vector β 1 =[α (1) β 1 (1) … β r (1) ] T and β 2 =[α (2) β 1 (2) … β r (2) ] T If they are not equal, e is the natural logarithm, then j is the regression change point. Since k(i) is continuous, there are constraints:
[0120] β 1 T x(j)=β 2 T x(j)
[0121] Step S32: According to the principle of least squares method, the weighted objective function of this data sample model is:
[0122]
[0123] In the formula, the weight of each item is i It is inversely proportional to the error variance of sample k(i).
[0124] Step S33: Find the minimum value of the objective function under the constraints to determine the estimate of the change point j:
[0125]
[0126] After finding the change point j, the data after the change point becomes more and more discrete and can be deleted as abnormal data. Therefore, the data can be divided into normal data and abnormal data, that is:
[0127]
[0128] Where W n is the normal data set for this wind speed range; W 0 is the abnormal data set of this wind speed interval. The same processing is performed on other wind speed intervals to obtain the normal data set and abnormal data set of the power curve of the wind turbine in the entire wind speed interval.
[0129] The above step S4: using a two-step quantile algorithm to identify the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, and then deleting the above abnormal data to obtain a normal data set, specifically includes:
[0130] Step S41: Arrange the wind speed-power data samples in the wind speed interval in ascending order of power and divide them into four equal parts. The three values P at the split point are 1 , P 2 and P 3 The intermediate data set in ascending power order is recorded as X = {x 1 ,x 2 ,…,x n}, the quantile calculation method is as follows:
[0131] (1) Calculate the second quartile, i.e. the median P 2 for:
[0132]
[0133] (2) Calculate the first and third quartiles P 1 and P 3 .
[0134] When n=2k,k=0,1,2…,from P 2 The data sample X is divided into two parts but does not include P 2 , calculate the median of these two parts respectively, and the two medians obtained are P 1 and P 3 ;
[0135] When n=4k+3,k=0,1,2…,there are:
[0136]
[0137] Among them, a and b are quantile parameters, with 0 <a<1,0<b<1,a+b=1;
[0138] Similarly, when n=4k+1, k=0, 1, 2…, we have:
[0139]
[0140] (3) Get the quantile range:
[0141] I PR =P 3 -P 1
[0142] Obtain inner limits for outliers in the sample data:
[0143] [F 1 ,Fu ]=[P 1 -1.5I PR ,P 3 +1.5I PR ]
[0144] The data outside the inner limit in sample X are all abnormal data, that is, the power abnormal data set is:
[0145] X 1 =[x i ,x i+1 ,…,x i+s |x <F 1 ∪x>F u ,0 <i<n,0<s<n,i,s∈R]
[0146] Step S42: Arrange the wind speed-power data samples in the wind speed interval in ascending order of wind speed and divide them into four equal parts. The three values Q at the dividing point are 1 , Q 2 and Q 3 The middle data set of wind speed in ascending order is recorded as Y = {y 1 ,y 2 ,…,y n}, the quantile calculation method is as follows:
[0147] (1) Calculate the second quartile, i.e. the median Q 2 for:
[0148]
[0149] (2) Calculate the first and third quartiles Q 1 and Q 3 .
[0150] When n=2k,k=0,1,2…,from Q 2 The data sample Y is divided into two parts (excluding Q 2 ), calculate the median of these two parts respectively, and the two medians obtained are Q 1 and Q 3 .
[0151] When n=4k+3,k=0,1,2…,there are:
[0152]
[0153] Among them, a and b are quantile parameters, with 0 <a<1,0<b<1,a+b=1。
[0154] Similarly, when n=4k+1, k=0, 1, 2…, we have:
[0155]
[0156] (3) Get the quantile range I QR :
[0157] I QR =Q 3 -Q 1
[0158] Obtain inner limits for outliers in the sample data:
[0159] [F 1 ′,F′ u ]=[Q 1 -1.5I QR ,Q 3 +1.5I QR ]
[0160] The data outside the inner limit in sample Y are all abnormal data, that is, the abnormal data set is:
[0161] Y 1 =[y i ,y i+1 ,…,y i+s |y <F 1 ′∪y>F′ u ,0 <i<n,0<s<n,i,s∈R]
[0162] The final cleaned data set is:
[0163] V=W n -X 1 -Y 1 =[(v i -x i ,p i -y i )|(v i -x i ,p i -y i )∈R,i∈R]
[0164] Use this data set as a sample data set for training;
[0165] The wind speed-power data V'=[(x′ i ,y′ i )|x′ i ,y′ i ∈R,i∈R] as the sample validation set for training, where x′ i The power data of wind speed-power data of the actual output of the wind turbine, y′ iThe wind speed data is the wind speed-power data of the actual output of the wind turbine.
[0166] The above step S5: using DCNN to perform data recovery, reconstructing the curve composed of wind speed-power data at some time points with missing data after cleaning, and obtaining the reconstructed wind speed-power data of the wind turbine at the time point without missing data, specifically includes:
[0167] Step S51: Figure 3 As shown in Figure 2, the process of data recovery using DCNN includes:
[0168] (1) Initialize the minimum loss function Loss best =∞, the training cycle is T, and the training cycle counter t=1; go to step (2); wherein the DCNN structure consists of an input layer, three hidden layers and an output layer; the loss function is: Where T is the training period;
[0169] (2) Determine whether t≤T holds: If so, set the patience value counter count=1 and go to step (3); otherwise, go to step (8);
[0170] (3) Determine whether count ≤ patience holds: If so, proceed to step (4); otherwise, Loss best =L(θ t ), go to step (7); where patience is a positive integer, which is the patience value used to monitor model training;
[0171] (4) Use the training sample data set V to train the deep convolutional neural network. The network parameters obtained by training in the tth training cycle are θ t , the corresponding deep convolutional neural network is recorded as Go to step (5);
[0172] (5) Use the training sample validation set V' to Verify and get the optimized deep convolutional neural network for the tth training cycle Go to step (6);
[0173] (6) Calculation The loss function L(θ t ): If L(θ t )>Loss best , then count = count + 1, go to step (3);
[0174] in, K represents the input sample set v' i -x' i The total number of samples, k is the sample index, Xk is the input sample set v' i -x' i The kth sample in Y k is the kth label in the output label set, X k With Y k Correspondingly, 1≤t≤T;
[0175] (7) t = t + 1, return to step (2);
[0176] (8) Loss best The corresponding network parameters are used as the optimal network model parameters The corresponding optimized deep convolutional neural network is used as the optimal network model
[0177] (9) Recover the wind speed data according to steps (1) to (8).
[0178] The present invention provides a wind power data recovery method based on the combination of change point grouping-two-step quantile and DCNN, which provides an effective method for processing wind turbine wind speed-power data without relying on large data volume samples, and reconstructs the data through DCNN to improve the accuracy of wind turbine wind speed-power data recovery, and has a good recognition effect on various types of abnormal data.
[0179] The above embodiments are provided only for the purpose of describing the present invention, and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the present invention should all be included within the scope of the present invention.
Claims
1. A wind power data recovery method based on change point grouping-two-step quantile combined with DCNN, characterized in that: include: Step S1: Analyze the factors affecting the controlled output of wind power, and classify the wind speed-power data forms according to different causes of abnormal data; Step S2: collecting wind speed-power data samples in the wind speed interval of the wind turbine generator set, thereby forming a wind speed-power data sample set of the wind turbine generator set; Step S3: using the change point grouping method to clean the obtained wind speed-power data samples, and remove the abnormal data of accumulation type at the bottom of the curve, the abnormal data of accumulation type in the middle of the curve, and some of the scattered abnormal data around the curve; Step S4: using a two-step quantile algorithm to identify the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, and then deleting all the abnormal data to obtain a normal data set; Step S5: using a deep convolutional neural network (DCNN) to perform curve optimization, reconstructing some of the missing wind speed-power data points after cleaning, and obtaining reconstructed wind speed-power data points without missing data; The step S2 specifically includes: Step S21: The data sample set within a certain wind speed interval is: , In the formula, and Respectively represent The wind speed and power of the data samples, and the data are arranged in descending order according to the power value, that is, , ; Step S22: Calculate the variance of each wind speed-power data point: : , in, is the power of the jth data point, Represents the average value of the power from the 1st data point to the i-th data point; Step S23: Calculate the variance change rate : ; In step S3, the bottom accumulation type abnormal data, the middle accumulation type abnormal data and the scattered type abnormal data around some curves are eliminated by using the change point grouping method, which specifically includes: Step S31: Use the change point grouping method to clean the data and remove abnormal data. Use the least square method to calculate the variance change rate. to process; set up The variable that is the order of the data samples is , with independent variables and the dependent variable The function of , yes A completely known non-random function of is a random variable, divided into two sections, each of which obeys a first-order linear model, and the regression coefficient is A mutation occurs at: , In the formula, and is the coefficient column vector, e is the natural logarithm, is the regression change point, is continuous, so there are constraints: , in, is a completely known non-random function at the regression change point; Step S32: According to the principle of least squares method, the weighted objective function of the data sample model in step S21 is: , In the formula, the weights of each item With sample The error variance is inversely proportional to Step S33: Find the minimum value of the objective function under the constraints to determine the change point Estimated value: , Find the change point After that, the data after the change point is deleted as abnormal data, so the data is divided into normal data and abnormal data, namely: , In the formula, is the normal data set of wind speed range in the data sample; The abnormal data set of the wind speed interval in the data sample is used, and the same processing is performed on other wind speed intervals, that is, the normal data set and abnormal data set of the wind speed-power curve of the entire wind speed interval of the wind turbine are obtained; The step S4: using a two-step quantile algorithm to identify the accumulated abnormal data on the upper part of the curve and the remaining scattered abnormal data around the curve, and then deleting the abnormal data to obtain a normal data set, specifically includes: Step S41: Arrange the wind speed-power data samples in the wind speed interval in ascending order of power and divide them into four equal parts. The three values at the split point are , and is the quantile, and the power ascending intermediate data set is recorded as , the quantile calculation method is as follows: (1) Calculate the second quartile, i.e. the median for: , (2) Calculate the first and third quartiles and : when When, from Data Sample Divided into two parts but not including , calculate the median of these two parts respectively, and the two medians obtained are and ; when When: , in is the quantile parameter, ; Similarly, When: , (3) Get the quantile range: , Obtain inner limits for outliers in the sample data: , The data outside the inner limit are all the accumulation type abnormal data on the upper part of the curve and the remaining scattered type abnormal data around the curve, that is, the power abnormal data set is: , Step S42: Arrange the wind speed-power data samples in the wind speed interval in ascending order of wind speed and divide them into four equal parts. The three values at the split point are , and is the quantile, and the middle data set of wind speed in ascending order is recorded as , the quantile calculation method is as follows: (1) Calculate the second quartile, i.e. the median for: , (2) Calculate the first and third quartiles and : when When, from Data Sample Divided into two parts does not include , calculate the median of these two parts respectively, and the two medians obtained are and ; when When: , in, is the quantile parameter, ; Similarly, When: , (3) Get the quantile range : , Obtain inner limits for outliers in the sample data: , sample The data outside the inner limit are all the accumulation type abnormal data on the upper part of the curve and the remaining scattered type abnormal data around the curve, that is, the wind speed abnormal data set is: , Delete the abnormal data to obtain the sample data set for training; , Where W n Normal data set of wind speed range, X1 is power abnormal data set, Y1 is wind speed abnormal data set; The wind speed-power data of the wind turbine's actual output is obtained through the wind outlet electric energy meter As the sample validation set for training, The power data is the wind speed-power data of the actual output of the wind turbine. Wind speed data which is the wind speed-power data of the actual output of the wind turbine; The step S5: using DCNN to perform data recovery, reconstructing the curve composed of wind speed-power data at some time points with missing data after cleaning, and obtaining the reconstructed wind speed-power data of the wind turbine at the time point without missing data is specifically implemented as follows; Step S51: The DCNN structure consists of an input layer, three hidden layers and an output layer. First, the process of recovering the power data includes: (1) The loss function is: , where T is the training period; Initialize the minimum loss function, let Loss best The value is infinite, the training cycle counter t=1; go to step (2); (2) Determine whether t is less than or equal to T: If so, set the patience value counter count to 1 and go to step (3); otherwise, go to step (8); (3) Determine whether count is less than or equal to patience: If so, proceed to step (4); otherwise, Loss best =L(θ t ), go to step (7); where patience is a positive integer, which is the patience value used to monitor model training; (4) Using the training sample data set Train the deep convolutional neural network, and the network parameters obtained by training in the tth training cycle are θ t , the corresponding deep convolutional neural network Go to step (5); (5) Use the training sample validation set right Verify and get the optimized deep convolutional neural network for the tth training cycle Go to step (6); (6) Calculation The loss function L(θ t ): If L(θ t ) is greater than Loss best , then the count count is increased by one and the process goes to step (3); in, , K represents the wind speed-power data set of the actual output of the wind turbine The total number of samples, k is the wind speed-power data set of the actual output of the wind turbine The sample index, X k The wind speed-power data set for the actual output of wind turbines The kth sample in Y k is the kth label in the output label set, X k With Y k Correspondingly, 1≤t≤T; (7) Increase t by 1 and return to step (2); (8) Loss best The corresponding network parameters are used as the optimal network model parameters , the corresponding optimized deep convolutional neural network is used as the optimal network model ; (9) Recover the wind speed data according to steps (1) to (8).
Citation Information
Patent Citations
Combined wind power plant data cleaning method
CN111291032A
Wind turbine generator abnormal operation data processing method
CN111522808A
Fan wind speed-power data combined processing method
CN113821502A
Wind power plant group model construction method
CN114036845A
Method of identification and compensation of inherent deviation of yaw error of wind turbine based on true power curve
US20210262439A1
