A new outlier correction algorithm for wind power forecasting

Abstract

The present invention relates to the technical field of data-driven industrial soft-sensing modeling, and provides a novel outlier-corrected wind power prediction algorithm, comprising the following steps: step 1, collecting wind speed historical data and wind power historical data as original data sets, and preprocessing obvious outliers in the wind speed sequence; step 2, using a new outlier detection algorithm to filter out outliers and then remove them, and then using a new outlier correction method to reconstruct missing values; step 3, using variational mode decomposition of wind speed and wind power sequences; step 4, constructing a multi-output gated cyclic unit model to synchronously predict wind speed and wind power, and at the same time proposing a new loss function; step 5, developing a fine-tuning correction module for constraining and correcting wind power prediction values. This scheme realizes the rapid and accurate estimation of wind power, thereby contributing to the optimal scheduling of power systems and reducing operating costs.

Description

Technical Field

[0001] The invention relates to the technical field of data-driven industrial soft-sensing modeling, and in particular to a novel outlier-corrected wind power prediction algorithm. Background Art

[0002] With the continuous growth of energy demand and the promotion of achieving carbon peak and carbon neutrality goals, wind energy as a renewable clean energy has attracted great attention. However, due to the randomness and uncertainty of wind energy, it is difficult to accurately estimate wind power; after large-scale wind power is connected to the power grid, it will cause a huge impact on the power grid, seriously affecting the safe and stable operation of the power grid, and also bringing severe challenges to the dispatching work of the power department. Therefore, it is of great theoretical value and practical significance to carry out research on high-precision and high-reliability soft measurement prediction methods for wind power.

[0003] However, the development of soft-sensing prediction methods for wind power has the following problems: (1) Due to the influence of wind turbulence and sensor failure, there are a large number of outliers in the data set; (2) The values ​​of historical process variables have complex randomness and non-stationary characteristics, making it difficult to accurately analyze their changes; (3) There are strong nonlinearities and time lags between the target variable and the historical process variables; (4) Soft-sensing algorithms that rely solely on data-driven models lack the constraints and corrections of domain knowledge, which has certain limitations on the accuracy of prediction results; (5) Soft-sensing algorithms sometimes produce prediction results that do not conform to physical laws. Therefore, it is urgent to develop a soft-sensing algorithm to detect and process outliers and combine relevant domain knowledge with time series neural networks. Summary of the invention

[0004] In order to solve the problems of outliers in measurement data, difficulty in capturing process change trends, lack of domain knowledge in soft measurement models, and prediction results that violate physical laws, the present invention proposes a new outlier correction wind power prediction algorithm to achieve rapid and accurate estimation of wind power, thereby contributing to the optimal scheduling of power systems and reducing operating costs.

[0005] To achieve the above object, the present invention adopts the following scheme: a new outlier-corrected wind power prediction algorithm, comprising the following steps:

[0006] Step 1: collect historical wind speed data and historical wind power data as the original data set, perform preliminary outlier detection and elimination on the historical wind speed data based on the box plot method, and then use the cubic spline interpolation method to complete the eliminated historical wind speed data;

[0007] Step 2: For the original data set that has been preliminarily processed in step 1, a new outlier detection algorithm is used to screen out outliers and then remove them, and then an outlier correction method is used to reconstruct missing values. The outlier detection algorithm is based on the K-means clustering algorithm and the Gaussian kernel density estimation method, and the new outlier correction method is based on the window smoothing method and theoretical power. The specific execution process of the deep outlier detection and elimination and missing value completion correction includes:

[0008] Step 2.1, using the K-means clustering algorithm to cluster data points with similar wind power values ​​into one category, and divide the sample points into multiple category clusters;

[0009] Step 2.2: In each category cluster, the Gaussian kernel density estimation method is used to analyze the wind speed value corresponding to each sample point to obtain the kernel density estimation value of each sample point, and the inverse of the kernel density estimation value is recorded as the outlier score;

[0010] Step 2.3, set the judgment threshold, and the sample outlier score exceeding the threshold is judged as an outlier;

[0011] Step 2.4, after removing the detected abnormal values, fill in the missing values ​​in two steps: first, use the window smoothing method to calculate the mean of the two samples before and after the missing value in the wind speed series and wind power series; second, according to the corresponding mathematical expression between wind speed and wind power, calculate the theoretical wind power value to correct the corresponding wind power value. The mathematical expression between wind speed and wind power is:

[0012] (1)

[0013] in, Indicates the wind speed at the hub; Indicates the cut-in wind speed; Indicates rated wind speed; Indicates the cut-out wind speed; Indicates rated power; Indicates the wind power generated under different wind speed conditions; represents the wind energy utilization coefficient; is the wind swept area of ​​the fan blades; is the air density;

[0014] Step 3: for the original data set after deep processing in step 2, the wind speed and wind power sequences are decomposed by variational mode, and the neural network model is used to mine the information in the wind speed historical data and the wind power historical data;

[0015] Step 4: Take wind power as the main prediction variable and wind speed as the auxiliary prediction variable to build a multi-output gated recurrent unit model to simultaneously predict wind speed and wind power; at the same time, a new loss function is proposed to enhance the model's attention to the wind power prediction task during training. The new loss function is:

[0016] (11)

[0017] in, is the adjustment factor, whose value range is (1, 2), used to adjust the loss ratio; is the loss value of wind power; is the loss value of wind speed; is the loss value of the forward propagation of the MGRU model;

[0018] Step 5: Develop a fine-tuning correction module based on the domain knowledge of the wind power generation process to constrain and correct the wind power forecast value. The specific execution process of the fine-tuning correction module includes:

[0019] Step 5.1, calculate the corresponding theoretical wind power value according to the wind speed forecast value, perform weighted recombination on the theoretical wind power value and the initially corrected wind power forecast value, and constrain and correct the wind power forecast value that deviates from the true value. The weighted recombination formula is:

[0020] (12)

[0021] in, is the weight coefficient, whose value range is (0, 1), which is used to adjust the value ratio; is the wind power value after constraint correction; is the theoretical wind power value; is the wind power forecast value after preliminary correction;

[0022] Step 5.2, correct the negative wind power forecast value to 0;

[0023] Step 5.3, fine-tune the power value based on the characteristics and mechanism knowledge of wind power curve: for the power value corresponding to the wind speed prediction value less than the credible range of the cut-in wind speed, set it to 0; for the power value corresponding to the wind speed exceeding the credible range of the rated wind speed, set it to the rated power value.

[0024] Specifically, in step 1, the specific execution process of preliminary detection and elimination of obvious abnormal values ​​in the wind speed sequence includes:

[0025] Step 1.1, calculate the lower quartile Q1, median Q2, upper quartile Q3 and interquartile range IQR of the wind speed data, where Q1 is the 25% quantile of the data, Q2 is the median value of the data, Q3 is the 75% quantile of the data, and IQR is the range between Q1 and Q3, and its value is Q3-Q1;

[0026] Step 1.2, calculate the upper and lower limits of the box plot. Points outside the upper and lower limits are considered outliers and are removed. The upper limit of the box plot is Q1-1.5*IQR, and the lower limit is Q3+1.5*IQR.

[0027] Specifically, in step 2, the data set includes 3,000 samples, 70% of the data set is used as training data, and the remaining 30% is used as test data.

[0028] Specifically, in step 3, the multi-output gated loop unit decomposes the wind speed and wind power sequences into 12 intrinsic mode function components respectively, and the specific execution process includes:

[0029] Step 3.1, initialize the number of decomposed modes , the optimization problem is constructed by the variational principle, the goal is to minimize the reconstruction error and modal bandwidth, and the specific form of the constrained variational expression is:

[0030] (2)

[0031] in, Indicates The center frequency of the sub-signal, is the Dirac distribution, Indicates about time The partial derivative of Indicates IMF components; represents the original signal; represents a complex exponential function;

[0032] Step 3.2, by introducing the Lagrange multiplier and the quadratic penalty factor , integrate the constraints into the optimization problem and form the Lagrangian function:

[0033]

[0034] Step 3.3, iteratively update each mode by alternating direction multiplier method , and The optimal solution of , the iterative update expression is:

[0035]

[0036] in, is the number of iterations, , , and They are , , and Fourier transform of

[0037] Step 3.4, By taking the inverse Fourier transform, we can get the expression of each component in the time domain, and thus get the decomposed sub-signals.

[0038] Specifically, in step 4, the working mechanism of the multi-output gated recurrent unit includes:

[0039] (7)

[0040] (8)

[0041] (9)

[0042] (10)

[0043] in, Input for the current moment, is the hidden state at the previous moment, To reset the gate, To update the gate, is the candidate hidden state at the current moment, The hidden state at the current moment, For different weight matrices, For different bias vectors, and represent the Sigmoid function and the hyperbolic tangent function respectively, and ⊙ represents the Hadamard product.

[0044] Specifically, the performance of the wind power prediction algorithm is evaluated based on the root mean square error, mean absolute error and determination coefficient.

[0045] (13)

[0046] (14)

[0047] (15)

[0048] in, represents the total number of samples in the test dataset, and They are The measured and predicted values ​​of samples, is the sample mean.

[0049] The beneficial effects of the present invention are as follows: first, this scheme develops a new outlier detection algorithm based on the K-means clustering algorithm and the Gaussian kernel density estimation method to screen out outliers in the data set, and develops a new outlier correction method based on the window smoothing method and theoretical power to reconstruct missing values, comprehensively evaluate the coupling relationship between wind speed and wind power series, and further reduce the negative impact of outliers on the model.

[0050] Moreover, while constructing a multi-output gated recurrent unit model to simultaneously predict wind speed and wind power, a new loss function is proposed. During the model training process, a larger weight is given to the wind power loss value, so that the model focuses more on the wind power prediction task during the training process.

[0051] Furthermore, in order to improve the estimation accuracy of the algorithm, relevant domain knowledge is encoded as constraints, and the model's predicted values ​​are constrained and corrected. A fine-tuning correction module is developed based on wind power field knowledge to make the prediction results more consistent with the physical laws and operating characteristics of actual wind power output, thereby improving the accuracy and reliability of the prediction.

[0052] In addition, MGRU is used as the basic learner for synchronous prediction of wind speed and wind power, so as to further explore the nonlinear interaction relationship between the two.

[0053] In summary, the wind power prediction algorithm of this scheme not only realizes the detection and processing of outliers, but also combines relevant domain knowledge with time series neural network to realize fast and accurate estimation of wind power, which helps to optimize the scheduling of power systems and reduce operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a framework diagram of the wind power prediction algorithm of the present invention;

[0055] Figure 2 This is a flow chart of the preliminary abnormal value detection and processing of the wind speed sequence of the present invention;

[0056] Figure 3 The original data in the embodiment of the present invention - wind power curve diagram;

[0057] Figure 4 is a graph of outlier detection results based on a new outlier detection algorithm in an embodiment of the present invention;

[0058] Figure 5 A wind power sequence and its components in an embodiment of the present invention;

[0059] Figure 6 The wind speed sequence and its component diagrams in the embodiment of the present invention;

[0060] Figure 7 This is a flow chart of the fine-tuning and correction module of the present invention;

[0061] Figure 8 It is a curve diagram of the actual value and predicted value of wind power according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the present invention clearer and more understandable, the present invention is described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the given embodiment is only one implementation method and does not represent all embodiments.

[0063] Embodiment 1

[0064] This example uses process data collected from a wind farm in eastern China. The original data set only contains historical data of two variables, wind speed and wind power, with a sampling interval of 10 minutes. There are 3,000 samples in the original data set. The historical data of the two variables are used as input, and the values ​​of the two variables in the next ten minutes are predicted.

[0065] like Figure 3 , it can be seen that there are obvious outliers around the wind power curve drawn using the original data. According to the causes of outliers, there are two main types. One is the horizontal dense points accumulated at the bottom of the wind power curve, whose wind speed is higher than the cut-in wind speed, but the corresponding wind power is still 0, which is usually caused by the shutdown of the wind turbine; the other is the discrete points around the wind power curve, which are usually caused by sensor failure, degradation or wind turbulence. These outliers will cause the model to have the risk of overfitting and affect the generalization ability of the model. Therefore, before training the model, it is necessary to identify and remove outliers to reduce the negative impact of outliers on the model. Secondly, due to the volatility and non-stationarity of wind speed and wind power series, it is difficult for the model to grasp their changing trends. Therefore, it is necessary to use a signal decomposition algorithm to decompose the original signal into multiple components with different frequency characteristics to help extract the potential patterns and trends in the signal, thereby improving the model's ability to capture data changes. At the same time, a neural network model with time series analysis capabilities is used to capture the dynamic characteristics of variables. In addition, the soft measurement algorithm based only on the data-driven model is a black box model, which does not combine the relevant knowledge in the field of industrial processes and may produce prediction values ​​that violate the mechanism. Therefore, in order to improve the estimation accuracy of the algorithm, it is necessary to encode the relevant field knowledge as constraints and perform constraint correction on the prediction value of the model.

[0066] Based on the above, this embodiment provides a new outlier correction wind power prediction algorithm, which realizes outlier detection and processing through a soft measurement algorithm and combines relevant field knowledge with a time series neural network. The algorithm specifically includes the following steps:

[0067] Step 1: Collect historical wind speed data and wind power data as the original data set, perform preliminary outlier detection and elimination on the historical wind speed data based on the box plot method, and then use the cubic spline interpolation method to complete the eliminated historical wind speed data.

[0068] In this embodiment, the first 70% of the original data set is used as training data, and the remaining 30% is used as test data. In this step, the initial outlier detection and elimination are first performed on the obvious outliers in the wind speed sequence to improve the stability of the original wind speed sequence, and then the missing values ​​are reconstructed using the cubic spline interpolation method to ensure the temporal continuity of the wind speed data.

[0069] The flowchart of preliminary outlier detection and processing of wind speed data is as follows: Figure 2 As shown, the specific implementation process includes:

[0070] Step 1.1, calculate the lower quartile Q1, median Q2, upper quartile Q3 and interquartile range (IQR) of the wind speed data. Q1 is the 25% quantile of the data, Q2 is the median value of the data, and Q3 is the 75% quantile of the data. These three constitute the box of the box plot. IQR is the range between Q1 and Q3, and its value is Q3-Q1.

[0071] Step 1.2, calculate the upper and lower limits of the box plot. The upper limit of the box plot is Q1-1.5*IQR, and the lower limit is Q3+1.5*IQR. Points outside the upper and lower limits are outlier points and are removed.

[0072] Step 2: For the original data set that has been preliminarily processed in step 1, a new outlier detection algorithm is used to screen out outliers and then remove them, and then a new outlier correction method is used to reconstruct missing values. The new outlier detection algorithm is based on the K-means clustering algorithm and the Gaussian kernel density estimation method, and the new outlier correction method is based on the window smoothing method and theoretical power.

[0073] like Figure 3 ,The wind power curve drawn using the original data has a large number of discrete points around and at the bottom of the curve. Therefore, for outliers, a new outlier detection algorithm is proposed to perform deep outlier detection, comprehensively evaluate the coupling relationship between wind speed and wind power series, and then a new outlier correction method is proposed. The specific implementation process includes:

[0074] Step 2.1, use the K-means clustering algorithm to cluster the sample points. The algorithm clusters the data points with similar wind power values ​​into one category and divides them into multiple category clusters.

[0075] Step 2.2: In each category cluster, the wind speed value corresponding to each sample point is analyzed and processed using the Gaussian kernel density estimation method to obtain the kernel density estimation value of each sample point, which is negated and recorded as the outlier score. The higher the score, the more serious the dispersion of the sample point.

[0076] Step 2.3, set the judgment threshold. If the sample outlier score exceeds the threshold, it is judged as an outlier. Figure 4 , shows the results of outlier detection, indicating that all obvious outlier points have been successfully identified.

[0077] Step 2.4, remove the detected outliers, and then fill in the missing values ​​in two steps. First, the window smoothing method is used to calculate the mean of the two samples before and after the missing values ​​in the wind speed sequence and wind power sequence, respectively, to effectively fill in the missing values ​​around the wind power curve. Secondly, due to the window smoothing method, there may be certain errors in the wind power interpolation values ​​corresponding to the outliers at the bottom of the power curve. Therefore, the theoretical power in the domain knowledge is used to correct the interpolated wind power value, and the theoretical wind power value is calculated according to the corresponding mathematical expression between wind speed and wind power to correct the corresponding wind power value. The mathematical expression between wind speed and wind power is:

[0078] (1)

[0079] in, Indicates the wind speed at the hub; Indicates the cut-in wind speed; Indicates rated wind speed; Indicates the cut-out wind speed; Indicates rated power; Indicates the wind power generated under different wind speed conditions; represents the wind energy utilization coefficient; is the wind swept area of ​​the fan blades; is the air density.

[0080] Step 3: For the original data set after deep processing in step 2, the wind speed and wind power series are decomposed using variational mode, and the neural network model is used to mine the information in the wind speed historical data and the wind power historical data.

[0081] The multi-output gated recurrent unit (MGRU) is used to decompose the wind speed and wind power sequence signals for more effective analysis and processing. Figure 5 and Figure 6 , VMD decomposes these two variables into 12 intrinsic mode functions (IMFs) components, and its specific implementation process includes:

[0082] Step 3.1, initialize the number of decomposed modes , the optimization problem is constructed by the variational principle, the goal is to minimize the reconstruction error and modal bandwidth, and the specific form of the constrained variational expression is:

[0083] (2)

[0084] in, Indicates The center frequency of the sub-signal, is the Dirac distribution, Indicates about time The partial derivative of Indicates IMF components; represents the original signal; represents the complex exponential function.

[0085] Step 3.2, by introducing the Lagrange multiplier and the quadratic penalty factor , integrate the constraints into the optimization problem to form the Lagrangian function:

[0086]

[0087] Step 3.3, iteratively update each mode by alternating direction multiplier method , and The optimal solution of , the iterative update expression is:

[0088]

[0089] in, is the number of iterations, , , and They are , , and The Fourier transform of .

[0090] Step 3.4, By taking the inverse Fourier transform, we can get the expression of each component in the time domain, and thus get the decomposed sub-signals.

[0091] Step 4: Take wind power as the main prediction variable and wind speed as the auxiliary prediction variable to construct a multi-output gated recurrent unit (MGRU) model to simultaneously predict wind speed and wind power. At the same time, a new loss function is proposed to enhance the model's attention to the wind power prediction task during training.

[0092] Use MGRU as the basic learner to simultaneously predict wind speed and wind power, thereby deepening

[0093] The nonlinear interaction between the two is explored. The working mechanism of the multi-output gated recurrent unit includes:

[0094] (7)

[0095] (8)

[0096] (9)

[0097] (10)

[0098] in, Input for the current moment, is the hidden state at the previous moment, To reset the gate, To update the gate, is the candidate hidden state at the current moment, The hidden state at the current moment, For different weight matrices, For different bias vectors, and represent the Sigmoid function and the hyperbolic tangent function respectively, and ⊙ represents the Hadamard product.

[0099] Since the main task of the model is to realize wind power prediction, a new loss function is proposed:

[0100] (11)

[0101] in, is the adjustment factor, whose value range is (1, 2), used to adjust the loss ratio; is the loss value of wind power; is the loss value of wind speed; is the loss value of the forward propagation of the MGRU model. During the model training process, a larger weight is given to the wind power loss value, so that the model focuses more on the wind power prediction task during the training process.

[0102] Step 5: Develop a fine-tuning correction module based on the domain knowledge of the wind power generation process to constrain and correct the wind power prediction value.

[0103] The soft measurement algorithm based only on the data-driven model is a black box model. It does not combine the relevant knowledge in the field of industrial processes and may produce prediction values ​​that violate the mechanism. Therefore, in order to improve the estimation accuracy of the algorithm, it is necessary to encode the relevant field knowledge as constraints and perform constraint correction on the prediction value of the model. In this embodiment, a fine-tuning correction module is developed based on the knowledge in the wind power field, such as Figure 7 , making the prediction results more consistent with the physical laws and operating characteristics of actual wind power output, and improving the accuracy and reliability of the prediction.

[0104] The specific execution process of the fine-tuning correction module includes:

[0105] Step 5.1, combining the mechanism relationship between wind speed and wind power, calculate the corresponding theoretical wind power value according to the wind speed forecast value, perform weighted recombination on the theoretical value and the wind power forecast value after preliminary correction, and constrain and correct the power forecast value that deviates significantly from the true value. The weighted recombination formula is:

[0106] (12)

[0107] in, is the weight coefficient, whose value range is (0, 1), which is used to adjust the value ratio; is the wind power value after constraint correction; is the theoretical wind power value; is the wind power forecast value after preliminary correction.

[0108] Step 5.2: According to the knowledge in the field of wind power, negative power generation is unreasonable. Therefore, if the power prediction value is negative, it is directly corrected to 0.

[0109] Step 5.3, combine the characteristics and mechanism knowledge of wind power curve to fine-tune the power value. When the wind speed is less than the cut-in wind speed, the wind turbine will not generate power; when the wind speed exceeds the rated wind speed and is less than the cut-out wind speed, the output power should be the rated wind speed. However, if Figure 3 The actual wind power curve is different from the theoretical curve. Under the same wind power, the wind speed has a certain range of variation, which is called the credible range. Based on this characteristic, the power value corresponding to the wind speed prediction value less than the cut-in wind speed credible range is set to 0; the power value corresponding to the rated wind speed credible range is set to the rated power value.

[0110] It should be noted that the performance of all algorithms in this embodiment is based on the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (CDR). ) for evaluation,

[0111] (13)

[0112] (14)

[0113] (15)

[0114] in, represents the total number of samples in the test dataset, and They are The measured and predicted values ​​of samples, is the sample mean.

[0115] Table 1 below shows the simulation results of each model on two data sets. The comparison shows the effectiveness of the developed outlier detection algorithm and missing value reconstruction method. At the same time, it can be seen from the performance of each data set that the performance of the MGRU model constructed in this embodiment is better than other algorithms in all indicators. Figure 8 To utilize the predicted value of the target variable by the wind power prediction algorithm in this embodiment, it is shown that the developed algorithm can effectively and dynamically follow the changes in the target variable value.

[0116] Table 1 Simulation results of different algorithms in various data sets

[0117]

Claims

1. A new outlier-corrected wind power prediction algorithm, characterized in that: The steps include: Step 1: collect historical wind speed data and historical wind power data as the original data set, perform preliminary outlier detection and elimination on the historical wind speed data based on the box plot method, and then use the cubic spline interpolation method to complete the eliminated historical wind speed data; Step 2: For the original data set that has been preliminarily processed in step 1, a new outlier detection algorithm is used to screen out outliers and then remove them, and then an outlier correction method is used to reconstruct missing values. The outlier detection algorithm is based on the K-means clustering algorithm and the Gaussian kernel density estimation method, and the new outlier correction method is based on the window smoothing method and theoretical power. The specific execution process of the deep outlier detection and elimination and missing value completion correction includes: Step 2.1, using the K-means clustering algorithm to cluster data points with similar wind power values ​​into one category, and divide the sample points into multiple category clusters; Step 2.2: In each category cluster, the Gaussian kernel density estimation method is used to analyze the wind speed value corresponding to each sample point to obtain the kernel density estimation value of each sample point, and the inverse of the kernel density estimation value is recorded as the outlier score; Step 2.3, set the judgment threshold, and the sample outlier score exceeding the threshold is judged as an outlier; Step 2.4, after removing the detected abnormal values, fill in the missing values ​​in two steps: first, use the window smoothing method to calculate the mean of the two samples before and after the missing value in the wind speed series and wind power series; second, according to the corresponding mathematical expression between wind speed and wind power, calculate the theoretical wind power value to correct the corresponding wind power value. The mathematical expression between wind speed and wind power is: (1) in, Indicates the wind speed at the hub; Indicates the cut-in wind speed; Indicates rated wind speed; Indicates the cut-out wind speed; Indicates rated power; Indicates the wind power generated under different wind speed conditions; represents the wind energy utilization coefficient; is the wind swept area of ​​the fan blades; is the air density; Step 3: for the original data set after deep processing in step 2, the wind speed and wind power sequences are decomposed by variational mode, and the neural network model is used to mine the information in the wind speed historical data and the wind power historical data; Step 4: Take wind power as the main prediction variable and wind speed as the auxiliary prediction variable to build a multi-output gated recurrent unit model to simultaneously predict wind speed and wind power; at the same time, a new loss function is proposed to enhance the model's attention to the wind power prediction task during training. The new loss function is: (11) in, is the adjustment factor, whose value range is (1, 2), used to adjust the loss ratio; is the loss value of wind power; is the loss value of wind speed; is the loss value of the forward propagation of the MGRU model; Step 5: Develop a fine-tuning correction module for constraining and correcting the wind power forecast value. The specific execution process of the fine-tuning correction module includes: Step 5.1, calculate the corresponding theoretical wind power value according to the wind speed forecast value, perform weighted recombination on the theoretical wind power value and the initially corrected wind power forecast value, and constrain and correct the wind power forecast value that deviates from the true value. The weighted recombination formula is: (12) in, is the weight coefficient, whose value range is (0, 1), which is used to adjust the value ratio; is the wind power value after constraint correction; is the theoretical wind power value; is the wind power forecast value after preliminary correction; Step 5.2, correct the negative wind power forecast value to 0; Step 5.3, fine-tune the power value based on the characteristics and mechanism knowledge of wind power curve: for the power value corresponding to the wind speed prediction value less than the credible range of the cut-in wind speed, set it to 0; for the power value corresponding to the wind speed exceeding the credible range of the rated wind speed, set it to the rated power value.

2. The novel outlier-corrected wind power prediction algorithm according to claim 1 is characterized by: In step 1, the specific execution process of preliminary detection and elimination of obvious outliers in the wind speed sequence includes: Step 1.1, calculate the lower quartile Q1, median Q2, upper quartile Q3 and interquartile range IQR of the wind speed data, where Q1 is the 25% quantile of the data, Q2 is the median value of the data, Q3 is the 75% quantile of the data, and IQR is the range between Q1 and Q3, and its value is Q3-Q1; Step 1.2, calculate the upper and lower limits of the box plot. Points outside the upper and lower limits are considered outliers and are removed. The upper limit of the box plot is Q1-1.5*IQR, and the lower limit is Q3+1.5*IQR.

3. The novel outlier-corrected wind power prediction algorithm according to claim 1 is characterized in that: In step 2, the data set includes 3,000 samples, 70% of the data set is used as training data, and the remaining 30% is used as test data.

4. The novel outlier-corrected wind power prediction algorithm according to claim 1 is characterized by: In step 3, the multi-output gated loop unit decomposes the wind speed and wind power sequences into 12 intrinsic mode function components respectively. The specific execution process is: include: Step 3.1, initialize the number of decomposed modes , the optimization problem is constructed by the variational principle, the goal is to minimize the reconstruction error and modal bandwidth, and the specific form of the constrained variational expression is: (2) in, Indicates The center frequency of the sub-signal, is the Dirac distribution, Indicates about time The partial derivative of Indicates IMF components; represents the original signal; represents a complex exponential function; Step 3.2, by introducing the Lagrange multiplier and the quadratic penalty factor , integrate the constraints into the optimization problem to form the Lagrangian function: Step 3.3, iteratively update each mode by alternating direction multiplier method , and The optimal solution of , the iterative update expression is: in, is the number of iterations, , , and They are , , and Fourier transform of Step 3.4, By taking the inverse Fourier transform, we can get the expression of each component in the time domain, and thus get the decomposed sub-signals.

5. The novel outlier-corrected wind power prediction algorithm according to claim 1 is characterized in that: In step 4, the working mechanism of the multi-output gated recurrent unit includes: (7) (8) (9) (10) in, Input for the current moment, is the hidden state at the previous moment, To reset the gate, To update the gate, is the candidate hidden state at the current moment, The hidden state at the current moment, For different weight matrices, For different bias vectors, and represent the Sigmoid function and the hyperbolic tangent function respectively, and ⊙ represents the Hadamard product.

6. The novel outlier-corrected wind power prediction algorithm according to claim 1 is characterized by: The performance of the wind power prediction algorithm is evaluated based on the root mean square error, mean absolute error and coefficient of determination. (13) (14) (15) in, represents the total number of samples in the test dataset, and They are The measured and predicted values ​​of samples, is the sample mean.

Images