Wind power cluster power day-ahead prediction method

Abstract

The invention relates to the technical field of wind power short-term prediction, and discloses a wind power cluster power day-ahead prediction method, which comprises the following steps of: summarizing space-time correlation characteristics of a wind power cluster into a long-distance convergence effect and a short-distance time sequence causal relationship between wind power plants; an improved set pair analysis method, a coincidence rate discrimination method and a variance discrimination method are used to obtain a convergence effect adjacency matrix surface, a Pearson correlation coefficient and a Granger causality test are used to obtain a time sequence causality adjacency matrix, and an error evaluation characteristic coefficient formula is proposed to construct an error adjacency matrix. A traditional graph convolutional neural network is improved, and a graph neural network structure based on a multi-channel attention mechanism is provided. The graph convolution prediction method considering the wind power cluster convergence effect and multi-channel attention mechanism fusion is high in prediction stability, high in prediction performance, effective in prediction result and high in practicability.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term wind power prediction, and in particular to a method for day-ahead prediction of wind power cluster power. Background Art

[0002] In recent years, the installed capacity of new energy has continued to increase rapidly, and the transformation of the world's energy structure is accelerating. Due to the randomness, intermittency and volatility of wind power, as the scale of wind power continues to expand and its share in the energy structure increases, the impact of wind power on the safe and stable operation of the power system will also increase. In order to ensure the balance of regional power, large-scale wind power cluster forecasting is imperative.

[0003] The wind power day-ahead forecast refers to the forecast for the next day from the forecast time, with a time resolution of 15 minutes. The significance of the wind power day-ahead forecast is to provide a reference for planning wind farm maintenance plans, spare capacity preparation plans, etc.

[0004] There are many existing studies on cluster wind power prediction, but there are still some shortcomings: there are studies on the output evolution law between wind farms in a wide area over a long period of time, but they are rarely applied; causal algorithms and correlation algorithms are difficult to characterize the spatiotemporal correlation between wind farms that are far apart; the description of the spatiotemporal correlation of wind power clusters is vague, and there is no clear definition or summary; the application of errors only stays at the correction of prediction results and probabilistic prediction, and there is a lack of research on the extraction and application of error characteristics. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the power of a wind power cluster before its arrival, which provides a definition summary of the spatiotemporal correlation of a wind power cluster based on the spatiotemporal characteristics of the wind power cluster, and provides a method for predicting the power of a wind power cluster before its arrival, which is clear in meaning, scientific, reasonable and effective, reliable and stable in prediction, can take into account the spatiotemporal correlation and convergence effect of the wind power cluster, and incorporates the wind power prediction error characteristics.

[0006] In order to achieve the above purpose, the technical solutions adopted are as follows:

[0007] The present invention provides a method for predicting wind power cluster power day-ahead, the method comprising:

[0008] Based on the wind power sequence, construct the adjacency matrix A s , simultaneous rate adjacency matrix A t , variance adjacency matrix A v ;

[0009] According to the adjacency matrix A s , simultaneous rate adjacency matrix A t and the variance adjacency matrix A v , construct the convergence effect adjacency matrix A a;

[0010] Use Pearson correlation coefficient to construct the adjacency matrix A r , and construct the causal adjacency matrix A based on Granger causality test g ; Wherein, the adjacency matrix A r and the causal adjacency matrix A g Used to characterize the relationship between wind farms in close proximity;

[0011] Based on the adjacency matrix A r and the Granger causal adjacency matrix A g , construct the temporal causal adjacency matrix A b ;

[0012] Construct the error characteristic adjacency matrix A e ; Wherein, the error characteristic adjacency matrix A e Used to characterize the errors between wind farms;

[0013] A graph convolutional neural network model based on the channel attention mechanism is constructed. In the graph convolutional neural network model, multiple graph convolution modules are designed to input adjacency matrices of different attributes respectively to form multiple channels. Each channel is used to extract different characteristic laws in the wind power cluster to obtain multiple features. The function definition is as follows:

[0014]

[0015]

[0016]

[0017] In the formula, The l+1th layer features representing the convergence effect, The l+1th layer features representing temporal causality, represents the l+1th layer feature of the error, f represents the graph convolution function, The l-th layer features representing the convergence effect, The l-th layer features representing temporal causality, The l-th layer features representing the error;

[0018] On the basis of multiple channels, the attention mechanism is introduced to learn the weight of each feature for prediction importance, and finally combine multiple characteristics to obtain the day-ahead prediction result of wind power cluster power. The function is defined as follows:

[0019]

[0020]

[0021] In the formula, Softmax(·) represents the normalization function, concat(·) represents the concatenation function, Output represents the day-ahead wind power cluster power forecast result, W a represents the weight of the convergence effect feature, W b represents the weight of temporal causal features, W e represents the weight of the error feature, H represents the fusion feature, represents the learnable parameter of the l+1th layer convergence effect, represents the learnable parameters of the l+1th layer of temporal causality, Represents the learnable parameters of the error features of the l+1th layer.

[0022] Furthermore, the adjacency matrix A is constructed by the following method: s :

[0023] The following formula is used to traverse and calculate the connection degree between each wind farm in the wind power cluster to obtain the connection degree matrix:

[0024]

[0025]

[0026]

[0027] Where, μ(A,B) is the connection matrix between wind power sequence A and wind power sequence B; {Δx i ,Δx i+1}, {Δx' i ,Δx' i+1} respectively represent the adjacent difference values ​​of the i-th identical characteristic element of wind power sequence A and wind power sequence B; {Δx j ,Δx j+1}, {Δx' j ,Δx' j+1} respectively represent the adjacent difference values ​​of the jth opposite characteristic elements of wind power sequence A and wind power sequence B; I is the difference coefficient; J is the opposition identification coefficient; N is the total number of features in the set; α i is the characteristic strength of the i-th identical set pair; β j is the characteristic strength of the jth opposite set pair; S' is the total strength of the same set pair; F' is the total strength of the uncertainty set pair; P' is the total strength of the opposite set pair;

[0028] In order from small to large, select the first m elements in the connection matrix, where m is the average number of connections of the simultaneous rate and variance, retain m element values, and set the rest to zero to form a matrix M;

[0029] Use matrix M to construct adjacency matrix A based on set pair analysis s, the calculation process is as follows:

[0030]

[0031]

[0032] In the formula, μ ij is the connection degree between the i-th and j-th wind farms; ∞ means that there is no connection between the two wind farms; τ is a constant used to avoid the edge weight being 0; S ij Represents the edge weights established by set pair analysis.

[0033] Furthermore, the simultaneous rate adjacency matrix A is constructed by the following method: t :

[0034] Traverse and calculate the daily synchronization rate of every two wind farms in the wind power cluster to form a matrix T k×i×j , the formula is as follows:

[0035]

[0036]

[0037]

[0038] In the formula, represents the power sequence of the i-th wind farm on the k-th day, represents the power sequence of the jth wind farm on the kth day, Cap i Cap represents the operating capacity of the i-th wind farm, j represents the operating capacity of the j-th wind farm; represents the simultaneous rate matrix of the wind power cluster on the kth day, represents the simultaneous rate of adding the power of wind farm No. i and No. j on day k;

[0039] For the matrix T k×i×j Each sub-matrix in is discriminated to obtain the discriminant matrix of each day, and the discriminant matrix of each day is summed up to obtain the simultaneous rate discriminant matrix H i×j , the formula is as follows:

[0040]

[0041]

[0042]

[0043]

[0044] In the formula, represents the discriminant matrix of the kth day, represents the judgment result of the i-th row and j-th column on the k-th day, h ij Represents the sum of the judgment results of the i-th row and j-th column of all days; represents the simultaneous rate of the power of wind farm No. j on day k;

[0045] Using the simultaneous rate discriminant matrix to perform adjacency matrix A of directed weighted graph t The specific formula for constructing is as follows:

[0046]

[0047]

[0048]

[0049] Where W ij represents the weight of wind farm No. i pointing to wind farm No. j; t ij Represents the edge weight established by the simultaneous rate.

[0050] Furthermore, the variance adjacency matrix A is constructed by the following method: v :

[0051] Construct a variance matrix and construct a variance adjacency matrix A based on the variance matrix v ; Among them, the construction formula of the variance matrix is ​​as follows:

[0052]

[0053]

[0054]

[0055]

[0056] Where V k×i×j represents the variance matrix, represents the normalized sequence of the sum of wind farm No. i and wind farm No. j, represents the sequence mean, n is the number of elements in the wind power sequence, w is the sequence number of the wind power sequence, represents the variance matrix of the wind power cluster on the kth day, It represents the sum of the variance of wind farms No. i and No. j on day k.

[0057] Furthermore, according to the adjacency matrix A s , simultaneous rate adjacency matrix A t and the variance adjacency matrix A v , the convergence effect adjacency matrix A is constructed by the following formula a :

[0058]

[0059]

[0060] In the formula, a ij represents the edge weight established by the convergence effect, s ij represents the edge weight constructed by the improved set pair analysis, t ij represents the edge weight constructed by the simultaneous rate, v ij Denotes the edge weights of the variance construction.

[0061] Furthermore, the adjacency matrix A is constructed using the Pearson correlation coefficient r ,include:

[0062] Traverse and calculate the daily synchronization rate of every two wind farms in the wind power cluster to form a matrix T k×i×j , the formula is as follows:

[0063]

[0064]

[0065]

[0066] In the formula, represents the correlation coefficient matrix of the wind power cluster on the kth day; represents the correlation coefficient between the power of wind farm No. i and No. j on day k; denote the sequence of the kth day for the i-th and j-th wind farms, respectively; They represent the mean values ​​of the i-th and j-th wind farms on the k-th day respectively; express;

[0067] For the matrix T k×i×j Each sub-matrix in is discriminated to obtain the discriminant matrix of each day, and the discriminant matrix of each day is summed up to obtain the simultaneous rate discriminant matrix H i×j , the formula is as follows:

[0068]

[0069]

[0070]

[0071]

[0072] In the formula, represents the discriminant matrix of the kth day, represents the judgment result of the i-th row and j-th column on the k-th day, h ij Represents the sum of the judgment results of the i-th row and j-th column of all days;

[0073] Using the simultaneous rate discriminant matrix to perform adjacency matrix A of directed weighted graph r 's construction.

[0074] Furthermore, the causal adjacency matrix A is constructed based on the Granger causality test. g ,include:

[0075] The Granger test is used to traverse and test whether the daily power series of each wind farm have Granger causality, so as to obtain the discriminant matrix, and the causal adjacency matrix A is constructed based on the discriminant matrix. g ; Among them, the calculation formula for using the Granger test to traverse and test whether the daily power series of two wind farms have Granger causality is as follows:

[0076]

[0077] In the formula, represents the Granger causality of the i-th wind farm to the j-th wind farm in the k-th day; represents the Granger test value of the i-th wind farm on the j-th wind farm in the k-th day; Fa represents the Granger test threshold.

[0078] Further, construct the error characteristic adjacency matrix A e ,include:

[0079] Based on the error smoothing evaluation coefficient, the error characteristic adjacency matrix A e ; Wherein, the error smoothing evaluation coefficient is expressed as:

[0080]

[0081] In the formula, represents the normalized value of the sum of the errors of wind farm No. i and No. j on the kth day; They represent the k-th day prediction error sequences of the i-th wind farm and the j-th wind farm respectively.

[0082] Furthermore, the method further comprises:

[0083] Select the normalized root mean square error NRMSE, the normalized mean absolute error NMAE and the coefficient of determination R 2 As the evaluation index of the prediction model performance, the calculation formulas are:

[0084]

[0085]

[0086]

[0087] Where n is the predicted power sequence length, y i,pre and i,ture are the predicted and actual values ​​of power, respectively. It represents the average value of actual power value, and Cap is the installed capacity.

[0088] The beneficial effects of the present invention are:

[0089] Starting from the spatiotemporal characteristics of wind power clusters, this paper proposes a feature extraction method that can simultaneously consider the short-range and long-range associations of wind farms. At the same time, based on the error characteristics of wind power, a method that can directly introduce the wind power prediction error into the feature extraction and prediction process is proposed. Finally, a multi-channel attention mechanism for graph convolutional networks is proposed to integrate adjacency matrix information of multiple different attributes, thereby improving the prediction accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 A flow chart of a method for day-ahead power prediction of a wind power cluster according to an embodiment of the present invention is shown.

[0091] Figure 2 A schematic diagram of set pair characteristic evaluation according to an embodiment of the present invention is shown.

[0092] Figure 3 A characteristic strength analysis diagram according to an embodiment of the present invention is shown; wherein (a) is a case where the opposite characteristic strength is the highest; (b) is a case where the opposite characteristic strength is relatively weak; (c) is a case where the opposite characteristic strength is the highest; (d) is a case where the opposite characteristic strength is the highest;

[0093] Figure 4 A structural diagram of a graph convolutional neural network based on a channel attention mechanism according to an embodiment of the present invention is shown;

[0094] Figure 5 The prediction result diagram of the cluster prediction strategy according to the embodiment of the present invention is shown; wherein (a) is a comparison of prediction curves of different methods in cluster A; (b) is a comparison of prediction curves of different methods in cluster B; (c) is a comparison of prediction curves of different methods in cluster C;

[0095] Figure 6 The prediction error distribution diagram of various prediction strategies according to the embodiments of the present invention is shown (the unit of frequency is "piece"); among them, (a) the prediction error frequency statistics of different methods in cluster A; (b) the prediction error frequency statistics of different methods in cluster B; (c) the prediction error frequency statistics of different methods in cluster C. DETAILED DESCRIPTION

[0096] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0097] The specific implementation of the present invention is further described in detail below in conjunction with the drawings and examples.

[0098] The embodiment of the present invention provides a method for predicting the power of a wind power cluster. Figure 1 As shown, the wind power cluster power day-ahead prediction method includes the following steps:

[0099] Step S10: Obtain the adjacency matrix A based on set pair analysis by using the improved set pair analysis method s .

[0100] Step S20: Obtain the simultaneous rate adjacency matrix A by calculating the simultaneous rate changes between two wind farms t .

[0101] Step S30: Establish the variance adjacency matrix A v , the set pair analysis adjacency matrix A s , simultaneous rate adjacency matrix A t Take the intersection of the three and the variance adjacency matrix to get the convergence effect adjacency matrix A a .

[0102] Step S40: Obtain the correlation coefficient adjacency matrix A through the Pearson correlation coefficient r ;

[0103] Step S50: Use Granger causality test to obtain the causal adjacency matrix A g , the correlation coefficient adjacency matrix A r and the causal adjacency matrix A g Take the intersection to get the temporal causal adjacency matrix A b .

[0104] Step S60: Using the initial prediction error, construct the error characteristic adjacency matrix A e .

[0105] Step S70: Input the NWP data of the wind power cluster, the convergence effect adjacency matrix, the temporal causal adjacency matrix and the error characteristic adjacency matrix into a graph convolutional neural network based on a multi-channel attention mechanism to obtain predicted power.

[0106] The following embodiment of the present invention will describe in detail the specific implementation process of the above seven steps.

[0107] For step S10, set pair analysis, also known as connection mathematics, is a new system analysis method that uses connection degree to deal with uncertainty problems. The core idea of ​​set pair analysis is to analyze and process the deterministic connection and uncertainty connection between the objective things being studied as an uncertain system. Among them, certainty includes two aspects, "sameness" and "opposition", while uncertainty refers only to "difference". Set pair analysis analyzes things and their systems through the three aspects of identity, difference and opposition. According to the relationship between the three elements of sameness, difference and opposition, the expression of the sameness, difference and opposition connection degree between two sets is:

[0108]

[0109] In the formula: μ(A, B) is the degree of connection between set A and set B; N is the total number of features in the set, N=S+P+F; S is the number of elements in the same state (same characteristics) in the two sets; P is the number of elements in the opposite state (opposite characteristics) in the two sets; F is the number of elements in the different state (different characteristics) in the two sets, that is, the number of features in the set that are neither identical nor opposite; I is the difference coefficient, which can be taken in the interval (-1,1) according to the specific situation; J is the opposition identification coefficient.

[0110] For set pair analysis, it is crucial to determine what characteristics each element in two sequences has. In the set pair analysis method defined in this paper, the total number of characteristics N+2 is the total number of sequence elements, that is, except for the first two elements in a sequence, all other sequence elements have the characteristics.

[0111] However, in the existing set pair analysis method, only the number of characteristics of the elements is counted, without considering the strength of the characteristics. Therefore, the improved set pair analysis adds the strength evaluation of the characteristics on the basis of the original set pair analysis, so as to more reasonably characterize the strength of the smoothing effect between the two sequences. The specific formula is as follows:

[0112]

[0113]

[0114]

[0115] Among them, {Δx i ,Δx i+1}, {Δx' i ,Δx' i+1} respectively represent the adjacent difference values ​​of the i-th identical characteristic element of sequence A and B. j ,Δx j+1}, {Δx' j ,Δx' j+1} respectively represent the adjacent difference values ​​of the jth opposite characteristic elements of sequences A and B. In this paper, I is taken as 0.5 and J is usually taken as -1.

[0116] In the improved set pair analysis method, the smaller the value obtained, the more obvious the smoothing effect between wind farms.

[0117] The steps of constructing an adjacency matrix using set pair analysis in this embodiment are as follows:

[0118] Step S11, using formula (2) to traverse and calculate the connection degree between any two wind farms in the wind power cluster to obtain a connection degree matrix;

[0119] Step S12, select the first m elements (m is the average number of connections of the simultaneity rate and variance) in the connection degree matrix in ascending order, retain these m element values, and set the rest to zero, to form a matrix M;

[0120] Step S13: Use the matrix M to construct an adjacency matrix A based on set pair analysis s , the specific calculation process is as follows:

[0121]

[0122]

[0123] Among them, μ ij is the connection degree between the i-th and j-th wind farms. ∞ means that there is no connection between the two wind farms. τ represents a minimum value, which is used to avoid the weight of the edge being 0. This paper takes 10 -5 .

[0124] In step S20, for the construction of the simultaneity adjacency matrix, the wind power simultaneity refers to the ratio of the highest active power of the wind farm during the statistical period to the total capacity of the grid-connected units during the period, which can be used to describe the evolution of the smoothing effect from the wind farm to the wind farm cluster. This paper uses the simultaneity to construct the adjacency matrix, but using the simultaneity of the entire sequence to construct the adjacency matrix will cause the time characteristics of the wind power output to be diluted or ignored. Therefore, this paper divides the wind power sequence into multiple subsequences and considers a method for constructing an adjacency matrix based on historical statistics. The specific method steps are as follows:

[0125] Step S21, traverse and calculate the daily synchronicity rate of every two wind farms in the wind power cluster to form a matrix T k×i×j , the formula is as follows:

[0126]

[0127]

[0128]

[0129] Among them, P i k represents the power sequence of the i-th wind farm on the k-th day, represents the power sequence of the jth wind farm on the kth day, Cap i Cap represents the operating capacity of the i-th wind farm, j Represents the operating capacity of the jth wind farm.

[0130] Step S22: Matrix T k×i×j Each sub-matrix in the discriminant matrix is ​​discriminated to obtain the discriminant matrix of each day, and the discriminant matrix of each day is summed up to obtain the final simultaneous rate discriminant matrix H i×j , the formula is as follows:

[0131]

[0132]

[0133]

[0134]

[0135] in, represents the discriminant matrix of the kth day, represents the judgment result of the i-th row and j-th column on the k-th day, h ij Represents the sum of the judgment results of the i-th row and j-th column of all days.

[0136] Step S23: Use the simultaneous rate discriminant matrix to calculate the adjacency matrix A of the directed weighted graph t The specific formula is as follows:

[0137]

[0138]

[0139]

[0140] Among them, W ij It represents the weight of wind farm No. i pointing to wind farm No. j, and ∞ represents no connection (no edge).

[0141] In step S30, for the variance adjacency matrix, variance is used in probability theory and statistics to measure the degree of dispersion of a random variable or a set of data. To some extent, variance can also represent the degree of fluctuation of a sequence. Using variance to construct the adjacency matrix A v The process is the same as that of the simultaneous rate. It is only necessary to replace the simultaneous rate matrix T in step S20 k×i×jThe construction process is replaced by the variance matrix S k×i×j The construction process, the specific calculation formula is as follows:

[0142]

[0143]

[0144]

[0145]

[0146] in, represents the normalized sequence of the sum of wind farm No. i and wind farm No. j, Represents the mean of the sequence, and n is the number of elements in the sequence.

[0147] The final convergence effect adjacency matrix A a By A s , A t and A v The three adjacency matrices are calculated using the following formula:

[0148]

[0149]

[0150] In step S40, for the correlation coefficient adjacency matrix, wind farms in close proximity usually have a high output similarity due to similar wind characteristics, and the output of wind farms in the same wind direction may also have a time-series lag relationship. Therefore, this embodiment uses the classic Pearson correlation coefficient and Granger causality test to characterize the correlation between close-range wind farms. At the same time, the corresponding adjacency matrix is ​​constructed, and finally the two adjacency matrices are combined to obtain the final time-series causal adjacency matrix.

[0151] The Pearson correlation coefficient is widely used to measure the degree of correlation between two variables, and its value is between -1 and 1. Use the Pearson correlation coefficient to construct the adjacency matrix A r The process is similar to that of the simultaneous rate, except that formula (9) and formula (12) in step S20 are replaced by formulas (23) and (24).

[0152]

[0153]

[0154] in, They represent the k-th day sequence of the i-th and j-th wind farms respectively. They represent the mean values ​​of the i-th and j-th wind farms on the k-th day respectively.

[0155] In step S50, for the causal adjacency matrix, the Granger causality test was pioneered by Clive WJGranger, the winner of the 2003 Nobel Prize in Economics, and is used to analyze the causal relationship between variables. Granger causality is specifically defined as: if the prediction effect of variable Y is better than the prediction effect of Y based on the past information of Y alone, that is, variable X helps to explain the future changes of variable Y, then variable X is considered to be the Granger cause of variable Y. The specific steps are as follows:

[0156] Step S51: regress the current y on all the lagged terms y and the remaining variables (if any), i.e., y on the lagged terms y of y t-1 ,y t-2 , ..., y t-q and other variables, but the lagged term x is not included in this regression, which is a constrained regression. Then the constrained residual sum of squares RSS is obtained from this regression R .

[0157] Step S52: Perform a regression with a lag term x, that is, add the lag term x to the previous regression equation. This is an unconstrained regression. The unconstrained residual sum of squares RSS is obtained from this regression. UR .

[0158] Step S53: The null hypothesis is H O :a1=a2=...=a q =0, that is, the lagged term x does not belong to this regression.

[0159] Step S54: In order to test this hypothesis, the F test is used, that is: It follows an F distribution with q and (nk) degrees of freedom. Here, n is the sample size, q is equal to the number of lags, i.e., the number of parameters to be estimated in the constrained regression equation, and k is the number of parameters to be estimated in the unconstrained regression equation.

[0160] Step S55: If the F value calculated at the selected significance level a exceeds the critical Fa value, the null hypothesis is rejected, and the lagged x term belongs to this regression, indicating that x is the cause of y.

[0161] Step S56: Similarly, in order to check whether y is the cause of x, the variables y and x can be replaced with each other, and steps S51-S55 can be repeated.

[0162] The Granger test is used to traverse and test whether the daily power series of each wind farm have Granger causality, as shown in formula (25), and the discriminant matrix of step S21 in step S20 is directly obtained. Finally, the Granger causality adjacency matrix A is constructed by a method similar to that in step S20. g .

[0163]

[0164] The final temporal causal adjacency matrix A b By A r and A g Construction, calculation method and smoothing effect adjacency matrix A a same.

[0165] In step S60, for the error characteristic adjacency matrix, there is also a smoothing effect on the errors between wind farms, and the manifestation of this smoothing effect is more obvious and direct. Figure 4 The normalized errors between several groups of wind farms are shown in Figure 1. It is not difficult to find that there are a large number of positive and negative offset periods between the prediction errors of different wind farms. Although this positive and negative offset period has no benefit for a single wind farm, it can achieve prediction complementarity between wind farms and reduce the prediction error of wind power clusters. At the same time, this situation also illustrates the correlation between the two wind farms from the perspective of error. An error smoothing evaluation coefficient is proposed to quantify this relationship. The smaller the error smoothing evaluation coefficient, the stronger the prediction correlation between the two wind farms. The specific expression is as follows:

[0166]

[0167] in, They represent the k-th day forecast error sequences of wind farm No. i and wind farm No. j respectively. If all forecast error sequences of two wind farms are used in the calculation, then it is the total error smoothing evaluation coefficient of the two wind farms.

[0168] According to the error correlation between wind farms, a method for constructing an adjacency matrix is ​​proposed. In order to eliminate the interference of sequence length on error smoothing, the error smoothing evaluation coefficient is calculated every day and historical statistics are performed, just like the correlation coefficient and Granger causality test. Error adjacency matrix A e The construction process and simultaneous rate matrix A t Similarly, just replace formula (9) in step S20 with formula (26).

[0169] At present, graph neural networks have many applications in image and text tasks, and also in the field of wind power prediction. Because the spatial distribution of wind power clusters and the relationship between wind farms are very similar to the nodes and edges in graph neural networks, it is particularly appropriate to apply graph neural networks to wind power cluster prediction. Although there are many improved models in graph neural networks, the significance and prediction effect of graph convolutional neural networks as one of the classic basic models in graph neural networks cannot be ignored. The model function is defined as:

[0170]

[0171] Among them, H 0 =X,X∈R N*D , N is the number of nodes in the graph, D is the dimension of each node feature vector, H l ∈R N*d is the feature of the lth layer, H l+1 ∈R N*h is the updated feature of layer l+1, and A is the adjacency matrix. D∈R N*N yes The degree matrix of W, d and h constitute the characteristic dimension of each node, representing the power or NWP time series data of a wind farm. l ∈R d*h is the weight parameter matrix of the lth layer. σ(·) is the nonlinear activation function.

[0172] Therefore, in step S70, a graph convolutional neural network based on a channel attention mechanism is proposed, and the specific real-time steps are as follows:

[0173] Step S71: Taking the wind farm as a node, design multiple graph convolution modules, input adjacency matrices of different attributes respectively, form a multi-channel model, each channel is used to extract different characteristic rules in the wind power cluster, and obtain multiple characteristics. The function definition is as follows:

[0174]

[0175]

[0176]

[0177] Among them, A a , A b and A e They represent the smoothing effect adjacency matrix, the causal adjacency matrix and the error characteristic adjacency matrix respectively.

[0178] Step S72: Based on the multi-channel, the attention mechanism is introduced to learn the weight of each feature for prediction importance, and finally combine multiple characteristics to obtain the final output. The function definition is as follows:

[0179]

[0180]

[0181] Among them, Softmax(·) represents the normalization function, and concat(·) represents the concatenation function.

[0182] In a certain embodiment, the method for predicting the power of a wind power cluster day-ahead further includes step S80 and step S90 after step S70.

[0183] Step S80: simulation calculation.

[0184] Simulation input: measured power data of a wind power cluster including multiple wind farms of sufficient length; day-ahead forecast NWP data corresponding to the wind power cluster including multiple wind farms; data sampling interval is 15 minutes; according to step S10 to step S70, the day-ahead power forecast result of the wind power cluster is obtained.

[0185] Step S90: Error analysis.

[0186] Select the normalized root mean square error (NRMSE), normalized mean absolute error (NMAE) and determination coefficient (R-square, R 2 ) is used as the evaluation index of the prediction model performance, and the calculation formulas are:

[0187]

[0188]

[0189]

[0190] Where n is the length of the predicted power sequence, y i,pre and i,ture are the predicted and actual values ​​of power, respectively. It represents the average value of actual power value, and Cap is the installed capacity.

[0191] According to step S80, the simulation input quantity is input, and the predicted power calculated by the model is compared with the measured power through the root mean square error formula (33), the mean absolute error formula (34) and the determination coefficient (35) in step 90 to calculate the error.

[0192] To further demonstrate the feasibility and advancement of the present invention, this embodiment is experimentally verified using three clusters, namely, cluster A (Wind farm cluster A), cluster B (Wind farm cluster B), and cluster C (Wind farm cluster C). The wind farm cluster information is shown in Table 1.

[0193] Table 1 Wind power cluster information

[0194]

[0195] Based on the wind power cluster information shown in Table 1, the implementation steps of the method proposed in the present invention are as follows:

[0196] Step 1: Figure 2 Demonstrates the methods and basis for distinguishing various relationships in set pair analysis. Figure 3 Shows how the strength of a feature can differ due to the difference between the values ​​of the two series. Figure 3 (a) and Figure 3 (b) are opposite characteristics, but Figure 3 The strengths of the two opposite characteristics in (a) and (b) are quite different. Figure 3 (a)Δx a1 =Δx' a1 And Δx a2 =Δx' a2 , which is the most ideal case of opposite characteristics. Figure 3 (b)Δx b1 >Δx' b1 And Δx b2 >>Δx' b2 , the opposite characteristics are weaker. d >>Δx c >Δx a ,but Figure 3 The opposite characteristics of (a), (c) and (d) are the same. Therefore, taking the historical power of each wind farm as the object, the connection degree between two wind farms in the wind power cluster is obtained by formulas (2)-(4). Then, the adjacency matrix A of the improved set pair analysis method is calculated using formulas (5) and (6): s .

[0197] Step 2: Use formulas (7)-(16) to normalize the power data of each wind farm in the wind power cluster with the installed capacity as the denominator to obtain the synchronization rate sequence, and then perform synchronization rate judgment in units of days to obtain the synchronization rate adjacency matrix A: t .

[0198] Step 3: With the simultaneous rate adjacency matrix A t The calculation process is similar to that of , replacing formulas (7)-(9) in formulas (7)-(16) with formulas (17)-(20) to obtain the variance adjacency matrix A v Then use formula (21) and formula (22) to obtain s , A t and A v Get the convergence effect adjacency matrix A a .

[0199] Step 4: Use Pearson correlation coefficient to construct adjacency matrix A rThe process is similar to that of the simultaneous rate adjacency matrix. Replace formula (9) and formula (12) in formula (7)-(16) with formula (23) and (24). Taking the power of each wind farm in the wind power cluster as input, the correlation coefficient adjacency matrix A can be obtained. r .

[0200] Step 5: Through the Granger causality test, the Granger test values ​​between the daily power values ​​of all wind farms in the wind power cluster are calculated to obtain a causal matrix similar to the form of formula (7) in step S20 above. Then, the causal adjacency matrix A is calculated according to the process of formulas (10)-(16): g , where the discriminant formula (12) needs to be replaced by the causal discriminant formula (25). Finally, the correlation coefficient adjacency matrix A is used r and the causal adjacency matrix A g , the causal temporal adjacency matrix A is obtained by taking the intersection of formula (21) and formula (22) b .

[0201] Step 6: Use the convergent effect adjacency matrix A obtained under the training set data a and the causal temporal adjacency matrix A b , input the test set NWP data for the first prediction, and obtain the prediction error value sequence. Then calculate the error characteristic adjacency matrix A through formulas (7)-(16): e , where formula (9) is replaced by the error evaluation characteristic coefficient formula (26).

[0202] Step 7: Combine the NWP features of each wind farm in the wind farm cluster under the test set and the three adjacency matrices A a , A b and A e Input the graph convolutional neural network based on the multi-channel attention mechanism, the structure is as follows Figure 4 , the final power prediction value is obtained, and the prediction results are evaluated using formulas (33)-(35).

[0203] Compared with the existing wind power cluster prediction strategy, the method of the present invention is more effective. Figure 5 , the prediction error is shown in Table 2. The RMSE and MAE of the method proposed in this paper are the lowest among all clusters, R 2is the highest. The remaining cluster prediction methods have different performances in each wind power cluster. However, from the average prediction error of the three clusters, the prediction effect of statistical upscaling is the worst, and its RMSE is the only one of all prediction methods that exceeds 9%. This is because in a wind power cluster, it is usually difficult to accurately represent the power trend of the entire cluster by selecting a few representative wind farms. However, since the number of wind farms in cluster A is small and the distribution is concentrated, the prediction effect of statistical upscaling in cluster A is better. The overall method ignores a lot of details within the cluster. Although the prediction error is reduced compared with statistical upscaling, its improvement in prediction is very limited. The cumulative method does not consider the correlation between wind farms. The introduction of cluster division solves the disadvantages of the overall method and the cumulative method to a certain extent, so it has a better prediction effect. However, cluster division does not take into account the correlation between clusters, and the classification method is not stable. The prediction effect of cluster division in cluster A is far better than the overall method and the cumulative method, but it does not have much advantage in clusters B and C. The graph convolution model takes the entire cluster as input and uses edges to represent the correlation between wind farms. This form truly takes into account the overall and local advantages and can achieve better prediction results. In clusters A, B, and C, the RMSE of this method is reduced by 1.55%, 1%, and 3.4% on average compared to other prediction methods. Because cluster C has the largest number of wind farms and the widest distribution, the improvement effect of this method is most obvious in cluster C.

[0204] Table 2 Prediction errors of different clusters under each method

[0205]

[0206] Figure 5 (a) The high output samples in the second half of cluster A and Figure 5 In (c), at the peak output of cluster C, the prediction curves of other methods are mostly too high, but the prediction curve of the method proposed in this paper can better track the change trend of the actual power curve. Therefore, the prediction error of the method proposed in this paper has very few positive errors (the error in which the predicted value is greater than the actual value). Figure 6 (a) Cluster A and Figure 6 (c) In cluster C, the prediction error of the proposed method accounts for a very small proportion of positive errors exceeding 5.5%, and is close to 0 in positive errors exceeding 20%. At the same time, the proportion of low errors is high, which shows that the power supply guarantee capability of the proposed method is stronger than other methods. However, it is worth mentioning that the proposed method has a higher proportion of negative errors (errors where the predicted value is less than the actual value), but this does not affect the power supply guarantee capability of the proposed method. The distribution law of prediction errors of the proposed method in cluster B is different from that in clusters A and B. Figure 6(b) shows that the proposed method has a certain proportion in each error range, but the proportion of small errors is higher, so the overall error is still the smallest. Figure 5 In (b), it is difficult to see the obvious difference between the prediction curve of the proposed method and the prediction curve of other methods. However, in general, the proposed method has achieved good prediction results in all three clusters, with the most obvious improvement in cluster C and a slight improvement in cluster B.

[0207] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention should be defined by the claims.

Claims

1. A method for predicting wind power cluster power day-ahead, characterized in that: The method comprises: Based on the wind power sequence, construct the adjacency matrix A s , simultaneous rate adjacency matrix A t , variance adjacency matrix A v ; According to the adjacency matrix A s , simultaneous rate adjacency matrix A t and the variance adjacency matrix A v , construct the convergence effect adjacency matrix A a ; Use Pearson's correlation coefficient to construct the adjacency matrix A r , and construct the causal adjacency matrix A based on Granger causality test g ; Wherein, the adjacency matrix A r and the causal adjacency matrix A g Used to characterize the relationship between wind farms in close proximity; Based on the adjacency matrix A r and the Granger causal adjacency matrix A g , construct the temporal causal adjacency matrix A b ; Construct the error characteristic adjacency matrix A e ; Wherein, the error characteristic adjacency matrix A e Used to characterize the errors between wind farms; A graph convolutional neural network model based on the channel attention mechanism is constructed. In the graph convolutional neural network model, multiple graph convolution modules are designed to input adjacency matrices of different attributes respectively to form multiple channels. Each channel is used to extract different characteristic laws in the wind power cluster to obtain multiple features. The function definition is as follows: In the formula, The l+1th layer features representing the convergence effect, The l+1th layer features representing temporal causality, represents the l+1th layer feature of the error, f represents the graph convolution function, The l-th layer features representing the convergence effect, The l-th layer features representing temporal causality, The l-th layer features representing the error; On the basis of multiple channels, the attention mechanism is introduced to learn the weight of each feature for prediction importance, and finally combine multiple characteristics to obtain the day-ahead prediction result of wind power cluster power. The function is defined as follows: In the formula, Softmax(·) represents the normalization function, concat(·) represents the concatenation function, Output represents the day-ahead wind power cluster power forecast result, W a represents the weight of the convergence effect feature, W b represents the weight of temporal causal features, W e represents the weight of the error feature, H represents the fusion feature, represents the learnable parameter of the l+1th layer convergence effect, represents the learnable parameters of the l+1th layer of temporal causality, Represents the learnable parameters of the error features of the l+1th layer.

2. The method according to claim 1, characterized in that The adjacency matrix A is constructed as follows s : The following formula is used to traverse and calculate the connection degree between each wind farm in the wind power cluster to obtain the connection degree matrix: Where, μ(A,B) is the connection matrix between wind power sequence A and wind power sequence B; {Δx i ,Δx i+1 }、{Δx i ',Δx i ' +1 } respectively represent the adjacent difference values ​​of the i-th identical characteristic element of wind power sequence A and wind power sequence B; {Δx j ,Δx j+1 }、{Δx' j ,Δx' j+1 } respectively represent the adjacent difference values ​​of the jth opposite characteristic elements of wind power sequence A and wind power sequence B; I is the difference coefficient; J is the opposition identification coefficient; N is the total number of features in the set; α i is the characteristic strength of the i-th identical set pair; β j is the characteristic strength of the jth opposite set pair; S' is the total strength of the same set pair; F' is the total strength of uncertainty set pairs; P' is the total strength of opposite set pairs; In order from small to large, select the first m elements in the connection matrix, where m is the average number of connections of the simultaneous rate and variance, retain m element values, and set the rest to zero to form a matrix M; Use matrix M to construct adjacency matrix A based on set pair analysis s , the calculation process is as follows: In the formula, μ ij is the connection degree between the i-th and j-th wind farms; ∞ means that there is no connection between the two wind farms; τ is a constant used to avoid the edge weight being 0; S ij Represents the edge weights established by set pair analysis.

3. The method according to claim 2, characterized in that The simultaneous rate adjacency matrix A is constructed by the following method t : Traverse and calculate the daily synchronization rate of every two wind farms in the wind power cluster to form a matrix T k×i×j , the formula is as follows: In the formula, represents the power sequence of the i-th wind farm on the k-th day, represents the power sequence of the jth wind farm on the kth day, Cap i Cap represents the operating capacity of the i-th wind farm, j represents the operating capacity of the j-th wind farm; represents the simultaneous rate matrix of the wind power cluster on the kth day, represents the simultaneity rate of the sum of the power of wind farm No. i and No. j on day k; For the matrix T k×i×j Each sub-matrix in is discriminated to obtain the discriminant matrix of each day, and the discriminant matrix of each day is summed up to obtain the simultaneous rate discriminant matrix H i×j , the formula is as follows: In the formula, represents the discriminant matrix of the kth day, represents the judgment result of the i-th row and j-th column on the k-th day, h ij Represents the sum of the judgment results of the i-th row and j-th column of all days; represents the simultaneity rate of wind farm No. j on day k; Using the simultaneous rate discriminant matrix to perform adjacency matrix A of directed weighted graph t The specific formula for constructing is as follows: Where W ij represents the weight of wind farm No. i pointing to wind farm No. j; t ij Represents the edge weight established by the simultaneous rate.

4. The method according to claim 3, characterized in that The variance adjacency matrix A is constructed as follows v : Construct a variance matrix and construct a variance adjacency matrix A based on the variance matrix v ; Among them, the construction formula of the variance matrix is ​​as follows: Where V k×i×j represents the variance matrix, represents the normalized sequence of the sum of wind farm No. i and wind farm No. j, represents the sequence mean, n is the number of elements in the wind power sequence, w is the sequence number of the wind power sequence, represents the variance matrix of the wind power cluster on the kth day, It represents the variance of the sum of the power of wind farm No. i and No. j on the kth day.

5. The method according to claim 4, characterized in that According to the adjacency matrix A s , simultaneous rate adjacency matrix A t and the variance adjacency matrix A v , the convergence effect adjacency matrix A is constructed by the following formula a : In the formula, a ij represents the edge weight established by the convergence effect, s ij represents the edge weight constructed by the improved set pair analysis, t ij represents the edge weight constructed by the simultaneous rate, v ij Denotes the edge weights of the variance construction.

6. The method according to claim 1, characterized in that Use Pearson correlation coefficient to construct the adjacency matrix A r ,include: Traverse and calculate the correlation coefficient of every two wind farms in the wind power cluster every day to form a matrix P k×i×j , the formula is as follows: In the formula, represents the correlation coefficient matrix of the wind power cluster on the kth day; represents the correlation coefficient between the power of wind farm No. i and No. j on day k; denote the sequence of the kth day for the i-th and j-th wind farms, respectively; They represent the mean values ​​of the i-th and j-th wind farms on the k-th day respectively; For the matrix T k×i×j Each sub-matrix in is discriminated to obtain the discriminant matrix of each day, and the discriminant matrix of each day is summed up to obtain the simultaneous rate discriminant matrix H i×j , the formula is as follows: In the formula, represents the discriminant matrix of the kth day, represents the judgment result of the i-th row and j-th column on the k-th day, h ij Represents the sum of the judgment results of the i-th row and j-th column of all days; Using the simultaneous rate discriminant matrix to perform adjacency matrix A of directed weighted graph r 's construction.

7. The method according to claim 1, characterized in that Constructing the causal adjacency matrix A based on Granger causality check g ,include: The Granger test is used to traverse and test whether the daily power series of each wind farm have Granger causality, so as to obtain the discriminant matrix, and the causal adjacency matrix A is constructed based on the discriminant matrix. g ; Among them, the calculation formula for using the Granger test to traverse and test whether the daily power series of two wind farms have Granger causality is as follows: In the formula, represents the Granger causality of the i-th wind farm to the j-th wind farm in the k-th day; represents the Granger test value of the i-th wind farm on the j-th wind farm in the k-th day; Fa represents the Granger test threshold.

8. The method according to claim 3, characterized in that Construct the error characteristic adjacency matrix A e ,include: Based on the error smoothing evaluation coefficient, the error characteristic adjacency matrix A e ; Wherein, the error smoothing evaluation coefficient is expressed as: In the formula, represents the normalized value of the sum of the errors of wind farm No. i and No. j on the kth day; They represent the k-th day prediction error sequences of the i-th wind farm and the j-th wind farm respectively.

9. The method according to any one of claims 1 to 8, characterized in that The method further comprises: Select the normalized root mean square error NRMSE, the normalized mean absolute error NMAE and the coefficient of determination R 2 As the evaluation index of the prediction model performance, the calculation formulas are: Where n is the predicted power sequence length, y i,pre and i,ture are the predicted and actual values ​​of power, respectively. It represents the average value of actual power value, and Cap is the installed capacity.

Images