Optimization method of Informer wind power prediction model based on health assessment

Abstract

The present invention discloses an optimization method for an Informer wind power prediction model based on health assessment, wherein the optimization method comprises at least two combinations of the following contents: 1) encoder optimization: improving the input x of the i-th layer (i‑1) The weights of strongly correlated features are optimized to optimize the input so that each encoding layer can directly retrieve the strongly correlated weights; 2) Decoder optimization: Improve the input c of the Lth layer L‑1 The weight of the strongly correlated features is used to optimize the input. In each decoding layer, the product of the output x and the health factor h is used as the input of the next layer to improve the decoding accuracy; 3) Embedding vector optimization: Improve the initial input x of the module (0) The weights of the strongly correlated features are used to optimize the embedding vector, and the optimized input is used as the initial input of the module; 4) Prediction process optimization: According to the initial input x of the module (0) The health matrix is ​​calculated with the optimized decoding output, and then the health assessment is performed on the decoding output to obtain the optimized decoding output.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power technology, and in particular to an optimization method for an Informer wind power prediction model based on health assessment. Background Art

[0002] The volatility of wind energy causes intermittent wind power, which may affect the operation of the power grid after being connected to the grid. In order to ensure the safety of the power grid, it is necessary to predict wind power so that it can be adjusted in time.

[0003] The early prediction method based on numerical weather prediction has been replaced by neural networks. Although neural networks have high prediction accuracy, they have slow prediction speed, complex structure, and difficulty in later maintenance. Finding new prediction models has become a research hotspot in recent years. Informer can replace traditional neural networks for power prediction due to its fast prediction speed and lightweight model. However, Informer has low prediction accuracy, and its prediction speed cannot meet actual needs, and its prediction performance is obviously insufficient.

[0004] In view of the shortcomings of Informer's prediction performance, researchers have optimized it from many aspects.

[0005] etc. optimized the encoding process, strengthened the sequence analysis ability of Informer, and improved the prediction accuracy; Zhao Dongyu, Feng Xia, Yang Z

[0006] etc. introduced adversarial mechanisms to enhance Informer's anti-interference ability and improve prediction accuracy; Yu Qiongfang, Liang Yingjie, Zhu Q, etc. improved sampling methods to enhance Informer's multivariate sampling ability and prediction speed. At present, most optimization methods cannot improve Informer's prediction accuracy and speed at the same time, the optimization effect is not ideal, and the complexity of Informer is increased, which affects its practical application.

[0007] Informer is a prediction model based on probabilistic sparse self-attention, proposed by Zhou Haoyi et al. in 2021. Unlike traditional neural networks, Informer adopts an encoder-decoder separation structure. During the prediction process, the encoded output is sent to the decoder for processing, the decoded output is obtained, and the prediction result is output through the fully connected layer.

[0008] Informer prediction process is as follows Figure 1 As shown. In Informer, there are 3 key points:

[0009] (1) Self-attention Distilling Operation: In the Encoder module, the depth of each encoding layer decreases, and the output length decreases accordingly, so as to extract features. At the end of the module, all features are connected as the input of the Decoder module.

[0010] (2) Generative Style Decoder: In the Decoder module, a new generative decoder is used, which only requires one forward derivation process to obtain the prediction result of the specified length, effectively avoiding the diffusion of cumulative errors.

[0011] (3) ProbSparse Self-attention Mechanism: Based on the conclusion that high-scoring dot products mainly contribute, Informer pays extra attention to strongly correlated features. During the self-attention process, only the strongly correlated features are calculated to reduce the prediction cost.

[0012] The existing Informer prediction model, while ensuring a certain degree of accuracy and speed, still has the following shortcomings:

[0013] 1. Frequent calculation: In the Encoder module, in order to increase the attention of strongly correlated features, self-attention calculations need to be performed frequently, which reduces the prediction speed;

[0014] 2. Decoding defects: In the Decoder module, sequence placeholders are used for auxiliary prediction, and the placeholders and feature data have the same weight. The unreasonable weight distribution leads to a decrease in prediction accuracy.

[0015] 3. Computational complexity: Informer needs to introduce additional modules or increase the feature dimension to complete the self-attention mechanism, which increases the computational complexity and reduces the advantage of lightweight model.

[0016] Health assessment is a weight optimization algorithm proposed by Lars Landberg in 2011. It converts the two-dimensional probability density function into a weighted health matrix based on the characteristics of the statistical probability distribution of the model output.

[0017] The health assessment process is as follows: Figure 2 As shown, specifically:

[0018] (1) Data collection: Collect wind turbine data such as wind speed, temperature, pressure, speed, and power through sensors and SCADA systems;

[0019] (2) Data cleaning: Clean the missing values, outliers, and transmission noise in the data to improve data reliability;

[0020] (3) Feature screening: From the cleaned data, feature data is screened out, and based on the dimension of the feature data, the dimension of the health matrix is ​​obtained;

[0021] (4) Status classification: Use feature data to calculate feature weights, and classify wind turbine status into normal, minor fault, and major fault based on the weight values;

[0022] The feature weights are calculated as follows:

[0023]

[0024] Where i is the feature number, from 1 to n, μ is the mean vector, and Σ is the covariance matrix.

[0025] (5) Matrix generation: Arrange the feature weights of the same state into a health matrix. Each matrix element represents the weight coefficient of the corresponding feature in that state.

[0026] Arrange the feature weights of formula (1) in order to obtain the weight matrix Q n×1 .

[0027] Using factor analysis on the characteristic data, we get the correlation matrix R 1×n , then multiply Q by R to get the health moment A, that is:

[0028] A n×n =Q n×1 R 1×n (2)

[0029] Where n is the dimension of the health matrix, which is usually 6 to 12. When n is greater than 12, dimensionality reduction is required.

[0030] The principles of health assessment are as follows:

[0031] Assume that the initial input z of the prediction model (0) , a coefficient matrix with the main diagonal elements being 1 and the remaining elements being between (0,1) D 0 is a diagonally dominant symmetric matrix, and D 0 The row and column labels and label order should be consistent.

[0032] In the prediction model, the i-th layer input z (i-1) The coefficient matrix D i , which can be regarded as i D 0 Multiplying together gives:

[0033] D i =[D0 ] i =D 0 ×D 0 ×…×D 0 (3)

[0034] z (0) Multiplying with the health matrix A, we get Right now but The coefficient matrix of The following calculations yield:

[0035]

[0036] In formula (4), A reduces D 0 The weak correlation weight is to adjust the corresponding main element to be less than 1 on the basis of maintaining diagonal dominance and symmetry. Instead of D 0 , then each time a forward calculation is performed, the weak correlation weight decreases once, thereby reducing the impact of weakly correlated features and improving the model prediction performance.

[0037] Since the prediction processes of different models are different, the impact of z and D on the prediction is also different. After optimization, the prediction accuracy or speed of the model may be improved.

[0038] The existing health assessment algorithm has two defects:

[0039] 1. High processing cost: When screening features, multi-dimensional wind turbine data is required, which increases the cost of data collection, processing and analysis;

[0040] 2. High complexity: In order to improve the reliability of the evaluation, a complex feature selection module is introduced, which increases the complexity of the evaluation. Summary of the invention

[0041] In view of the above-mentioned technical problems to be solved, the present invention provides an optimization method for the Informer wind power prediction model based on health assessment, which can monitor in real time, reduce rescue risks, shorten rescue time, improve rescue success rate, and ensure that rescue work is carried out in an orderly manner.

[0042] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0043] An optimization method for an Informer wind power prediction model based on health assessment, the optimization method comprising at least two combinations of the following:

[0044] 1) Encoder optimization: improve the input x of the i-th layer (i-1) The input is optimized so that each encoding layer directly retrieves the strongly correlated weights.

[0045] 2) Decoder optimization: improve the input c of the Lth layer L-1 The weights of strongly correlated features are used to optimize the input. In each decoding layer, the product of the output x and the health factor h is used as the input of the next layer to improve the decoding accuracy.

[0046] 3) Embedding vector optimization: improve the initial input x of the module (0) The weights of the strongly correlated features are used to optimize the embedding vector, and the optimized input is used as the initial input of the module;

[0047] 4) Prediction process optimization: Based on the initial input x of the module (0) The health matrix is ​​calculated with the optimized decoding output, and then the health assessment is performed on the decoding output to obtain the optimized decoding output.

[0048] As a further improvement of the above technical solution:

[0049] Preferably, the encoder optimization is specifically:

[0050] Assume that the Encoder module consists of multiple encoding layers and convolutional layers, and the input of the i-th layer is x (i-1) , the output is x (i) :

[0051]

[0052] In formula (5), Conv represents the convolution calculation of the convolution layer; Attention represents the probabilistic sparse self-attention function.

[0053]

[0054] In formula (6), softmax represents the normalization function; Indicates x (i-1) The probability sparse matrix obtained after sampling; represents variance; K, K T , V represents the adjustment matrix of the sampling key value; p represents the function output;

[0055] Raise x (i-1) In , the weights of strongly correlated features are obtained as:

[0056]

[0057] In formula (7), Replace x (i-1) ,Let each encoding layer,directly retrieve the strongly correlated weights.

[0058] Preferably, the decoder optimization is specifically as follows:

[0059] Assume that the Decoder module consists of L decoding layers. In each decoding layer, the input c L-1 And the feature data y is processed by multiple decoding sub-layers, that is,

[0060] c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (···Decoder1(y,c0)···)) (10)

[0061] In the formula, Decoder i (·) represents the decoding calculation of the i-th layer, c0 is the first layer input of the Decoder module, and c i-1 is the input of the i-th layer, and y represents the feature data at the current moment;

[0062] Raise c L-1 In the equation, the weight of the strongly correlated feature is

[0063] c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (···(Decoder1(y,c0)·h)···)) (11)

[0064] In the formula, h is the health factor, which is composed of the health matrix A n×n Take n=1 and perform compression transformation to obtain; in each decoding layer, the output x is multiplied by the health factor h and used as the input of the next layer.

[0065] Preferably, the embedding vector optimization is specifically:

[0066] The initial input x of the Encoder module (0) , improve x (0) In the equation, the weight of the strongly correlated feature is

[0067]

[0068] In the formula, x (0) is the unoptimized embedding vector, is the optimized embedding vector; after optimization, Will replace x (0) , as the initial input of the Encoder module.

[0069] Preferably, the prediction process is optimized, specifically:

[0070] Embedding vector x (0)After continuous processing by the Encoder module and the Decoder module, the decoded output c is obtained. L Then, use x (0) and c L Calculate the health matrix Right now

[0071]

[0072] Health assessment for c L To proceed, that is:

[0073]

[0074] in, is the optimized decoding output; Calculate the required weight matrix By x (0) The correlation matrix is ​​obtained by the correlation coefficient method. By c L Obtained by correlation coefficient method.

[0075] The optimization method of the Informer wind power prediction model based on health assessment provided by the present invention has the following advantages compared with the prior art:

[0076] (1) The optimization method of the Informer wind power prediction model based on health assessment of the present invention adopts health assessment to optimize the Informer model and improve the prediction accuracy and speed. First, a health assessment is performed based on the operating data to obtain the health matrix of the wind turbine; then, the encoding method is optimized through the health matrix to improve the prediction speed of the Informer; then, the decoding method, embedding vector and prediction process are improved using the health matrix to improve the prediction accuracy of the Informer.

[0077] (2) The optimization method of the Informer wind power prediction model based on health assessment of the present invention, the Informer optimization model designed based on health assessment, can accelerate the convergence of the Informer and improve the prediction accuracy and speed of the model. The Encoder combined with the optimized Informer model is a high-quality wind power prediction model. Compared with the unoptimized model, the Informer optimization model based on health assessment has lower data cost and better optimization effect when predicting wind power. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is the Informer prediction flow chart in the prior art.

[0079] Figure 2 The figure is a flowchart of health assessment in the prior art.

[0080] Figure 3 Embedded vector structure diagram when implementing the application of the present invention.

[0081] Figure 4 Before and after optimization for the application and implementation of the present invention Comparison chart.

[0082] Figure 5 This is the predicted effect diagram of KC Unit 5 when verifying the present invention.

[0083] Figure 6 This is the predicted effect diagram of KC Unit 33 during the verification of the present invention.

[0084] Figure 7 This is the predicted effect diagram of Longyuan Unit 1 during the verification of the present invention.

[0085] Figure 8 This is the predicted effect diagram of Longyuan Unit 12 during the verification of the present invention.

[0086] Fig. 9 This is the predicted effect diagram of the 5th 1-36 unit when verifying the present invention.

[0087] Fig.10 This is the predicted effect diagram of Unit 5, 2-30 when verifying the present invention.

[0088] Fig.11 This is the predicted effect diagram of GEFC Unit 5 when verifying the present invention.

[0089] Fig.12 This is the predicted effect diagram of GEFC Unit 6 when verifying the present invention. DETAILED DESCRIPTION

[0090] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0091] like Figures 1 to 4 As shown, the optimization method of the Informer wind power prediction model based on health assessment of the present invention includes the following contents:

[0092] 1) Encoder optimization

[0093] Assume that the Encoder module consists of multiple encoding layers and convolutional layers, and the input of the i-th layer is x (i-1) , the output is x (i) In each encoding layer, x (i-1) We advance forward through Multi-Head Self-Attention and then use a convolutional layer to reduce the output dimension to obtain x (i),Right now

[0094]

[0095] In formula (5), Conv represents the convolution calculation of the convolution layer; Attention represents the probabilistic sparse self-attention function, that is,

[0096]

[0097] In formula (6), softmax represents the normalization function; Indicates x (i-1) The probability sparse matrix obtained after sampling; represents variance; K, K T , V represents the adjustment matrix of the sampling key value; p represents the function output.

[0098] In x (i-1) In the example, all features have equal weights, so we need to traverse x (i-1) After that, the strongly correlated weights are sampled. Combined with the health matrix A, the x (i-1) The weights of medium and strong correlated features, i.e.

[0099]

[0100] In formula (7), Replace x (i-1) , which allows each encoding layer to directly retrieve strongly correlated weights, accelerating , thereby improving the encoding speed of the Encoder module and the prediction speed of the Informer.

[0101] Demonstration of the optimization effect of the Encoder module:

[0102] In formula (6), the matrix The i-th query element q i , obtained through the following operations:

[0103]

[0104] In the formula, q i , k i , v i represent K, the i-th row in V;

[0105] p(k j |q i )=k(q i ,k j ) / ∑ l k(q i ,k l ), k(q i,k j ) represents an asymmetric exponential kernel Represents variance.

[0106] In equations (6) and (8), the sparse metric M(q i , K) Top-u query, that is:

[0107]

[0108] Where, L K Represents the input sequence length of matrix K.

[0109] In sparse self-attention, the length of query and key is equal, i.e., L Q =L K = L, the computational complexity is O(L ln L). However, x (i-1) All features are weighted equally. In order to increase the weight of strongly correlated features, M(q i ,K) query requires secondary dot product calculation, which reduces the prediction speed of Informer.

[0110] From equations (3) and (4), we know that in equation (7), the optimized The weight of strongly correlated features is increased. As the initial input of the Encoder module, we can let M(q i ,K) query, only one dot product calculation is required, and the computational complexity of sparse self-attention is reduced to This improves the encoding speed of the Encoder module and the prediction speed of the Informer module.

[0111] 2) Decoder optimization

[0112] Assume that the Decoder module consists of L decoding layers. In each decoding layer, the input c L-1 And the feature data y is processed by multiple decoding sub-layers, that is,

[0113] c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (···Decoder1(y,c0)···))(10)

[0114] In the formula, Decoder i (·) represents the decoding calculation of the i-th layer, c0 is the first layer input of the Decoder module, and c i-1is the input of the i-th layer, and y represents the feature data at the current moment.

[0115] Due to c L-1 All feature elements of have equal weights. Combined with the promotion c L-1 In the equation, the weight of the strongly correlated feature is

[0116] c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (···(Decoder1(y,c0)·h)···)) (11)

[0117] In the formula, h is the health factor, which is composed of the health matrix A n×n Take n = 1 and perform compression transformation to obtain. In each decoding layer, the output x is multiplied by the health factor h and used as the input of the next layer, thereby improving the decoding accuracy.

[0118] Since the decoding accuracy of the Decoder module is improved, the prediction accuracy of the Informer is improved after optimization.

[0119] Demonstration of the optimization effect of the Decoder module:

[0120] The first layer input of the Decoder module can be expressed as

[0121]

[0122] In the formula, is the final output of the Encoder module. is a sequence placeholder with a scalar value of 0 (corresponding to Figure 2 The auxiliary input is all 0 part), L token Indicates the length of the sequence finally output by the Encoder, L y Indicates the length of the placeholder sequence, d model Represents the feature dimension, and Concat represents vector concatenation. By setting the placeholder and setting the masked dot product to -∞, Informer applies the masked multi-head attention to the probabilistic sparse self-attention of the Decoder module.

[0123] Although the application of masked multi-head attention prevents In , the features of each position focus on the autoregression phenomenon of the features at the next moment, but and The weights of are equal, so that each time the forward calculation of formula (10) is performed, The accumulated error will diffuse once, reducing the prediction accuracy of Informer.

[0124] The optimization scheme of formula (11) is equivalent to improving The weight of The weight of, from formula (3) and formula (4), we can know that each time the forward calculation of formula (11) is performed, The weight of will decrease once, thereby improving the prediction accuracy of Informer.

[0125] 3) Embedding vector optimization

[0126] The embedding vector refers to the initial input x of the Encoder module (0) Since the Encoder module reduces the feature dimension layer by layer and extracts strongly related features, now combined with the improvement of x (0) In the equation, the weight of the strongly correlated feature is

[0127]

[0128] In the formula, x (0) is the unoptimized embedding vector, is the optimized embedding vector. After optimization, Will replace x (0) , as the initial input of the Encoder module.

[0129] Because the weight of strongly correlated features in the embedding vector is increased, the prediction accuracy of Informer is improved after optimization.

[0130] Demonstration of the optimization effect of the embedding vector module:

[0131] Assume that the tth sequence input is X t , and X t The global timestamp type is p and the feature dimension is d model , then by embedding at a fixed position, the embedding vector of the local context is retained, which is obtained by the following two equations:

[0132]

[0133]

[0134] In the formula, j∈{1,…,[d model / 2]}. POS represents a learnable embedding SE using a global timestamp, which has a limited vector length (up to 60, in minutes). To align the dimensions, the Informer model uses a 1D convolutional filter with kernel width = 3 and stride = 1, and converts the scalar context Projection to d model Dimensional vector In. That is, the embedding vector It can be expressed as follows:

[0135]

[0136] Where i∈{1,…,L x}, α is the amplitude factor between balancing scalar projection and local / global embedding. If the sequence has been normalized, α=1.

[0137] The embedding vector structure is as follows Figure 3 As shown. Since the trigonometric functions sin and cos have their own sampling weights, equations (14) and (15) give Its weight distribution will change compared with the original data, that is, Figure 3 In , the weight distribution of local timestamp and global timestamp is inconsistent, which causes system error. From equations (3), (4), (8), (9) and (12), we know that The systematic error will gradually expand in the subsequent encoding-decoding process, affecting the prediction accuracy of Informer.

[0138] Formula (13) combines the deviation of feature distribution to optimize the embedding vector. From Formula (3) and Formula (4), it can be seen that after optimization, the system error caused by the change of feature distribution will gradually decrease in the subsequent encoding-decoding process, thereby improving the prediction accuracy of Informer.

[0139] 4) Prediction process optimization

[0140] Depend on Figure 2 From equations (3) to (12), we know that the embedding vector x (0) , after continuous processing by the Encoder module and the Decoder module, the decoded output c is obtained L After that, the distribution of strongly correlated features may change. Now, combined with the change of health assessment data source, using x (0) and c L Calculate the health matrix Right now

[0141]

[0142] And the health assessment is changed to c L To proceed, that is:

[0143]

[0144] in, This is the optimized decoding output. Calculate the required weight matrix By x (0) The correlation matrix is ​​obtained by the correlation coefficient method. By cL Obtained by correlation coefficient method.

[0145] After optimization, Informer's prediction accuracy increases due to the combination of changes in feature distribution.

[0146] Demonstration of optimization effect of prediction process:

[0147] Assume x (0) and c L The feature order of is the same, the feature dimension is n, and the strongly correlated features are continuously distributed. And, x (0) Among them, the features from e to f are strongly correlated features, and c L In , the g to h features are strongly correlated features, then:

[0148] 1≤e≤g <f≤h≤n (19)

[0149] From equations (3) to (16), it can be seen that the prediction process does not involve the increase or decrease of the number of strongly correlated features. Therefore, the following constraints exist:

[0150] fe=hg=k, k is a positive integer (20)

[0151] For the embedding vector x (0) , only retain the main diagonal elements, and the simplified coefficient matrix with all other elements being 0 is

[0152] D x(0) =diag[p1…p e-1 1…1 p f+1 …p n ] (twenty one)

[0153] Similarly, for the decoded output c L , its simplified coefficient matrix is

[0154]

[0155] and In , all elements that are not 1 have values ​​between (0,1).

[0156] From formula (16) and Figure 3 Know that x (0) Contains global timestamp information. Through reasonable sampling, it can be considered that x (0) The strongly correlated feature distribution of is the same as the original data. From equations (3) to (7), (10) and (11), we know that after multiple encoding and decoding operations, In , all elements that are not 1 tend to 0. However, the simplified coefficient matrix corresponding to the original data is The simplified coefficient matrix corresponding to the prediction result is When e < g in Equation (19), that is, when the feature distributions are different, the prediction accuracy of Informer decreases.

[0157] By adopting the optimization schemes of Equations (17) and (18), it is possible to and for the feature offset situation, increase the values of the elements from the e-th to the (g - 1)-th in

[0158] When k = 2, before and after optimization As Figure 4 shown. Figure 4 In

[0159] 5) Combined optimization

[0160] Since the optimization of the Encoder module mainly focuses on speed improvement, and the optimization of the Decoder module, embedding vectors, and prediction process mainly focuses on accuracy improvement. To avoid over-optimization, the present invention selects 3 combined optimization models, namely Encoder-Decoder, Encoder-embedding vectors, and Encoder-prediction process, for training and testing.

[0161] Optimization verification:

[0162] 1) Error index setting

[0163] Adopt R 2 (Coefficient of determination), MSE (Mean Square Error), and MAE (Mean Absolute Error), a total of 3 error indicators, as shown in Equations (23)-(25):

[0164]

[0165]

[0166]

[0167] In the formula, y i is the true value, is the predicted value, is the true average value. R 2 represents the fitting degree, and the closer the value is to 1, the higher the fitting degree; MSE and MAE represent the dispersion degree, and the closer the value is to 0, the smaller the prediction fluctuation.

[0168] If the optimization is effective, ideally, R 2As the MSE and MAE increase, they will also decrease. However, due to different models, the degree of prediction dispersion is different, and in order to increase the overall fit, reducing the fit of specific points will increase the cost of accident handling. Therefore, when the two indicators increase by more than 10% and the remaining one decreases by no more than 1%, it is also considered to be effective.

[0169] 2) Simulation test

[0170] This verification method uses four datasets: Baidu KDD CUP 2022 dataset, "China Software Cup" - Longyuan Wind Power Track dataset, the dataset of the 5th National Industrial Internet Data Innovation Application Competition, and the Task15 dataset of Global EnergyForecasting Competition 2014. Each dataset has 17,000-21,000 data. A total of 8 models were selected for simulation testing, including unoptimized, Encoder optimized, Decoder optimized, Encoder-Decoder optimized, embedded vector optimized, prediction process optimized, Encoder-embedded vector optimized, and Encoder-prediction process optimized. For each model, simulation tests were performed for 3 rounds of training and 6 rounds of training respectively. The first 80% of the dataset is used as the training set, and the remaining 20% ​​is the test set.

[0171] In order to reduce the impact of dimensional differences, all prediction results are standardized as follows:

[0172]

[0173] In the formula, z0 represents the unstandardized prediction result, z1 represents the standardized prediction result, and z s represents the variance of the prediction results, Represents the average of all prediction results.

[0174] The components of the Informer model are detailed in Table 1.

[0175] Table 1 Details of Informer model components

[0176]

[0177] The prediction results using the KDD CUP dataset are as follows: Figure 5 and Figure 6 The prediction results using the Longyuan wind power data set are shown in Figure 7 and Figure 8 The prediction results of the fifth competition data set are shown in Fig. 9 and Fig.10 The prediction results using the GEFC 2014 dataset are shown in Fig.11 and Fig.12 shown.

[0178] The comparison of prediction results of each data set is shown in Tables 2 to 9.

[0179] Table 2 Prediction indicators of KDD CUP Unit 5

[0180]

[0181]

[0182] Table 3 Prediction indicators of KDD CUP Unit 33

[0183]

[0184] Table 4 Prediction indicators of Longyuan Wind Power Unit 1

[0185]

[0186]

[0187] Table 5 Prediction indicators of Longyuan Wind Power Unit 12

[0188]

[0189]

[0190] Table 6 Prediction indicators of Unit 36 ​​of 5th Wind Farm No. 1

[0191]

[0192] Table 7 Prediction indicators of Unit 30 of 5th Wind Farm No. 2

[0193]

[0194]

[0195] Table 8 Prediction indicators of GEFC Unit 5

[0196]

[0197] Table 9 Prediction indicators of GEFC Unit 6

[0198]

[0199]

[0200] Figure 5-Figure 12In the figure, V1 represents the prediction effect without optimization, V2 represents the prediction effect with encoder optimization, V3 represents the prediction effect with decoder optimization, V4 represents the prediction effect with encoder-decoder optimization, V5 represents the prediction effect with embedding optimization, V6 represents the prediction effect with prediction process optimization, V7 represents the prediction effect with encoder-prediction process optimization, and V8 represents the prediction effect with encoder-embedded vector optimization. All prediction effect graphs are drawn after 6 rounds of training.

[0201] Depend on Figure 5-Figure 12 It is known that, except for the independent optimization of Encoder, the prediction accuracy of the other optimization models has increased compared with the non-optimized model. It can be seen from Tables 2 to 9 that the proposed 7 optimization models have achieved at least one optimization effect of improving the prediction accuracy by 15% or improving the prediction speed by 100%. Moreover, the three Encoder combined with optimization models have simultaneously achieved the optimization effects of improving the prediction accuracy by 15% and improving the prediction speed by 100%. In addition, the selected combined optimization model only needs half the number of training rounds, and its prediction accuracy is close to or exceeds that of the non-optimized model. This shows that the Informer optimization based on health assessment is reasonable and effective. The Encoder combined with optimized Informer model is a high-quality wind power prediction model.

[0202] The above implementation cases are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Although the present invention has been disclosed as above with preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modification, equivalent changes and modifications made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. An optimization method for an Informer wind power prediction model based on health assessment, characterized in that: The specific methods of health assessment are: (1) Data collection: Through sensors and SCADA systems, wind turbine data including wind speed, temperature, pressure, speed, and power are collected; (2) Data cleaning: Clean the missing values, outliers, and transmission noise in the data to improve data reliability; (3) Feature screening: From the cleaned data, feature data is screened out, and based on the dimension of the feature data, the dimension of the health matrix is ​​obtained; (4) Status classification: Use feature data to calculate feature weights, and classify the wind turbine status into normal, minor fault, and major fault based on the weight values; The feature weights are calculated as follows: Where i is the feature number, from 1 to n, μ is the mean vector, Σ is the covariance matrix, and T represents the vector transpose; (5) Matrix generation: Arrange the feature weights of the same state into a health matrix, where each matrix element represents the weight coefficient of the corresponding feature in that state; Arrange the feature weights of formula (1) in order to obtain the weight matrix Q n×1 ; Using factor analysis on the characteristic data, we get the correlation matrix R 1×n , then multiply Q by R to get the health matrix A, that is: A n×n =Q n×1 R 1×n (2) In the formula, n is the dimension of the health matrix, ranging from 6 to 12. When n is greater than 12, dimensionality reduction is required. The optimization method comprises at least two combinations of the following: 1) Encoder optimization: improve the input x of the i-th layer (i-1) The weights of the strongly correlated features are optimized using the health matrix A, so that each encoding layer directly retrieves the strongly correlated weights; 2) Decoder optimization: improve the input c of the Lth layer L-1 The weights of the strongly correlated features are used to optimize the input using the health matrix A. A is first compressed and transformed to obtain the health factor h. In each decoding layer, the product of the output x and the health factor h is used as the input of the next layer to improve the decoding accuracy. 3) Embedding vector optimization: Use the health matrix A to improve the initial input x of the module (0) The weights of the strongly correlated features are used to optimize the embedding vector, and the optimized input is used as the initial input of the module; 4) Prediction process optimization: Based on the initial input x of the module (0) The health matrix is ​​calculated with the optimized decoding output, and then the health assessment is performed on the decoding output to obtain the optimized decoding output.

2. The optimization method of the Informer wind power prediction model based on health assessment according to claim 1 is characterized in that: The encoder optimization is specifically as follows: Assume that the Encoder module consists of multiple encoding layers and convolutional layers, and the input of the i-th layer is x (i-1) , the output is x (i) : x (i) =Conv(Attention(x (i-1) )) (5) In formula (5), Conv represents the convolution calculation of the convolution layer; Attention represents the probabilistic sparse self-attention function, In formula (6), softmax represents the normalization function; Indicates x (i-1) The probability sparse matrix obtained after sampling; represents variance; K, K T , V represents the adjustment matrix of the sampling key value; p represents the function output; Raise x (i-1) In , the weights of strongly correlated features are obtained as: In formula (7), x1 (i-1) Replace x (i-1) ,Let each encoding layer,directly retrieve the strongly correlated weights.

3. The optimization method of the Informer wind power prediction model based on health assessment according to claim 2 is characterized in that: The decoder optimization is specifically as follows: Assume that the Decoder module consists of L decoding layers. In each decoding layer, the input c L-1 And the feature data y is processed by multiple decoding sub-layers, that is, c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (···Decoder1(y,c0)···)) (10) In the formula, Decoder i (·) represents the decoding calculation of the i-th layer, c0 is the first layer input of the Decoder module, and c i-1 is the input of the i-th layer, and y represents the feature data at the current moment; Raise c L-1 In the equation, the weight of the strongly correlated feature is c L =Decoder(y,c L-1 )=Decoder L (Decoder L-1 (···(Decoder1(y,c0)·h)···)) (11) In the formula, h is the health factor, which is composed of the health matrix A n×n Take n=1 and perform compression transformation to obtain; in each decoding layer, the output x is multiplied by the health factor h and used as the input of the next layer.

4. The optimization method of the Informer wind power prediction model based on health assessment according to claim 3 is characterized in that: The embedding vector optimization is specifically as follows: The initial input x of the Encoder module (0) , improve x (0) In the equation, the weight of the strongly correlated feature is In the formula, x (0) is the unoptimized embedding vector, is the optimized embedding vector; after optimization, Will replace x (0) , as the initial input of the Encoder module.

5. The optimization method of the Informer wind power prediction model based on health assessment according to claim 4 is characterized in that: The prediction process optimization is specifically as follows: Embedding vector x (0) After continuous processing by the Encoder module and the Decoder module, the decoded output c is obtained. L Then, use x (0) and c L Calculate the health matrix Right now Health assessment for c L To proceed, that is: in, is the optimized decoding output; Calculate the required weight matrix By x (0) The correlation matrix is ​​obtained by the correlation coefficient method. By c L Obtained by correlation coefficient method.

Images