Wind power prediction method based on optimized neural network and health assessment

Abstract

The present invention discloses a wind power prediction method based on an optimized neural network for health assessment, which relates to the technical field of wind power prediction methods. The prediction method comprises the following steps: step S1, obtaining wind turbine data, performing health assessment and constructing a health matrix; step S2, calculating improved power weights based on the health matrix of step S1, using the matrix approximation hyperparameters as WSA starting values, accelerating the SVR optimization process, and using neural networks to optimize the predicted power. The wind power prediction method based on an optimized neural network for health assessment of the present invention first performs a health assessment based on wind turbine data to obtain a health matrix; then, the network prediction strategy is optimized based on the matrix weight calculation results; then, the matrix approximation hyperparameters are used to accelerate the optimization process; finally, the optimized BP and WSA-SVR neural networks are used to predict wind power, while improving the prediction accuracy and speed.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction methods, and in particular to a health assessment-based optimized neural network wind power prediction method. Background Art

[0002] The volatility of wind energy causes instability in wind power, which in turn affects grid operation after connection. To ensure grid security, wind power forecasting and timely regulation are necessary.

[0003] Internationally, there has been considerable research on wind power prediction technology, which can be broadly divided into two types: physical models and statistical models. Constructing physical models is relatively complex, and because these models are affected by environmental factors, climate factors, and the wind turbine's own status information, it is difficult to construct a reasonable and effective prediction model for wind power, and stable prediction accuracy cannot be guaranteed. Therefore, models based on statistical analysis are more widely used in the wind power prediction process. Early wind power prediction methods often used numerical weather prediction (NWP) and power curves. Due to the low measurement accuracy of NWP and the large transmission delay, the prediction accuracy of this method is not high. The power curve is derived from the wind turbine operating conditions. When environmental changes affect the operation of the wind turbine, causing the actual power to deviate from the rated power, the prediction accuracy is reduced.

[0004] To improve accuracy, neural networks are often used for power prediction. Neural networks use historical data to train network models.

[0005] Wind power forecasting has shown significant performance improvements compared to similar methods, but it also suffers from issues such as slow prediction speed and high data costs. Finding efficient optimization methods to improve neural network prediction performance has become a research hotspot in recent years. Optimization algorithms can enhance network capabilities while simultaneously improving prediction speed and accuracy. Currently, most optimization algorithms can only improve either accuracy or speed, resulting in unsatisfactory results. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: the present invention provides an optimized neural network wind power prediction method based on health assessment. Through health assessment, the wind turbine health matrix is ​​obtained, the power weight calculation method is improved, and the network prediction strategy is adjusted, which can simultaneously improve the prediction accuracy and speed.

[0007] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0008] A wind power prediction method based on an optimized neural network for health assessment includes the following steps:

[0009] Step S1, obtaining wind turbine data, performing health assessment and constructing a health matrix;

[0010] In step S2, based on the health matrix of step S1, the improved power weight is calculated, and the matrix approximation hyperparameter is used as the starting value of WSA to accelerate the SVR optimization process and optimize the predicted power using the neural network.

[0011] As a further improvement of the above technical solution:

[0012] Preferably, the process of constructing the health matrix in step S1 is:

[0013] S1-1, data collection: collect wind turbine data such as wind speed, temperature, pressure, speed, power, etc.

[0014] S1-2, data cleaning: data cleaning is performed to improve data reliability by targeting missing values, outliers, and noise;

[0015] S1-3, feature screening: From the cleaned data, feature data is screened to obtain the approximate hyperparameters of the neural network and the dimensions of the health matrix;

[0016] S1-4, state classification: calculate the power generation efficiency and classify the wind turbine state categories based on the characteristic data;

[0017] S1-5, matrix generation: Arrange the power generation efficiencies of the same state into a matrix according to their values.

[0018] Preferably, in step S1, the health matrix A is:

[0019] A n×n =Q n×1 V 1×n

[0020] Where n is the dimension of the health matrix, Q n×1 is the efficiency matrix, V 1×n is the characteristic correlation matrix of power.

[0021] Preferably, in step S2, the method for obtaining the original predicted power is:

[0022] First, get the weight matrix C n×1 ,Right now:

[0023] C n×1 =A n×n B n×1

[0024] Where B n×1 is the feature matrix;

[0025] Then, the original predicted power P i , expanded to power matrix D 1×n ,Right now:

[0026] D 1×n =

P i P i ··· P i

[0027] Next, combine D and C to obtain the reconstructed prediction power Right now:

[0028]

[0029] Where λ is related to the characteristic data and the fan operating conditions, and its value ranges from 0.01 to 0.1.

[0030] Preferably, the improved power weight calculation formula is:

[0031] Δω=α×δ×x i

[0032] Where Δω represents the change in weight, α is the learning rate, δ is the error of the current node, and x i is the input of the current node.

[0033] Preferably, the method to accelerate the SVR optimization process is:

[0034] Assume that the input of the i-th data is x i , the output is y i , the predicted value is f(x i ), define the error variable e i , which represents the gap between the actual value and the predicted value, that is:

[0035] e i =f(x i )-y i

[0036] While minimizing the sum of squared errors, we limit each e i Does not exceed the given threshold ε, and the constraint condition is:

[0037] minimize:0.5×ω T ×ω+T1×∑e i

[0038] subject:|e i |≤ε,i=1,2,…,n

[0039] Where ω is the power weight, which is the same as the element of the weight matrix C, T1 is the regularization parameter, and ω T is the transpose of ω, and the WSA algorithm is used to optimize T1;

[0040] The SVR prediction value is:

[0041] f(x i )=∑α i ×K(x i ,x)+e i

[0042] Where, α i is the Lagrange multiplier, obtained by solving the Lagrange function equation by SVR; K(x i ,x) represents the inner product kernel function, e i Error variable.

[0043] The optimized neural network wind power prediction method based on health assessment provided by the present invention has the following advantages over the existing technology:

[0044] The present invention's health assessment-based optimized neural network wind power forecasting method improves prediction accuracy and speed by optimizing the neural network through health assessment. First, a health assessment is performed based on wind turbine data to obtain a health matrix. The network prediction strategy is then optimized using the matrix weight calculation results. Matrix hyperparameter approximation is then used to accelerate the optimization process. Finally, the optimized BP and WSA-SVR neural networks are used to predict wind power, improving both prediction accuracy and speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the health assessment process diagram of the present invention.

[0046] Figure 2 It is a schematic diagram of the neural network optimization solution based on the health matrix of the present invention.

[0047] Figure 3 This is a diagram of the unoptimized BP prediction effect in the comparative example of the present invention.

[0048] Figure 4 This is a diagram showing the optimized BP prediction effect in the comparative example of the present invention.

[0049] Figure 5 It is a diagram of the prediction effect of the unoptimized WSA-SVR in the comparative example of the present invention.

[0050] Figure 6 It is a WSA-SVR prediction effect diagram after optimization in the comparative example of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of this specification to clearly and completely describe the technical solutions in the embodiments of this specification. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this invention.

[0052] Figure 1 and Figure 2 An embodiment of the wind power prediction method based on an optimized neural network for health assessment of the present invention is shown, comprising the following steps:

[0053] Step S1: Obtain wind turbine data, perform health assessment and construct a health matrix.

[0054] Wind speed and power generation efficiency are positively correlated. When wind speed decreases, the wind turbine rotates slowly and the power generation efficiency is low. When wind speed increases, the wind turbine rotates quickly and the power generation efficiency is high. Applying the multivariate Gaussian distribution model to calculate the power generation efficiency, the health matrix can be obtained, such as Figure 1 The specific process is as follows:

[0055] S1-1, data collection: collect wind turbine data such as wind speed, temperature, pressure, speed, power, etc.

[0056] S1-2, data cleaning: perform data cleaning for missing values, outliers, and noise to improve data reliability.

[0057] In order to reduce the order error of data of different magnitudes, data normalization is first performed. Considering that some categories of data have negative values, the maximum-minimum normalization algorithm is as follows:

[0058]

[0059] Where x i Represents the current data, x min Indicates the minimum value of the data, x max Indicates the maximum value of the data.

[0060] S1-3, feature screening: Use the cleaned data to perform principal component analysis (PCA), select the top n highest-scoring data in the analysis results as feature data, perform health assessment based on the feature data, and obtain the approximate hyperparameters of the neural network and the dimensions of the health matrix.

[0061] The dimension of the health matrix can be determined by formula (2)

[0062] n=Kind i (2)

[0063] Where Kind iIndicates the number of features. Each feature data contains 1 feature. Kind i The count is increased by 1.

[0064] S1-4, status classification: Calculate the power generation efficiency through characteristic data, and classify the wind turbine status into normal, minor fault, serious fault and other categories according to the efficiency value.

[0065] The power generation efficiency Q is calculated by formula (3): n×1 .

[0066]

[0067] Where i ranges from 1 to n, μ is the mean vector, Σ is the covariance matrix, and i is the feature number.

[0068] S1-5, Matrix Generation: Arrange the power generation efficiencies of the same state into a matrix according to their values. Each matrix element represents the power generation efficiency of the wind turbine in that state.

[0069] Arrange the calculated power generation efficiency in order of characteristics to obtain the efficiency matrix Q n×1 ,Right now:

[0070] Q n×1 =

f(x1) f(x2)··· f(x n )

[0071] Where, f(x i ) represents the power generation efficiency of the i-th feature under the same type of state.

[0072] The factor analysis method is used on the measurement data to obtain the characteristic correlation matrix V of the power 1×n ,Right now:

[0073] V 1×n =

p1 p2 ··· p n

[0074] Where p i represents the correlation coefficient between the i-th feature and power.

[0075] Using Q T After arranging V, multiply Q by V to obtain the health matrix A, that is:

[0076] A n×n =Q n×1 V 1×n (6)

[0077] Where n is the dimension of the health matrix, which is usually 6-12. When n is greater than 12, dimensionality reduction is required.

[0078] There are two shortcomings in the health assessment process:

[0079] (1) High data cost: Calculating power generation efficiency requires a variety of wind turbine data, which increases the cost of data collection, processing, and analysis;

[0080] (2) High complexity: In order to improve the reliability of evaluation, the feature selection module is introduced, which increases the complexity of the matrix and is not conducive to improvement and maintenance.

[0081] In order to improve the evaluation process and reduce the evaluation cost, first, the factor analysis method is used to simplify the feature selection module and reduce the matrix complexity; then, the feature data is dimensional compressed and decoupled to reduce the data processing cost.

[0082] Step S2: Based on the health matrix of step S1, an optimization algorithm is designed to calculate the improved power weight and output the original predicted power using a neural network.

[0083] There are currently two deficiencies in optimizing neural network algorithms: first, the prediction speed is not significantly improved, and the prediction time is long; second, the optimization is insufficient, and the accuracy is not significantly improved.

[0084] Because power weight significantly impacts the prediction accuracy of the Back Propagation Network (BP), this embodiment uses a health matrix to calculate the improved power weight of the BP to improve prediction accuracy. Because the prediction speed of Support Vector Regression (SVR) is significantly affected by parameter optimization, the Whale Swarm Algorithm (WSA) is introduced as an SVR optimization algorithm to accelerate the optimization process and improve prediction speed. Matrix approximation hyperparameters are used as WSA starting values ​​to accelerate the SVR optimization process.

[0085] Based on the health matrix in step S1, the improved power weight is calculated and the predicted power is output using the neural network. This reduces the prediction error and improves the network prediction accuracy and speed without increasing the data processing cost. This includes the following:

[0086] S2-1, power prediction: After accelerating the optimization process by approximating hyperparameters, a neural network is used to perform power prediction and obtain the original predicted power.

[0087] From formula (2) and formula (3), we know that n features correspond to the health matrix A n×n Arrange the current moment data into the feature matrix B according to the order of matrix Q n×1 ,Right now:

[0088] B n×1 =

q1 q2 ··· q n

[0089] Where q i Represents the correlation coefficient between the i-th feature and the power at the time to be predicted.

[0090] First, matrices A and B are multiplied to obtain the weight matrix C n×1 ,Right now:

[0091] C n×1 =A n×n B n×1 (8)

[0092] Then, the original predicted power P i , expanded to power matrix D 1×n All elements in D are original predicted power P i ,Right now:

[0093] D 1×n =

P i P i ··· P i

[0094] Next, combine D and C to obtain the reconstructed prediction power Right now:

[0095]

[0096] Where λ is related to the characteristic data and the fan operating conditions, and its value ranges from 0.01 to 0.1.

[0097] S2-2, design optimization algorithm.

[0098] For BP network, the hidden layer error can be expressed as:

[0099] δ h =f′(x)×∑(ω×δ o ) (11)

[0100] Where, δ h represents the hidden layer error; δ o represents the estimated system error, which is a given value; f′ is the derivative of the activation function, x is the hidden layer input, ω is the power weight, which is the same as the element of the weight matrix C, and Σ represents the summation of the output layer nodes.

[0101] The update calculation of BP power weight is as follows:

[0102] Δω=α×δ×x i (12)

[0103] Where Δω represents the change in weight, α is the learning rate, δ is the error of the current node, and x iis the input of the current node.

[0104] In formula (6), the matrix A is dimensional compressed, A is compressed into a matrix of 1 row and n columns, and converted into a vector form, ω is calculated, and the interference term of ω is reduced, so that ω×δ o The value of is significantly reduced, reducing δ h , reduce prediction error and improve BP prediction accuracy.

[0105] Let the health matrix be x i , by using formula (12), the update calculation amount can be reduced, and the weight convergence can be accelerated to obtain the approximate optimal learning rate and improve the BP prediction speed.

[0106] The optimization process of SVR is as follows:

[0107] Assume that the input of the i-th data is x i , the output is y i , the predicted value is f(x i ). Define the error variable e i , which represents the gap between the actual value and the predicted value. That is:

[0108] e i =f(x i )-y i (13)

[0109] During the optimization process, while minimizing the sum of squared errors, each e is limited i Not exceeding the given threshold ε, we can obtain the constraint conditions composed of equations (14) and (15):

[0110] minimize:0.5×ω T ×ω+T1×∑e i (14)

[0111] subject:|e i |≤ε,i=1,2,…,n (15)

[0112] Where ω is the power weight, which is the same as the element of the weight matrix C, T1 is the regularization parameter, and ω T is the transpose of ω, and the WSA algorithm is used to optimize T1.

[0113] The calculation of the SVR prediction value is as follows:

[0114] f(x i )=∑α i ×K(x i ,x)+e i (16)

[0115] Where, α iis the Lagrange multiplier, obtained by solving the Lagrange function equation by SVR; K(x i ,x) represents the inner product kernel function, e i Error variable.

[0116] The approximate regularization parameter of the health matrix allows WSA to start optimizing near the global optimal value, obtain the optimal hyperparameters, reduce the optimization time, accelerate the prediction process, and improve the prediction speed of SVR.

[0117] S2-3, optimize predicted power.

[0118] Based on S2-2 and S2-3, according to the health matrix, the weight matrix C calculated by the health matrix in formula (8) is n×1 Take the transpose and get And C T Expressed as a vector, the improved power weight ω is obtained. Using the improved ω, the steps of equations (13) to (16) are performed, and the optimized predicted power P1 is obtained from equation (16).

[0119] Comparative Example

[0120] The data used is from Baidu KDD CUP 2022, Unit 1 of the test data set, with a total of 40 days of data. BP and WSA-SVR networks were selected for optimization. Data preprocessing used forward and backward interpolation to ensure the consistency of input data. The prediction results before and after BP network optimization are as follows: Figure 3 and Figure 4 shown.

[0121] Using R 2 There are three error indicators: coefficient of determination, MSE (mean square error), and MAE (mean absolute error), as shown in formulas (17) to (19):

[0122]

[0123]

[0124]

[0125] Where y i is the true value, is the predicted value, is the mean of the true values.

[0126] The comparison of prediction errors before and after optimization is shown in Table 1.

[0127] Table 1 Comparison of neural network prediction errors

[0128]

[0129] From Table 1 and Figure 3 , Figure 4 It can be seen that after BP network optimization, the prediction accuracy is improved by 40% and the prediction time is reduced by 20%. Figure 5 , Figure 6 It can be seen that after the WSA-SVR network is optimized, the prediction accuracy is improved by 10% and the prediction time is reduced by 45%.

[0130] The optimization algorithm of the present invention simplifies the BP neural network weight correction step and significantly improves the hysteresis of the prediction. The matrix approximation regularization parameter is used as the starting value for the optimization, which can improve the optimization efficiency.

[0131] During the WSA-SVR optimization process, the approximate regularization parameters of the health matrix are passed to the WSA-SVR network, allowing the WSA to start from a near-global optimal value, reducing the WSA-SVR prediction time. After optimization, the WSA-SVR accuracy is significantly improved.

[0132] The above examples are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modifications, equivalent variations, and modifications to the above examples that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A wind power prediction method based on an optimized neural network for health assessment, characterized in that: The following steps are involved: Step S1, obtaining wind turbine data, performing health assessment and constructing a health matrix; Step S2, based on the health matrix of step S1, calculate the improved power weight, use the matrix approximation hyperparameter as the starting value of the whale group algorithm, accelerate the SVR optimization process, and use the neural network to optimize the predicted power; The process of constructing the health matrix in step S1 is as follows: S1-1, data collection: collect wind turbine data such as wind speed, temperature, pressure, speed, power, etc. S1-2, data cleaning: data cleaning is performed to improve data reliability by targeting missing values, outliers, and noise; S1-3, feature screening: From the cleaned data, feature data is screened to obtain the approximate hyperparameters of the neural network and the dimensions of the health matrix; S1-4, state classification: calculate the power generation efficiency and classify the wind turbine state categories based on the characteristic data; S1-5, matrix generation: Arrange the power generation efficiency of the same state into a matrix according to the value; In step S1, the health matrix A ,for: ； Where, n is the health matrix dimension, is the efficiency matrix, is the characteristic correlation matrix of power; In step S2, the method for obtaining the original predicted power is: First, get the weight matrix ,Right now: ； Where, is the feature matrix; Then, the original predicted power , expanded to power matrix ,Right now: ； Next, gather D and C , and obtain the reconstructed prediction power ,Right now: ； Where, λ It is related to characteristic data and fan operating conditions, and its value ranges from 0.01 to 0.

1.

2. The wind power prediction method based on optimized neural network of health assessment according to claim 1 is characterized in that: The improved power weight calculation formula is: ； Where, represents the change in weight, is the learning rate, is the error of the current node, is the input of the current node.

3. The wind power prediction method based on optimized neural network of health assessment according to claim 1 is characterized in that: The method for accelerating the SVR optimization process is: Assume that i The input of data is , the output is , the predicted value is f ( ), define the error variable , which represents the gap between the actual value and the predicted value, that is: ； While minimizing the sum of squared errors, we limit each Does not exceed a given threshold , the constraints are: ； ； Where, ω is the power weight, and the weight matrix C The elements are the same, is the regularization parameter, for ω The transpose of Conduct optimization; The SVR prediction value is: ； Where, is the Lagrange multiplier, obtained by solving the Lagrange function equation by SVR; represents the inner product kernel function, Error variable.

Images