Wind power generation system power prediction method based on stacked sparse auto-encoder network Hammerstein model

Abstract

The invention discloses a power prediction method for a wind power generation system based on a stacked sparse auto-encoder network Hammerstein model, and aims to solve the problems that only nonlinear mapping is modeled and dynamic characteristics of the system are neglected and parameter coupling of the Hammerstein model leads to complex identification in an existing method, the Hammerstein model is constructed, an ARMAX model is utilized to describe a dynamic linear module, and the dynamic linear module is used to predict the power of the wind power generation system based on the stacked sparse auto-encoder network Hammerstein model. A static nonlinear module is described by stacking the sparse auto-encoder network; designing a zero-mean Gaussian signal input proxy model, realizing series module decoupling based on covariance function characteristics, and eliminating parameter coupling; a self-adaptive multi-strategy grey wolf optimization algorithm is adopted to determine the number of neurons in a network hidden layer, training is performed in combination with a sparse criterion function, and the feature extraction capability is improved. According to the method, synchronous capture of static nonlinear and dynamic linear characteristics is realized, the calculation complexity is reduced, the model identification precision, prediction precision and robustness are improved, and the method is suitable for accurate prediction of the power of a wind power system.

Description

Technical Field

[0001] This invention relates to a power prediction method for wind power generation systems based on the Hammerstein model of a stacked sparse autoencoder network. Background Technology

[0002] Power prediction modeling for wind power generation systems has a wide range of applications in the power and energy sectors. Research on wind power generation prediction and assessment of the impact of wind power generation on the power grid both require prediction of power generation. Power generation can be used to evaluate the energy and operating status of wind power generation systems and is of great value in determining whether wind turbine generators are operating effectively.

[0003] Wind power generation systems often exhibit complex inherent mechanisms characterized by strong nonlinearity, uncertainty, and large time delays. Traditional mathematical models struggle to accurately describe these systems, and existing, relatively mature linear system theories often fail to yield satisfactory results. In recent years, significant progress has been made in the identification and modeling of wind power generation systems, resulting in various distinctive theories and methods. Widely applied techniques include data-driven and deep learning-based wind power system modeling methods. For example, multinomial models and state-space models are used to establish wind speed-power relationships, as are deep belief networks and long short-term memory networks. However, these techniques only consider the nonlinear mapping relationship between input and output data, neglecting the dynamic characteristics of the system, leading to low identification and modeling accuracy. In summary, current wind power prediction research still faces the following two main challenges:

[0004] 1. Wind power generation systems are highly nonlinear and nonstationary dynamic systems, and their operating characteristics are influenced by a variety of coupled environmental factors. It is difficult to construct high-precision prediction models using a single neural network or traditional statistical methods. Establishing a mathematical model that can simultaneously capture the system's nonlinear and dynamic characteristics is fundamental to achieving high-precision power prediction.

[0005] 2. In parameter estimation of the Hammerstein model for wind power prediction, existing methods often lead to coupling between the parameters of nonlinear and linear modules, forming parameter product terms. This necessitates additional decomposition techniques to separate the parameters, which not only increases computational complexity but may also reduce model identification accuracy. A key challenge is how to utilize effective parameter estimation methods to reduce computational complexity while improving the estimation accuracy and robustness of system parameters.

[0006] To overcome the limitations of the above methods, this invention uses a stacked sparse autoencoder network Hammerstein model for wind power prediction. The designed Hammerstein model integrates the stacked sparse autoencoder network and the autoregressive moving average model, and takes into account the nonlinear and dynamic linear characteristics of the wind power system. It is more suitable for describing the wind power system and has important theoretical and practical significance for the accurate prediction of wind farm output power. Summary of the Invention

[0007] This invention provides a power prediction method for wind power generation systems based on the Hammerstein model of a stacked sparse autoencoder network to address the problems existing in the prior art.

[0008] The technical solutions adopted in this invention are as follows:

[0009] A power prediction method for wind power generation systems based on the Hammerstein model of stacked sparse autoencoder networks includes the following steps:

[0010] S1: Obtain wind speed and power data from the wind power generation system;

[0011] S2: Use the interquartile range method to clean outliers from the wind speed and power data to obtain cleaned data;

[0012] S3: Based on the cleaned data, establish a proxy model constructed by a stacked sparse autoencoder network, input the zero-mean Gaussian signal into the proxy model, and obtain Gaussian signal output data;

[0013] S4: Use the ARMAX model to describe the dynamic linear modules of the Hammerstein model, and use the stacked sparse autoencoder network to describe the static nonlinear modules of the Hammerstein model, and construct a Hammerstein model wind power generation system.

[0014] S5: Based on the Gaussian signal output data, the order of the ARMAX model is determined using the Hankel matrix and singular value decomposition technique, and the parameters of the ARMAX model are identified using correlation analysis. Based on the cleaned data, the number of hidden layer neurons in the stacked sparse autoencoder network is determined using the adaptive multi-strategy gray wolf optimization algorithm, and the weights and biases of the stacked sparse autoencoder network are updated using the backpropagation algorithm to obtain the trained stacked sparse autoencoder network.

[0015] S6: The stacked sparse autoencoder network trained by S5 is used as the static nonlinear module of the Hammerstein model, and the ARMAX model identified by S5 is used as the dynamic linear module of the Hammerstein model. Input wind speed data to obtain power prediction results.

[0016] Furthermore, in S2, the quartile method includes:

[0017] Based on the vertical direction, the wind speed step size is defined. The wind speed data is averaged, grouped, and sorted in ascending order. The quartiles are calculated to determine the normal range of wind speed data.

[0018] Based on the horizontal power step size, the power data is averaged, grouped, and sorted in ascending order. The quartiles are then calculated to determine the normal power data range.

[0019] Outliers that are outside the normal range for both wind speed and power data are removed.

[0020] Furthermore, the stacked sparse autoencoder network has a four-layer structure, with each layer represented as follows:

[0021] The first layer performs linear transformations and activation calculations: h 1 (t)=W1 T u(t)+b1, where u(t) is a zero-mean Gaussian signal, W1 T b1 is the first-layer encoding weight matrix, b1 is the first-layer encoding bias vector, and h 1 (t) represents the initial activation value vector of the first layer of hidden neurons;

[0022] The second layer performs k-sparser processing: Determines the index set Γ = supp k (h 1 (t)), where Γ is the set of the k neurons with the largest activation values ​​in the second layer, and the initial activation value vector of the hidden neurons in the second layer is h. 2 (t)=W2 T h 1 (t)+b2,W2 T b1 is the encoding weight matrix of the second layer, and b2 is the encoding bias vector of the second layer.

[0023] The third layer maps the features back to the reconstructed data: Where W1 is the decoding weight matrix of the first k-SAE, This is the decoding bias vector for the first k-SAE. To reconstruct the data input;

[0024] The fourth layer reconstructs the features from the first layer: Among them, W2 T This is the decoding weight matrix for the second k-SAE. This is the decoding bias vector for the second k-SAE. It is the first hidden layer h 1 The reconstructed value of (t).

[0025] Furthermore, in S3,

[0026] The Gaussian signal output data is used as the output of the static nonlinear module of the Hammerstein model. This output is defined as an intermediate variable v(t), and the intermediate variable v(t) and the zero-mean Gaussian signal u(t) satisfy the following covariance relationship:

[0027] R vu (τ)=b0R u (τ),

[0028] Where: b0 is a constant, and τ is a time constant;

[0029] R vu (τ)=E(v(t)u(t-τ)) is the cross-covariance function of the intermediate variable v(t) and the zero-mean Gaussian signal u(t), R u (τ)=E(u(t)u(t-τ)) is the autocovariance function of the zero-mean Gaussian signal u(t).

[0030] Furthermore, the stacked sparse autoencoder network is trained using the following criterion function with sparsity properties:

[0031]

[0032] Among them, E k-SAE Let L represent the loss function of the k-sparse autoencoder, L represent the data length, ^ represent the estimated value of the identification, and ε represent a pre-set small threshold.

[0033] Furthermore, the adaptive multi-strategy gray wolf optimization algorithm specifically includes:

[0034] Definition and Encoding: Using the number of neurons in the two hidden layers of a stacked sparse autoencoder network as the parameter to be optimized, a location vector is constructed:

[0035] X i = [h1, h2],

[0036] h min ≤h1,h2≤h max ,

[0037] Where h1 is the number of neurons in the first hidden layer, h2 is the number of neurons in the second hidden layer, and h1 and h2 are positive integers. min and h max The minimum and maximum number of neurons are preset;

[0038] Population initialization: The population is initialized using an infinitely folded SIN chaotic model, the expression of which is:

[0039]

[0040] -1≤x n ≤1,x n ≠0

[0041] h i =h min +x n (h max -h min ), i = 1, 2

[0042] Where, x n The nth value in the chaotic sequence;

[0043] Convergence factor control: The search range is dynamically adjusted using a nonlinear convergence factor;

[0044] Location update: A neighboring wolf pack is constructed based on Euclidean distance, and multi-neighbor learning is performed within the range of the neighboring wolf pack to update the individual location;

[0045] Location Update: Calculate the position X of each individual gray wolf i based on Euclidean distance. iold (t) and X inew1 The domain range R of (t+1) i (t)=||X iold (t)-X inew1 (t+1)||, where R i (t) represents the territory of the i-th gray wolf;

[0046] According to the R of each individual gray wolf i (t), construct the neighboring gray wolf pack N for each individual gray wolf. i (t) is:

[0047] N i (t)={X j (t)|D i (X iold (t),X inew1 (t+1))≤R i (t)}

[0048] Where N i (t) is the set of neighbors of the i-th gray wolf, X j (t) represents the position of the j-th gray wolf;

[0049] From the neighbor set N i Two individuals X are randomly selected from (t). n (t) and X r (t), generating new candidate positions based on dimension learning:

[0050] X inew2 (t+1)=X iold(t)+rand(X n (t)-X r (t))

[0051] The new position of an individual is updated based on its fitness value, as follows:

[0052]

[0053] in X is the fitness function used to evaluate performance. i (t+1) represents the position of the i-th gray wolf in the next moment;

[0054] Elite perturbation: A generalized backward learning strategy is used to perturb the position of the leader wolf in the later stages of each search. The fitness values ​​before and after the perturbation are compared to determine whether to update the position.

[0055] Furthermore, in S6, the power prediction result is obtained by inputting the wind speed data into the static nonlinear module of the Hammerstein model to obtain intermediate variables, and then inputting the intermediate variables into the dynamic linear module of the Hammerstein model to obtain the power prediction result.

[0056] The present invention has the following beneficial effects:

[0057] 1) The method of this invention integrates the mechanism characteristics of wind power generation system and constructs a wind power generation system based on the Hammerstein model, thereby improving the accuracy and generalization ability of the model.

[0058] 2) The method of this invention utilizes the covariance function characteristics of Gaussian signals in static nonlinear systems to decouple the serial modules of the Hammerstein model, thus solving the problem of unmeasurable intermediate variable information in the Hammerstein model.

[0059] 3) The method of this invention uses the adaptive multi-strategy gray wolf optimization algorithm to optimize the parameters in the stacked sparse autoencoder network, which improves efficiency, solution accuracy and convergence speed, thereby achieving better performance. Attached Figure Description

[0060] Figure 1 This is a model diagram of a wind power generation system based on the Hammerstein model.

[0061] Figure 2 This is a diagram of the stacked sparse autoencoder network structure of the present invention.

[0062] Figure 3 This is a flowchart of the identification process of the present invention.

[0063] Figure 4 This is a nonlinear fitting graph from the present invention.

[0064] Figure 5 This is a comparison chart of the output power of the present invention. Detailed Implementation

[0065] The invention will now be further described with reference to the accompanying drawings.

[0066] This invention discloses a power prediction method for wind power generation systems based on a Hammerstein model using a stacked sparse autoencoder network, comprising the following steps:

[0067] S1: Obtain wind speed and power data from the wind power generation system;

[0068] S2: Use the interquartile range method to clean outliers from the wind speed and power data to obtain cleaned data;

[0069] S3: Based on the cleaned data, establish a proxy model constructed by a stacked sparse autoencoder network, input the zero-mean Gaussian signal into the proxy model, and obtain Gaussian signal output data;

[0070] S4: Use the ARMAX model to describe the dynamic linear modules of the Hammerstein model, and use the stacked sparse autoencoder network to describe the static nonlinear modules of the Hammerstein model, and construct a Hammerstein model wind power generation system.

[0071] S5: Based on the Gaussian signal output data, the order of the ARMAX model is determined using the Hankel matrix and singular value decomposition technique, and the parameters of the ARMAX model are identified using correlation analysis. Based on the cleaned data, the number of hidden layer neurons in the stacked sparse autoencoder network is determined using the adaptive multi-strategy gray wolf optimization algorithm, and the weights and biases of the stacked sparse autoencoder network are updated using the backpropagation algorithm to obtain the trained stacked sparse autoencoder network.

[0072] S6: The stacked sparse autoencoder network trained by S5 is used as the static nonlinear module of the Hammerstein model, and the ARMAX model identified by S5 is used as the dynamic linear module of the Hammerstein model. Input wind speed data to obtain power prediction results.

[0073] The following is a detailed description of each step.

[0074] Data cleaning using the quartile method, as detailed below:

[0075] The wind speed step size was defined based on the vertical direction, and the wind speed data were averaged and grouped. Then, the data were sorted in ascending order. The steps for calculating the quartiles are as follows:

[0076] For each set of wind speed data, data points Q1, Q2, and Q3 are defined. The sequence is divided into four parts and sorted in ascending order: P = [p1, p2, ..., p...]. p Q2 can be calculated as follows:

[0077]

[0078] Then, calculate Q1 and Q3. Q2 divides P into two subsequences, and calculates the medians Q'2 and Q”2 of the two subsequences respectively, thus obtaining Q1 = Q'2 and Q3 = Q”2.

[0079] Let p = 4k + 1 (k ∈ R) + If ), then the following is given:

[0080]

[0081] Let p = 4k + 3 (k ∈ R) + If ), then the following is given:

[0082]

[0083] Based on Q1 and Q3, define I QR for:

[0084] I QR =Q3-Q1,

[0085] This allows us to obtain the effective value [F]. b ,F t The scope of ] is:

[0086] [F b ,F t ]=[Q1-1.5I QR Q3+1.5I QR ].

[0087] For the horizontal direction, a power step size was defined, and the power data were averaged and grouped. Then, wind speeds were sorted in ascending order. Using the data from the previous step, quartiles were calculated for the vertical direction. At this point, data cleaning is complete.

[0088] Remove data points where the wind speed is below the specified cut-in wind speed and the power generation is zero. Then, find data points where the wind speed exceeds the cut-in wind speed, and where the power output is close to zero or is itself close to zero, and exclude them from the dataset.

[0089] The stacked sparse autoencoder network in this invention has a four-layer structure, and each layer is represented as follows:

[0090] (1) The first layer performs linear transformation and activation calculation, as follows:

[0091] h1 (t)=W1 T u(t)+b1,

[0092] Where u(t) is the input signal, W1 T b1 is the first-layer coding matrix, b1 is the first-layer coding bias vector, and h 1 (t) represents the initial activation value vector of the first layer of hidden neurons.

[0093] (2) The second layer undergoes k-sparse processing, and L1 regularization is used to restrict the activation of neurons in the hidden layer. By introducing sparsity constraints, the most discriminative features in the data are focused, thereby improving the model's generalization ability. When the input data is noisy, sparsity constraints enable the model to learn robust representations and effectively remove noise interference, as shown below:

[0094] Γ=supp k (h 1 (t)),

[0095]

[0096] h 2 (t)=W2 T h 1 (t)+b2,

[0097] Where Γ represents the set of the k neurons with the highest activation values ​​in the second layer, (Γ) c W2 is the complement of Γ. T b1 is the encoding weight matrix of the second layer, b2 is the encoding bias vector of the second layer, and h is the encoding weight matrix of the second layer. 2 (t) The initial activation vector of the hidden neurons in the second layer.

[0098] (3) The third layer will h 1 The features in (t) are mapped back to the reconstructed data. The process is represented as:

[0099]

[0100] Where W1 is the decoding weight matrix of the first k-SAE, This is the decoding bias vector for the first k-SAE. To reconstruct the data input.

[0101] (4) The fourth layer performs processing on the features h of the first layer. 1 The estimate of (t) is expressed as:

[0102]

[0103] Among them, W2 T This is the decoding weight matrix for the second k-SAE. This is the decoding bias vector for the second k-SAE. It is the first hidden layer h 1 The estimated value of (t).

[0104] The static nonlinear module and dynamic linear module of the Hammerstein model are decoupled using the covariance function properties of Gaussian signals, specifically as follows:

[0105] A Gaussian signal is designed, and the corresponding output is obtained by passing the Gaussian signal through a surrogate model constructed by a stacked sparse autoencoder network, thereby obtaining the Gaussian signal input and output data.

[0106] For a static nonlinear module, assuming the input is a zero-mean Gaussian signal, the product of the autocovariance function of the Gaussian input signal u(t) (i.e., the zero-mean Gaussian signal) and a constant is equal to the covariance function of the intermediate variable v(t) and the Gaussian input signal u(t), that is:

[0107] R vu (τ)=b0R u (τ)

[0108] Where: b0 is a constant, τ is a time constant; R vu (τ)=E(v(t)u(t-τ)) represents the cross-covariance function between the input u(t) and the intermediate variable v(t), R u (τ)=E(u(t)u(t-τ)) is the autocovariance function of the input u(t).

[0109] Utilize b0R u (τ) replaces R vu (τ) can decouple the static nonlinear module and the dynamic linear module of the Hammerstein model.

[0110] In this invention, u G (t) consists of a set of zero-mean Gaussian signals, i.e., u G (t)=[u G1 (t),u G2 (t)] T After passing through the Hammerstein model, the output is y. G (t). Due to the properties of the covariance function It can be seen that the problem of the unmeasurable intermediate variable information v(t) in the Hammerstein model has been solved, that is, by utilizing... Replace the unknown covariance matrix This achieves the decoupling of the static nonlinear module and the linear module.

[0111] A wind power generation system model is established using the Hammerstein model, which is as follows:

[0112] v(t) = f(u(t)),

[0113]

[0114] Where u(t) is the input of the Hammerstein model, and in the wind power generation system, u(t) represents the variables related to the input of the wind power generation system, such as wind speed.

[0115] f(·) is a static nonlinear block, i.e., a stacked sparse autoencoder network.

[0116] v(t) is an intermediate variable after passing through the static nonlinear module, used to characterize the static nonlinear characteristics of the wind power generation system.

[0117] y(t) is the output of the Hammerstein model. In a wind power generation system, y(t) represents variables related to the output of the wind power generation system, such as power.

[0118] This is the denominator polynomial of the dynamic linear module, used to describe the dynamic characteristics of the wind power generation system.

[0119] The numerator polynomial of the dynamic linear module works together with the denominator polynomial to describe the dynamic characteristics of the system.

[0120] z -1 This represents the unit shift operator, used to construct the dynamic equations of a discrete-time system.

[0121] n a and n b The orders of the denominator and numerator polynomials are respectively, and they are predetermined model parameters.

[0122] e(t) is the noise term, representing the unmodeled dynamic or measured noise in the wind power generation system.

[0123] t represents time.

[0124] To improve the model's generalization and anti-interference capabilities, based on the Hammerstein model of the stacked sparse autoencoder network constructed above, the following criterion function with sparsity properties is defined:

[0125]

[0126] Among them, E k-SAELet L represent the loss function of the k-sparse autoencoder, L represent the data length, ^ represent the estimated value of the identification, and ε represent a pre-set small threshold.

[0127] In S5: Based on the Hankel matrix and singular value decomposition technique, the order of ARMAX is adaptively calculated to be 2. The order of the ARMAX model is then adaptively determined using the Hankel matrix and singular value decomposition technique.

[0128] Construct the Hankel matrix as follows:

[0129]

[0130] Where g = [g1, g2, ..., g m [ ] represents the impulse response.

[0131] Perform singular value decomposition on the Hankel matrix:

[0132]

[0133] Where U and V are orthogonal matrices, and S = diag(σ1,σ2,…,σ) l ).

[0134] The system order is determined by calculating the ratio of adjacent singular values:

[0135]

[0136] The expression for the ARMAX model, identified using correlation analysis, is as follows:

[0137]

[0138] Multiplying both sides of the formula by u1(t-τ) and calculating the expected value, we get:

[0139]

[0140] Among them, R yu (τ)=E(y(t)u(t-τ)), R eu (τ)=E(e(t)u(t-τ)).

[0141] Based on the independence between the input signal and the noise signal, R can be obtained. eu (τ)=0, therefore we can obtain:

[0142]

[0143] in

[0144]

[0145] in R = [R yu (1),R yu (2),…,R yu (q)],

[0146] The final result is:

[0147]

[0148] Stacked sparse autoencoder networks, through their multi-layered encoder structure, perform nonlinear transformations and feature extraction on input wind speed data layer by layer, aiming to learn a high-order abstract mapping from wind speed to power. The performance of this network is highly dependent on its parameter configuration. However, due to the complex network structure and the high-dimensional, non-convex parameter space, traditional gradient descent algorithms are prone to getting trapped in local optima and suffer from low training efficiency.

[0149] To address the aforementioned problems, this invention employs an adaptive multi-strategy gray wolf optimization algorithm for global parameter optimization search in stacked sparse autoencoder networks. The core idea of ​​this algorithm is to simulate the social hierarchy and cooperative hunting mechanism of a gray wolf population, guiding the population to efficiently and robustly seek optimization within a complex parameter space by dynamically adjusting the balance between "exploration" and "exploitation." It effectively avoids local optima and achieves global optimal search for key network parameters by adaptively switching between multiple strategies such as encirclement, pursuit, and attack, and updating the social hierarchy in real time based on individual fitness. The adaptive multi-strategy gray wolf optimization algorithm optimizes the number of hidden layer neurons in the stacked sparse autoencoder network, expressed as:

[0150] (1) Definition and Encoding:

[0151] Using the number of neurons in the two hidden layers of the stacked sparse autoencoder network as the parameter to be optimized, a location vector is constructed, the expression of which is:

[0152] X i = [h1, h2],

[0153] h min ≤h1,h2≤h max ,

[0154] Where h1 represents the number of neurons in the first hidden layer, h2 represents the number of neurons in the second hidden layer, and h1 and h2 are positive integers. min and h max The minimum and maximum number of neurons are preset.

[0155] (2) Initializing the population using the chaotic model:

[0156] Population initialization is performed using the infinitely folded SIN chaotic model, expressed as follows:

[0157]

[0158] -1≤x n ≤1,x n ≠0,

[0159] h i =h min +x n (h max -h min ), i = 1, 2,

[0160] Where x n It is the nth value of the chaotic sequence.

[0161] (3) Nonlinear convergence factor:

[0162] A nonlinear convergence factor is used to adapt the optimization of the chaotic model. The expression for the nonlinear convergence factor is:

[0163] a(t) = a init -(a init -a fin (1 / T) max ) 2 ,

[0164] In the formula: a init With a fin Let T be the initial value of the convergence factor. ma x represents the maximum number of iterations;

[0165] (4) Dimensional learning individual updates:

[0166] To improve the search capabilities of the wolf pack, a position update method based on dimensional learning is introduced among individual wolves. The specific strategy is as follows:

[0167] Based on the current positions and corresponding fitness values ​​of all gray wolf individuals, the three best individuals are selected and named α wolf, β wolf, and δ wolf, respectively. Their positions are denoted as X. α (t), X β (t) and X δ (t).

[0168] Candidate positions facing the alpha wolf:

[0169] X1(t)=X α (t)-A1|C1X α (t)-X iold (t)|,

[0170] Candidate positions facing the beta wolf:

[0171] X2(t)=X β (t)-A2|C2X β (t)-X iold (t)|,

[0172] Candidate positions facing the δ wolf:

[0173] X3(t)=X δ (t)-A3|C3X δ (t)-X iold (t)|,

[0174] Where A i and C i Let A be a random coefficient. i =2ar1-a,C i =2r², i = 1, 2, 3, r1 and r2 are random numbers in [0, 1, ..., 2], a represents the control parameter, and a ∈ [0, 2]. i (t) is the position vector of the i-th gray wolf.

[0175] Calculate the target position X inew1 (t+1):

[0176]

[0177] First, calculate the position X of each individual gray wolf i using Euclidean distance. iold (t) and X inew1 The domain range R of (t+1) i (t) is:

[0178] R i (t)=||X iold (t)-X inew1 (t+1)||,

[0179] Where R i (t) represents the territory of the i-th gray wolf.

[0180] Secondly, based on the R of each individual gray wolf i (t), construct the neighboring gray wolf pack N for each individual gray wolf. i (t) is:

[0181] N i (t)={X j (t)|D i (X iold (t),X inew1 (t+1))≤R i (t)},

[0182] Where N i(t) is the set of neighbors of the i-th gray wolf, X j (t) represents the position of the j-th gray wolf.

[0183] From the neighbor set N i Two individuals X are randomly selected from (t). n (t) and X r (t), generating new candidate positions based on dimension learning:

[0184] X inew2 (t+1)=X iold (t)+rand(X n (t)-X r (t)),

[0185] Finally, the individual's new position is updated based on its fitness value, as shown below:

[0186]

[0187] in X is the fitness function used to evaluate performance. i (t+1) represents the position of the i-th gray wolf in the next moment.

[0188] (5) Generalized reverse learning mutation:

[0189] To avoid local optima, a generalized back-learning strategy is used to perturb the position of the leader wolf in the later stages of each search, which is expressed as:

[0190]

[0191]

[0192] In the formula, X α,i X β,i X δ,i Let α, β, and δ be the positions of the leading wolves in the i-th dimension. i Let b be the minimum value of the current population in the i-th dimension. i is the maximum value of the current population in the i-th dimension, and k is a scaling factor that takes the value [0,1]. After perturbing the leader wolves α, β, and δ, the fitness values ​​before and after the perturbation are compared to determine whether the leader wolves α, β, and δ have updated their positions.

[0193] Based on the backpropagation algorithm, the expression for identifying stacked sparse autoencoder networks is as follows:

[0194] (1) The loss function of the k-sparse autoencoder is defined as follows:

[0195]

[0196] (2) Based on the backpropagation technique, the gradient of the bias between the weights w1 and b2 from the output layer to the hidden layer is calculated, and the expression is as follows:

[0197]

[0198] (3) Calculate h 1 The gradient of (t) is expressed as follows:

[0199]

[0200] (4) Because h 1 (t) is already k-sparse, so only the k largest activation values ​​are retained, and their corresponding gradients should also propagate in the selected set Γ. The rest should be set to zero, so h 1 The gradient of (t) is:

[0201]

[0202] Where Γ is the mask matrix, which takes the value 1 only at the Γ position and 0 at the other positions, and ⊙ represents the Hadamard product.

[0203] (5) Next, the weights from the hidden layer to the input layer will be calculated. The gradient of bias b1:

[0204]

[0205] (6) Next, the parameter θ is adjusted using the gradient descent algorithm. k-SAE Update:

[0206]

[0207] Where η is the learning rate.

[0208] In S6: Based on the identified Hammerstein model, the wind speed of the wind power generation system is used as the input of the Hammerstein model to obtain the corresponding predicted power output.

[0209] In the Hammerstein model wind power prediction method using the stacked sparse autoencoder network described above, a mechanistic function model of the wind power generation system is constructed. This model not only retains the physical characteristics and good generalization of the mechanistic properties, but also introduces a stacked sparse autoencoder network. Through this nested structure, the model can still improve prediction accuracy and optimize computational efficiency, thereby enhancing overall performance, even with limited or low-quality data.

[0210] In the above identification process, Gaussian signals and wind power data are used to separate and identify the parameters of the static nonlinear module and the dynamic linear module of the Hammerstein model, which improves the identification accuracy and has good predictive performance.

[0211] The invention will be further described below in conjunction with the parameters.

[0212] like Figures 1 to 3 The Hammerstein model identification based on the wind power generation system model is as follows:

[0213] (1) According to the wind power generation system model structure diagram, since the wind turbine and generator are nonlinear, the low-speed shaft, gearbox, and high-speed shaft are approximately linear within the normal operating range. The Hammerstein model consists of a static nonlinear module and a dynamic linear module connected in series, so it can be used to build a Hammerstein-type wind power generation system:

[0214] (2) A quartile data cleaning technique involving horizontal and vertical dimensions was used to eliminate outlier data points, and a dataset consisting of 3,628 data pairs involving wind speed and power was finally obtained.

[0215] (3) A surrogate model of the wind power generation system was established using a stacked sparse autoencoder network. The number of hidden layers in the network was determined using an adaptive multi-strategy gray wolf optimization algorithm, with the first and second hidden layers having 29 and 42 layers, respectively. Then, the k value in the stacked sparse autoencoder network was set to 6. The surrogate model was trained using data from 1000 sets of wind power generation systems, with 50 iterations. The training model achieved the smallest mean square error, which was 7.385E-4.

[0216] (4) The Gaussian signal u(t) in this invention is 1000 Gaussian signals with a mean of 0 and a variance of 0.05. The corresponding output y(t) is obtained by the trained surrogate model. First, the order of the dynamic linear module is selected as 2 using the singular values ​​of the Hankel matrix. Then, the unknown parameters of the dynamic linear module of the Hammerstein model are identified by the correlation analysis method, namely [a1, a2, b1, b2] = [-0.952, 0.6187, 1.4334, -0.9014].

[0217] (5) A static nonlinear module of the Hammerstein model is established using a stacked sparse autoencoder network. The network model is trained based on 2447 sets of data. The number of hidden layers, i.e., the first hidden layer and the second hidden layer, is determined using an adaptive multi-strategy gray wolf optimization algorithm. The parameter is h. 1 =25,h 2=33, and the k value in the network is set to 6. The training yields the smallest mean squared error, with an error of 1.0247E5(kW).

[0218] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

[0219] like Figure 4 The recognition result diagram of the Hammerstein model shown below includes the following steps:

[0220] This invention utilizes wind speed and power data from wind power generation systems to train a Hammerstein model. The trained model achieves the smallest mean square error, at 1.0247E5 (kW). The training results clearly demonstrate that the Hammerstein model designed in this invention achieves high accuracy, providing support for the modeling and prediction of wind power generation systems.

[0221] like Figure 5 As shown, a power prediction method for wind power generation systems based on the Hammerstein stacked sparse autoencoder network includes the following steps:

[0222] Based on the remaining 181 sets of original data, and after the identification of the Hammerstein model, the power of the wind power generation system is predicted, and the real power curve is constructed. It can be clearly seen from the prediction results that the Hammerstein wind power generation system power prediction method based on stacked sparse autoencoder network designed in this invention can achieve high prediction accuracy.

[0223] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A power prediction method for wind power generation systems based on the Hammerstein model of a stacked sparse autoencoder network, characterized in that: Includes the following steps: S1: Obtain wind speed and power data from the wind power generation system; S2: Use the interquartile range method to clean outliers from the wind speed and power data to obtain cleaned data; S3: Based on the cleaned data, establish a proxy model constructed by a stacked sparse autoencoder network, input the zero-mean Gaussian signal into the proxy model, and obtain Gaussian signal output data; S4: Use the ARMAX model to describe the dynamic linear modules of the Hammerstein model, and use the stacked sparse autoencoder network to describe the static nonlinear modules of the Hammerstein model, and construct a Hammerstein model wind power generation system. S5: Based on the Gaussian signal output data, the order of the ARMAX model is determined using the Hankel matrix and singular value decomposition technique, and the parameters of the ARMAX model are identified using correlation analysis. Based on the cleaned data, the number of hidden layer neurons in the stacked sparse autoencoder network is determined using the adaptive multi-strategy gray wolf optimization algorithm, and the weights and biases of the stacked sparse autoencoder network are updated using the backpropagation algorithm to obtain the trained stacked sparse autoencoder network. S6: The stacked sparse autoencoder network trained by S5 is used as the static nonlinear module of the Hammerstein model, and the ARMAX model identified by S5 is used as the dynamic linear module of the Hammerstein model. Input wind speed data to obtain power prediction results.

2. The wind power generation system power prediction method based on the Hammerstein model of stacked sparse autoencoder network as described in claim 1, characterized in that: In S2, the quartile method includes: Based on the vertical direction, the wind speed step size is defined. The wind speed data is averaged, grouped, and sorted in ascending order. The quartiles are calculated to determine the normal range of wind speed data. Based on the horizontal power step size, the power data is averaged, grouped, and sorted in ascending order. The quartiles are then calculated to determine the normal power data range. Outliers that are outside the normal range for both wind speed and power data are removed.

3. The power prediction method for wind power generation systems based on the Hammerstein model of stacked sparse autoencoder network as described in claim 1, characterized in that: The stacked sparse autoencoder network has a four-layer structure, with each layer represented as follows: The first layer performs linear transformations and activation calculations: ,in, The signal is a zero-mean Gaussian signal. This is the first layer encoding weight matrix. This is the first layer encoding bias vector. This is the initial activation value vector of the first layer of hidden neurons; The second layer performs k-sparseness processing: determining the index set. ,in, Let k be the set of the k neurons with the largest activation values ​​in the second layer. The initial activation value vector of the hidden neurons in the second layer is: , This is the encoding weight matrix for the second layer. This is the encoding bias vector for the second layer; The third layer maps the features back to the reconstructed data: ,in, This is the decoding weight matrix for the first k-SAE. This is the decoding bias vector for the first k-SAE. To reconstruct the data input; The fourth layer reconstructs the features from the first layer: ,in, This is the decoding weight matrix for the second k-SAE. This is the decoding bias vector for the second k-SAE. For the first hidden layer The reconstructed value.

4. The wind power generation system power prediction method based on the Hammerstein model of stacked sparse autoencoder network as described in claim 1, characterized in that: In S3, The Gaussian signal output data is used as the static nonlinear module output of the Hammerstein model, and this output is defined as an intermediate variable. The intermediate variable With zero-mean Gaussian signal The following covariance relationship must be satisfied: ， in: It is a constant. It is a time constant; intermediate variables With zero-mean Gaussian signal The cross-covariance function, Zero-mean Gaussian signal The autocovariance function.

5. The power prediction method for wind power generation systems based on the Hammerstein model of stacked sparse autoencoder network as described in claim 1, characterized in that: Stacked sparse autoencoder networks are trained using the following criterion function with sparsity properties: ， in, Let L represent the loss function of the k-sparse autoencoder, L represent the data length, ^ represent the estimated value of the identification, and ε represent a pre-set small threshold.

6. The wind power generation system power prediction method based on the Hammerstein model of stacked sparse autoencoder network as described in claim 1, characterized in that: The adaptive multi-strategy gray wolf optimization algorithm specifically includes: Definition and Encoding: Using the number of neurons in the two hidden layers of a stacked sparse autoencoder network as the parameter to be optimized, a location vector is constructed: ， ， in, This represents the number of neurons in the first hidden layer. This represents the number of neurons in the second hidden layer. and It is a positive integer. and The minimum and maximum number of neurons are preset; Population initialization: The population is initialized using an infinitely folded SIN chaotic model, the expression of which is: in, For the chaotic sequence One value; Convergence factor control: The search range is dynamically adjusted using a nonlinear convergence factor; Location update: A neighboring wolf pack is constructed based on Euclidean distance, and multi-neighbor learning is performed within the range of the neighboring wolf pack to update the individual location; Location Update: Calculate the position of each individual gray wolf i based on Euclidean distance. and Scope of fields ,in Let i be the territory of the i-th gray wolf; According to the R of each individual gray wolf i (t), construct the neighboring gray wolf pack N for each individual gray wolf. i (t) is: in Let i be the set of neighbors of the i-th gray wolf. Let j be the position of the j-th gray wolf; From the neighborhood set Two individuals are randomly selected from the data. and Generate a new candidate location based on dimensionality learning: The new position of an individual is updated based on its fitness value, as follows: in The fitness function is used to evaluate performance. Let i be the position of the i-th gray wolf in the next moment; Elite perturbation: A generalized backward learning strategy is used to perturb the position of the leader wolf in the later stages of each search. The fitness values ​​before and after the perturbation are compared to determine whether to update the position.

7. The wind power generation system power prediction method based on the Hammerstein model of stacked sparse autoencoder network as described in claim 1, characterized in that: In S6, the power prediction result is obtained by inputting the wind speed data into the static nonlinear module of the Hammerstein model to obtain intermediate variables, and then inputting the intermediate variables into the dynamic linear module of the Hammerstein model to obtain the power prediction result.

Images