A wind power system power prediction method fusing a sparse deep belief network hammerstein model and error correction

Abstract

The application discloses a wind power system power prediction method fusing a sparse deep belief network Hammerstein model and error correction, acquires wind speed data and power data of a wind power system, cleans outliers by using a quartile method, maps an original input space to a high-dimensional feature space by using a Gaussian kernel function, establishes a proxy model constructed by a sparse deep belief network, and designs zero-mean Gaussian signal input to acquire input and output data; the application describes a dynamic linear module and a static nonlinear module of the Hammerstein model by using an autoregressive moving average model and a sparse deep belief network respectively, establishes a Hammerstein model structure; the application adaptively calculates an order of the autoregressive moving average model and identifies parameters by using a Hankel matrix and a singular value decomposition technology, simultaneously updates weights and biases of the sparse deep belief network by using a contrast divergence algorithm and a back propagation algorithm; the application combines the identified modules to perform power prediction, and performs error correction on the predicted power by using a random forest algorithm.

Description

Technical Field

[0001] This invention relates to a wind power system power prediction method that integrates a sparse deep belief network Hammerstein model and error correction. Background Technology

[0002] Power prediction for wind power systems is a key technology for power system dispatching and operation control. Wind power systems have complex characteristics such as strong nonlinearity, uncertainty, and large time delay, making it difficult for traditional mathematical mechanistic models to accurately describe such systems.

[0003] Existing technologies mainly employ data-driven and deep learning-based methods to establish wind speed-power mapping relationships, such as using multinomial models, state-space models, or deep belief networks and long short-term memory networks. However, these methods only consider the nonlinear mapping relationship between input and output data and do not simultaneously consider the dynamic characteristics of the system, resulting in limited accuracy in identification and modeling.

[0004] Furthermore, in Hammerstein model parameter estimation, existing methods often lead to the coupling of parameters between nonlinear and linear modules, forming parameter product terms. Additional decomposition techniques are required to separate the parameters, which increases computational complexity and may reduce the model's identification accuracy.

[0005] Therefore, how to establish a mathematical model that can simultaneously capture the nonlinear and dynamic characteristics of the system, and how to effectively decouple and identify the parameters of the series modules of the Hammerstein model, are technical problems that urgently need to be solved in the field of wind power prediction. Summary of the Invention

[0006] This invention provides a wind power prediction method that integrates a sparse deep belief network Hammerstein model and error correction to address the problems existing in the prior art.

[0007] The technical solutions adopted in this invention are as follows:

[0008] A wind power system power prediction method integrating a sparse deep belief network Hammerstein model and error correction includes the following steps:

[0009] S1: Obtain wind speed and power data from the wind power generation system;

[0010] S2: Use the quartile method to clean outliers from wind speed and power data;

[0011] S3: Based on the wind speed and power data after cleaning, a Gaussian kernel function is used to map the original input space to a high-dimensional feature space, and a proxy model of the wind power generation system constructed by a sparse deep belief network is established; a zero-mean Gaussian signal is designed and input into the proxy model to obtain the corresponding Gaussian signal output;

[0012] S4: Describe the dynamic linear module of the Hammerstein model using an autoregressive moving average model, and describe the static nonlinear module of the Hammerstein model using a sparse deep belief network, thus establishing the Hammerstein model structure.

[0013] S5: Adaptively calculate the order of the autoregressive moving average model using the Hankel matrix and singular value decomposition technique. Based on the input and output data of the zero-mean Gaussian signal, identify the parameters of the autoregressive moving average model using correlation analysis. Based on wind speed and power data, update the weights and biases of the sparse deep belief network using the contrastive divergence algorithm and backpropagation algorithm to obtain the optimal parameters.

[0014] S6: The updated sparse deep belief network is used as the static nonlinear module of the Hammerstein model structure, and the identified autoregressive moving average model is used as the dynamic linear module of the Hammerstein model structure. The power is predicted using the Hammerstein model structure to obtain the predicted power, and the error is corrected by the random forest algorithm.

[0015] Furthermore, in S2, the quartile method includes vertical data cleaning based on wind speed data grouping and horizontal data cleaning based on power data grouping.

[0016] Furthermore, in S3, the Gaussian kernel function is defined as:

[0017] ,

[0018] in, , For the i-th and j-th input sample vectors, The squared Euclidean distance between the two samples is given. For kernel width parameter, This is the kernel function value.

[0019] Furthermore, in S3, the covariance function characteristics of the zero-mean Gaussian signal are used to decouple the static nonlinear module and the dynamic linear module of the Hammerstein model structure. The covariance function characteristics are expressed as follows:

[0020] ,

[0021] in, It is a constant. It is a time constant; For input With intermediate variables The cross-covariance function; Let u(t) be the autocovariance function of the input u(t).

[0022] Furthermore, in S3, the sparse deep belief network includes a multi-layer neural network structure, with each layer undergoing unsupervised pre-training using a contrastive divergence algorithm, and global fine-tuning performed using a backpropagation algorithm after pre-training.

[0023] Furthermore, in S5, the order of the autoregressive moving average model is adaptively calculated, specifically by determining the order by calculating the ratio of adjacent singular values.

[0024] Furthermore, in S5, the parameters of the autoregressive moving average model are identified, including the denominator and numerator polynomial coefficients of the dynamic linear module.

[0025] Furthermore, in S6, the random forest algorithm is used to correct the error of the predicted power, including constructing a training sample set for the random forest correction model. The input features of the training sample set include the current input wind speed, historical predicted values ​​of the Hammerstein model structure, historical residuals, and operating condition features.

[0026] Furthermore, in S6, the error correction includes:

[0027] Calculate the predicted residuals:

[0028] ,

[0029] The random forest regression algorithm is used to learn the nonlinear mapping relationship between the input features and the predicted residuals, resulting in a residual prediction model:

[0030] ,

[0031] The final corrected power prediction value is:

[0032] ,

[0033] in, For the first The actual power of each sample The initial predicted power for the Hammerstein model structure is... This represents a random forest regressor composed of multiple decision trees. The predicted residual value;

[0034] The training objective of random forest is to minimize the corrected mean squared error.

[0035] ,

[0036] Where N is the total number of training samples.

[0037] Furthermore, in S5, the optimal parameters are obtained, including defining a criterion function with sparsity properties, wherein the criterion function is:

[0038] ,

[0039] in, The fitting error of the Hammerstein model structure;

[0040] For data length;

[0041] For a specific moment;

[0042] The actual power at time t;

[0043] The power is estimated for identification at time t; A pre-set, smaller threshold; Identify estimated values ​​for intermediate variables after passing through the static nonlinear module;

[0044] The identified static nonlinear mapping function; Let be the input wind speed at time t; The estimated value for the numerator polynomial of the dynamic linear module; is the estimated value of the denominator polynomial for the dynamic linear module; z is the unit shift operator; is the noise term; ESK−DBN is the loss function of the sparse deep belief network; The input wind speed is the estimated value at time t. To support the set; The support set operator for taking the top k largest elements; This represents the state vector of the neurons in the first hidden layer. To support the collection The complement of the set corresponds to the neuron state.

[0045] The present invention has the following beneficial effects:

[0046] 1) By integrating sparse deep belief networks and autoregressive moving average models to construct the Hammerstein model, the nonlinear and dynamic linear characteristics of the wind power system are considered, which improves the prediction accuracy and generalization ability of the model.

[0047] 2) By utilizing the covariance function characteristics of zero-mean Gaussian signals in static nonlinear systems, the static nonlinear module and dynamic linear module of the Hammerstein model are decoupled, solving the problem of unmeasurable intermediate variable information in the Hammerstein model and reducing the computational complexity of parameter identification.

[0048] 3) The random forest algorithm is used to correct the error of the predicted power of the sparse deep belief network Hammerstein model. By constructing a random forest regressor to learn the nonlinear mapping relationship between the prediction residual and the input features, the systematic bias of the model is effectively compensated, and the accuracy and robustness of power prediction are improved. Attached Figure Description

[0049] Figure 1 This is a model diagram of a wind power generation system based on the Hammerstein model.

[0050] Figure 2 This is a structural diagram of a deep belief network.

[0051] Figure 3 This is a flowchart of the identification process of the present invention.

[0052] Figure 4 This is a comparison chart of the error correction power of the random forest in this invention. Detailed Implementation

[0053] The invention will now be further described with reference to the accompanying drawings.

[0054] This invention discloses a wind power system power prediction method that integrates a sparse deep belief network Hammerstein model and error correction, the steps of which are as follows (e.g.) Figure 3 ):

[0055] S1: Obtain wind speed and power data from the wind power generation system;

[0056] S2: Use the quartile method to clean outliers from wind speed and power data;

[0057] S3: Based on the wind speed and power data after cleaning, a Gaussian kernel function is used to map the original input space to a high-dimensional feature space, and a proxy model of the wind power generation system constructed by a sparse deep belief network is established; a zero-mean Gaussian signal is designed and input into the proxy model to obtain the corresponding Gaussian signal output;

[0058] S4: Describe the dynamic linear module of the Hammerstein model using an autoregressive moving average (ARMA) model, and describe the static nonlinear module of the Hammerstein model using a sparse deep belief network, thus establishing the Hammerstein model structure.

[0059] S5: Adaptively calculate the order of the autoregressive moving average model using Hankel matrix and singular value decomposition techniques; identify the parameters of the autoregressive moving average model (ARMA model) based on the input and output data of the zero-mean Gaussian signal using correlation analysis; update the weights and biases of the sparse deep belief network based on wind speed and power data using contrastive divergence algorithm and backpropagation algorithm.

[0060] S6: The updated sparse deep belief network is used as the static nonlinear module of the Hammerstein model structure, and the identified autoregressive moving average model is used as the dynamic linear module of the Hammerstein model structure. Power prediction is performed using the Hammerstein model structure to obtain the predicted power, and the predicted power is corrected for error using the random forest algorithm. Figure 4 .

[0061] The following is a detailed description of each step.

[0062] (1) Remove data points where the wind speed is lower than 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, where the power output is close to zero or is close to zero, and remove them from the dataset.

[0063] (2) A data cleaning technique incorporating both horizontal and vertical quartiles is proposed. The wind speed step size is defined based on the vertical direction, and the wind speed data is averaged and grouped. Then, the data is sorted in ascending order. The calculation steps for the quartiles are as follows:

[0064] For each group of wind speed data, data points were defined. ,and The sequence is divided into four parts. Sort in ascending order. , The calculation can be performed as follows:

[0065] ,

[0066] Then, we calculate and . Will Divide the data into two subsequences and calculate the median of each subsequence. and Then get and .

[0067] set up Then the following is given:

[0068] ,

[0069] set up Then the following is given:

[0070] ,

[0071] according to and ,definition for:

[0072] ,

[0073] This allows us to obtain valid values. The scope is:

[0074] ,

[0075] 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.

[0076] The Gaussian kernel function is used to map the original input space to a high-dimensional feature space. The kernel function is defined as follows:

[0077] ,

[0078] in, , For the i-th and j-th input sample vectors, The squared Euclidean distance between the two samples is given. For kernel width parameter, This is the kernel function value.

[0079] Then the feature matrix after kernel function mapping It is directly used as input to a sparse deep belief network, which is composed of multiple sparse restricted Boltzmann machines stacked together, such as... Figure 2 The first layer of the sparse restricted Boltzmann machine introduces sparse regularization constraints, and the activation probabilities of neurons in each layer are expressed as follows:

[0080] ,

[0081] ,

[0082] ,

[0083] in, Let j be the state of the j-th hidden layer neuron. This represents the state of the i-th visible layer neuron. The input vector for the visible layer is the feature vector after kernel mapping. This represents the hidden layer state vector. It is the Sigmoid activation function. For the bias term of the j-th hidden layer neuron, For the bias term of the i-th visible layer neuron, Let be the connection weights between the i-th neuron in the visible layer and the j-th neuron in the hidden layer. The sparsity penalty coefficient, Let N be the average activation probability of the j-th hidden layer neuron, and N be the total number of samples. Let be the target sparsity, and be a pre-defined constant with a value range of (0,1).

[0084] The output of the first hidden layer is then used as the input to the second sparse restricted Boltzmann machine (SRM). Multiple SRMs are stacked layer by layer to construct a deep belief network. The training of each layer uses the same sparsity constraints and activation probability calculation formula as the first layer. Each layer undergoes unsupervised pre-training using the contrastive divergence algorithm. After pre-training, backpropagation is used to globally fine-tune the entire network.

[0085] In S3: The covariance function characteristics of the Gaussian signal are used to decouple the static nonlinear module and the dynamic linear module of the Hammerstein model, specifically as follows:

[0086] A zero-mean Gaussian signal is designed, and the corresponding output is obtained from the surrogate model constructed by the sparse deep belief network. Thus, the input and output data of the Gaussian signal are obtained.

[0087] 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) and a constant is equal to the covariance function of the intermediate variable v(t) and the input u(t), i.e.:

[0088] ,

[0089] in: It is a constant. It is a time constant; Indicates input With intermediate variables The cross-covariance function, It's a loss enter The autocovariance function.

[0090] use replace This allows for the decoupling of the static nonlinear module and the dynamic linear module in the Hammerstein model.

[0091] In this invention, It consists of a set of zero-mean Gaussian signals, that is, after passing through the Hammerstein model, the output is... Due to the properties of the covariance function It can be seen that the intermediate variable information of the Hammerstein model The problem of immeasurability was solved, that is, by utilizing... Replace the unknown covariance matrix This achieves decoupling between the static nonlinear module and the linear module.

[0092] like Figure 1 A wind power generation system model is established using the Hammerstein model. The Hammerstein model is as follows:

[0093] ,

[0094] ,

[0095] in, As input to the Hammerstein model, in a wind power generation system, This represents variables related to the inputs of a wind power generation system, such as wind speed.

[0096] f(·) represents a static nonlinear block, i.e., a sparse deep belief network.

[0097] These are intermediate variables after passing through the static nonlinear module, used to characterize the static nonlinear characteristics of the wind power generation system.

[0098] The output of the Hammerstein model, in a wind power generation system, This represents variables related to the output of a wind power generation system, such as power.

[0099] This is the denominator polynomial of the dynamic linear module, used to describe the dynamic characteristics of the wind power generation system.

[0100] The numerator polynomial of the dynamic linear module works together with the denominator polynomial to describe the dynamic characteristics of the system.

[0101] This represents the unit shift operator, used to construct the dynamic equations of a discrete-time system.

[0102] and The orders of the denominator and numerator polynomials are respectively, and they are predetermined model parameters.

[0103] e(t) is the noise term, representing the unmodeled dynamic or measured noise in the wind power generation system.

[0104] t represents time.

[0105] To improve the model's generalization and robustness, based on the Hammerstein sparse deep belief network model constructed above, the following criterion function with sparsity properties is defined:

[0106] ,

[0107] in, The fitting error of the Hammerstein model structure;

[0108] For data length;

[0109] For a specific moment;

[0110] The actual power at time t;

[0111] The power is estimated for identification at time t; A pre-set, smaller threshold; Identify estimated values ​​for intermediate variables after passing through the static nonlinear module;

[0112] The identified static nonlinear mapping function; Let be the input wind speed at time t; The estimated value for the numerator polynomial of the dynamic linear module; is the estimated value of the denominator polynomial for the dynamic linear module; z is the unit shift operator; is the noise term; ESK−DBN is the loss function of the sparse deep belief network; The input wind speed is the estimated value at time t. To support the set; The support set operator for taking the top k largest elements; This represents the state vector of the neurons in the first hidden layer. To support the collection The complement of the set corresponds to the neuron state.

[0113] In S5: The order of ARMA is adaptively calculated to be 2 based on the Hankel matrix and singular value decomposition technique. The order of the ARMA model is adaptively determined using the Hankel matrix and singular value decomposition technique as follows:

[0114] Construct the Hankel matrix as follows:

[0115] ,

[0116] in It is an impulse response.

[0117] Perform singular value decomposition on the Hankel matrix:

[0118] ,

[0119] Where U and V are orthogonal matrices. .

[0120] The system order is determined by calculating the ratio of adjacent singular values:

[0121] ,

[0122] ,

[0123] The expression for identifying the ARMA model using correlation analysis is as follows:

[0124] ,

[0125] Multiply both sides of the formula And by calculating the expected value, we get:

[0126] ,

[0127] in, , .

[0128] Based on the independence between the input signal and the noise signal, we can obtain Therefore, we can obtain:

[0129] ,

[0130] ,

[0131] in .

[0132] ,

[0133] in , , .

[0134] The final result is:

[0135] ,

[0136] ,

[0137] Sparse deep belief networks, through their multi-layered stacked restricted Boltzmann machine structure and the nonlinear mapping capability of Gaussian kernel functions, 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 training process of this network consists of two stages: unsupervised pre-training and supervised fine-tuning.

[0138] The contrastive divergence algorithm is used for unsupervised pre-training, and the parameters... The parameter update formula is:

[0139] ,

[0140] ,

[0141] ,

[0142] in, For learning rate, This represents the initial state of the i-th visible layer neuron. After one step of Gibbs sampling, the reconstructed state of the i-th visible layer neuron. It is the product of the initial visible layer activation probabilities and the corresponding hidden layer activation probabilities. Reconstruct the product of the activation probabilities of the visible layer and the corresponding hidden layer. This is the momentum term for the visible layer bias.

[0143] The backpropagation algorithm is used to perform global fine-tuning of the entire network. For the fine-tuning process, the loss function is defined as follows:

[0144] ,

[0145] Then, using the backpropagation algorithm for fine-tuning, the gradient between the output layer and the hidden layer can be obtained as follows:

[0146] ,

[0147] Similarly, the gradient between hidden layers is:

[0148] ,

[0149] Therefore, the weight update formula is:

[0150] ,

[0151] Where η' represents the learning rate of the BP algorithm, This represents the estimated output of the inverse of the linear module. This represents the output of the i-th neuron in the (r-1)-th layer. The error term of the j-th output neuron, The actual output value of the j-th neuron in the output layer.

[0152] Sparse deep belief networks, through their multi-layered sparse restricted Boltzmann machine stacked structure and the nonlinear mapping capability of Gaussian kernel functions, 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 network's performance is highly dependent on the configuration of hyperparameters such as the number of neurons in its hidden layers, sparsity parameters, and kernel width. However, due to the complex network structure and high-dimensional nonconvex parameter space, traditional gradient descent algorithms are prone to getting trapped in local optima and suffer from low training efficiency.

[0153] To address the aforementioned issues, this invention employs a random forest error correction method to compensate for the prediction bias of the sparse deep belief network Hammerstein model. The core concept of this method is to construct a random forest regressor, using the initial predicted power from the Hammerstein model, historical wind speed data, and historical prediction errors as input features, and learning the nonlinear mapping relationship between the prediction residuals and the features. By integrating the prediction results of multiple decision trees, the random forest can effectively capture complex error patterns, while also exhibiting strong resistance to overfitting and robustness to outliers. This method does not require a pre-set error model and can adaptively correct system biases, thereby effectively improving the accuracy and stability of the final power prediction.

[0154] In S6, evaluation metrics are used to perform error analysis on the prediction results to obtain the optimal parameters, achieving accurate power prediction. Random forest error correction is then applied to the predicted power.

[0155] The initial prediction power of the Hammerstein model for the k-th sample is: The actual power is Calculate the predicted residuals:

[0156] ,

[0157] The training sample set for constructing the random forest correction model is used, with the input features being the feature vectors that influence the error. The output is the target residual. The input features include: the current wind speed. Historical predictions of the Hammerstein model , ..., historical residuals , ...and operating characteristics.

[0158] The random forest regression algorithm is used to learn the nonlinear mapping relationship between input features and residuals, resulting in a residual prediction model:

[0159] ,

[0160] in, This represents a random forest regressor composed of multiple decision trees. This represents the predicted residual value.

[0161] The final corrected power prediction value is:

[0162] ,

[0163] The training objective of random forest is to minimize the corrected mean squared error.

[0164] ,

[0165] Where N is the total number of training samples.

[0166] The invention will be further described below in conjunction with the parameters.

[0167] The Hammerstein model identification based on the wind power generation system model is as follows:

[0168] (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 working 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.

[0169] (2) A quartile data cleaning technique involving horizontal and vertical dimensions was used to eliminate outlier data points, and a dataset consisting of 3,626 data pairs involving wind speed and power was finally obtained.

[0170] (3) A surrogate model of the wind power generation system was established using a sparse deep belief network. The number of layers in the first and second hidden layers were 20 and 10, respectively. The sparsity ρ value in the sparse deep belief network was then set to 0.125. The surrogate model was trained using data from 3300 sets of wind power generation systems, with 1000 iterations. The training model had the smallest mean square error, which was 8.964E3.

[0171] (4) Gaussian signal in this invention Given 1000 Gaussian signals with a mean of 0 and a variance of 0.3, the trained surrogate model outputs the following: 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 correlation analysis, i.e., [a1, a2, b1, b2] = [0.8385, 0.5064, -1.3443, -0.6040].

[0172] (5) A static nonlinear module of the Hammerstein model was built using a sparse deep belief network, and a dynamic linear module of the Hammerstein model was built using ARMA. The network model was trained using 3300 sets of data, and the remaining 326 sets of data were used for prediction. The number of neurons in the first hidden layer and the second hidden layer was set to h1=20 and h2=10, respectively, and the sparsity ρ value in the network was set to 0.125. The mean squared error obtained from the training was the smallest, with an error of 1.3196E5 (kW).

[0173] 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 wind power system power prediction method integrating a sparse deep belief network Hammerstein model and error correction, characterized in that: Includes the following steps: S1: Obtain wind speed and power data from the wind power generation system; S2: Use the quartile method to clean outliers from wind speed and power data; S3: Based on the wind speed and power data after cleaning, a Gaussian kernel function is used to map the original input space to a high-dimensional feature space, and a proxy model of the wind power generation system constructed by a sparse deep belief network is established; a zero-mean Gaussian signal is designed and input into the proxy model to obtain the corresponding Gaussian signal output; S4: Describe the dynamic linear module of the Hammerstein model using an autoregressive moving average model, and describe the static nonlinear module of the Hammerstein model using a sparse deep belief network, thus establishing the Hammerstein model structure. S5: Adaptively calculate the order of the autoregressive moving average model using the Hankel matrix and singular value decomposition technique. Based on the input and output data of the zero-mean Gaussian signal, identify the parameters of the autoregressive moving average model using correlation analysis. Based on wind speed and power data, update the weights and biases of the sparse deep belief network using the contrastive divergence algorithm and backpropagation algorithm to obtain the optimal parameters. S6: The updated sparse deep belief network is used as the static nonlinear module of the Hammerstein model structure, and the identified autoregressive moving average model is used as the dynamic linear module of the Hammerstein model structure. The power is predicted using the Hammerstein model structure to obtain the predicted power, and the error is corrected by the random forest algorithm.

2. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S2, the quartile method includes vertical data cleaning based on wind speed data grouping and horizontal data cleaning based on power data grouping.

3. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S3, the Gaussian kernel function is defined as: ， in, , For the i-th and j-th input sample vectors, The squared Euclidean distance between the two samples is given. For kernel width parameter, This is the kernel function value.

4. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S3, the covariance function characteristic of the zero-mean Gaussian signal is used to decouple the static nonlinear module and the dynamic linear module of the Hammerstein model structure. The covariance function characteristic is expressed as: ， in, It is a constant. It is a time constant; For input With intermediate variables The cross-covariance function; Let u(t) be the autocovariance function of the input u(t).

5. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S3, the sparse deep belief network includes a multi-layer neural network structure. Each layer is pre-trained in an unsupervised manner using the contrastive divergence algorithm, and after pre-training, the backpropagation algorithm is used for global fine-tuning.

6. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S5, the order of the autoregressive moving average model is adaptively calculated, specifically by determining the order by calculating the ratio of adjacent singular values.

7. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S5, the parameters of the autoregressive moving average model are identified, including the denominator and numerator polynomial coefficients of the dynamic linear module.

8. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S6, the random forest algorithm is used to correct the error of the predicted power, including building a training sample set for the random forest correction model. The input features of the training sample set include the current wind speed, historical predictions of the Hammerstein model structure, historical residuals, and operating conditions.

9. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S6, error correction includes: Calculate the predicted residuals: ， The random forest regression algorithm is used to learn the nonlinear mapping relationship between the input features and the predicted residuals, resulting in a residual prediction model: ， The final corrected power prediction value is: ， in, For the first The actual power of each sample The initial predicted power for the Hammerstein model structure is... This represents a random forest regressor composed of multiple decision trees. The predicted residual value; The training objective of random forest is to minimize the corrected mean squared error. ， Where N is the total number of training samples.

10. The wind power system power prediction method integrating the sparse deep belief network Hammerstein model and error correction as described in claim 1, characterized in that: In S5, the optimal parameters are obtained, including defining a criterion function with sparsity properties. The criterion function is as follows: ， in, The fitting error of the Hammerstein model structure; For data length; For a specific moment; The actual power at time t; The power is estimated for identification at time t; A pre-set, smaller threshold; Identify estimated values ​​for intermediate variables after passing through the static nonlinear module; The identified static nonlinear mapping function; Let be the input wind speed at time t; The estimated value for the numerator polynomial of the dynamic linear module; is the estimated value of the denominator polynomial for the dynamic linear module; z is the unit shift operator; is the noise term; ESK−DBN is the loss function of the sparse deep belief network; The input wind speed is the estimated value at time t; To support the set; The support set operator for taking the top k largest elements; This represents the state vector of the neurons in the first hidden layer. To support the collection The complement of the set corresponds to the neuron state.

Images