Cross-wind-field adaptive wind power prediction method

Abstract

The invention discloses a cross-wind-field adaptive wind power prediction method, which comprises the following steps: S1, acquiring cross-wind-field data, and preprocessing the cross-wind-field data; s2, inputting the cross-wind-field data after data cleaning into an encoder embedding layer to obtain encoding features; s3, the coding features are input into a Mamba module for time sequence information modeling, and modeling features are generated; and S4, inputting the modeling features into a full connection layer to obtain a wind power prediction result. According to the method, a Token-Embedding and Mamba collaborative architecture is provided, unified space mapping of different wind field data is realized through an encoder embedding layer, double-path fusion is adopted, distribution offset between wind fields is effectively eliminated, the accuracy of a prediction result of cross-wind field data is remarkably improved, historical information and future trend can be learned at the same time, and the method has the advantages of being high in practicability and easy to popularize. And the method has higher adaptability in a complex wind power change mode.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind farm data processing, and in particular relates to a cross-wind farm adaptive wind power prediction method. Background Art

[0002] Existing multi-wind farm joint prediction technologies are mainly divided into statistical models, traditional machine learning models, and deep learning models. However, applying existing technologies to cross-wind farm prediction still faces great challenges. Specific issues include: Statistical models (such as ARIMA) rely on linear assumptions and can only be modeled through linear combinations of historical series, failing to capture the nonlinear mapping relationship between wind speed and power. Furthermore, traditional ARIMA models are trained on data from a single wind farm and lack a mechanism for cross-farm feature transfer. This significantly increases prediction errors when applied to different wind farms.

[0003] ② Traditional machine learning models (such as SVM) are highly dependent on manual feature engineering, and their processing of time series data is "independent and incoherent". They are unable to model dependencies across time steps. In addition, in cross-wind farm scenarios, the feature correlation distribution of different wind farms is inconsistent, resulting in the need to readjust the feature engineering logic when adding new wind farms, resulting in high migration costs.

[0004] ③ In deep learning models, LSTM models use a unidirectional recursive structure. In cross-wind farm scenarios, single-directional time series modeling cannot adapt to the dynamic changes in different wind farms, resulting in significant errors in cross-farm predictions. Transformer models, based on a self-attention mechanism, have a computational complexity of O(N²) (N is the sequence length). When inputting large-scale data from multiple wind farms, memory usage and computational costs increase, making joint training across multiple wind farms difficult.

[0005] Therefore, the core defects of existing technologies can be summarized as follows: the model cannot adapt to the wind speed differences in different geographical and climatic zones; the model structure needs to be reconstructed when the model is used for a new wind farm. In view of this situation, the present invention proposes a wind power prediction method that is adaptive across wind farms. Summary of the Invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a cross-wind farm adaptive wind power prediction method that solves the problem that the model used in the traditional method cannot adapt to the wind speed differences in different geographical climate zones.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a wind power prediction method adaptively across wind farms, comprising the following steps: S1. Acquire cross-wind farm data and pre-process the cross-wind farm data; S2. Input the cleaned cross-wind farm data into the encoder embedding layer to obtain the encoding features; S3. Input the encoded features into the second Mamba module to perform temporal information modeling and generate modeling features; S4. Input the modeling features into the fully connected layer to obtain the wind power prediction results.

[0008] Furthermore: In S1, the preprocessing method is specifically as follows: the cross-wind farm data is cleaned by the quartile method and K-means, the missing values ​​of the cross-wind farm data are filled by an interpolation model based on the attention mechanism, and the cross-wind farm data is Z-score normalized. The cross-wind farm data includes wind speed, wind direction and temperature.

[0009] Furthermore: in S2, the encoder embedding layer includes a Token-Embedding layer, a first Mamba module and a multi-layer encoder structure connected in sequence. The multi-layer encoder structure includes several encoder layers connected in sequence. Each encoder layer includes a multi-head stream attention mechanism module, a feedforward network and a residual connection and a layer normalization layer connected in sequence.

[0010] Furthermore: In S2, the method for obtaining the coding feature is specifically: S21. Input the cleaned cross-wind farm data into the Token-Embedding layer, project the input data into a unified feature space and add position encoding to generate intermediate codes. The intermediate codes are divided by channel dimension, and 50% of the intermediate codes are input into the first Mamba module to obtain time series feature vectors. The remaining intermediate codes are combined with the time series feature vectors to obtain the initial coding features. S22. Input the initial coding features into the multi-layer encoder structure to obtain coding features.

[0011] Further: In S21, the initial coding feature is obtained The specific expression is:

[0012] Where, is the input embedding module of the encoder, is the position encoding module, This is the cross-wind farm data after data cleaning; In S22, the first l The hidden state matrix calculated by the encoder layer The specific expression is:

[0013] Where, is the coding unit, , is the number of encoder layers.

[0014] Furthermore: In S3, the first Mamba module and the second Mamba module have the same structure, both including linear projection, convolution, activation function, custom S6 module and residual connection; Among them, the custom S6 module processes the input sequence through linear transformation and discretization process. x Generate dynamic parameters B, C and Δ through linear transformation; The specific expressions of the formula for generating dynamic parameters B and C are:

[0015] Where, For input The associated matrix, , used to generate dynamic parameters B and C; Through the activation function From the parameters and Calculate the dynamic parameter Δ;

[0016]

[0017] Where, is the matrix associated with the input used to calculate the dynamic parameters, is the linear transformation function used to generate the matrix related to the dynamic parameters B and C, is a broadcast function used to perform dimension expansion or adaptation operations on the results of linear transformation. is the Δ activation function used to calculate the dynamic parameters.

[0018] The beneficial effects of the above further solution are: the present invention innovatively designs the dynamic parameterization of the S6 module and the selective state space mechanism, which breaks through the problem of error accumulation in traditional models.

[0019] The beneficial effects of the present invention are: (1) The present invention provides a cross-wind farm adaptive wind power prediction method, proposes a collaborative architecture of Token-Embedding and Mamba, realizes unified spatial mapping of different wind farm data through the encoder embedding layer, and adopts dual-path fusion. A part of the feature vector output by the Token-Embedding layer is directly input into the Mamba module to retain the original wind farm characteristics; the other part of the feature vector after Token-Embedding mapping is combined with the part that passes through the Mamba module, effectively eliminating the distribution offset between wind farms, and significantly improving the accuracy of the prediction results of cross-wind farm data.

[0020] (2) The present invention can learn historical information and future trends simultaneously, and has stronger adaptability under complex wind power change patterns. It adopts a dual mechanism combining the first Mamba module and the second Mamba module, providing better long-range dependency capture capabilities, so that the prediction results can be closer to the actual power change trend.

[0021] (3) The performance of the Transformer and Mamba models in the existing technologies is not as good as that of the model of the present invention. The generalization capability of the present invention across wind farms is outstanding, and a single model can cover 4 wind farms, reducing the deployment volume by 75%.

[0022] (4) This invention adopts the Flow-Attention+Mamba architecture. The competition mechanism based on network flow theory reduces the computational complexity from O(N²) to O(N). The joint training mode reduces the manual operation and maintenance costs, making it suitable for large-scale application of wind farm clusters. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of a cross-wind farm adaptive wind power prediction method of the present invention. DETAILED DESCRIPTION

[0024] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0025] like Figure 1 As shown, in one embodiment of the present invention, a cross-wind farm adaptive wind power prediction method includes the following steps: S1. Acquire cross-wind farm data and pre-process the cross-wind farm data; S2. Input the cleaned cross-wind farm data into the encoder embedding layer to obtain the encoding features; S3. Input the encoded features into the second Mamba module to perform temporal information modeling and generate modeling features; S4. Input the modeling features into the fully connected layer to obtain the wind power prediction results.

[0026] In S1, the preprocessing method is as follows: the cross-wind farm data is cleaned by quartile method and K-means, the missing values ​​of the cross-wind farm data are filled by the interpolation model based on the attention mechanism, and the cross-wind farm data is Z-score normalized. The cross-wind farm data includes wind speed, wind direction and temperature.

[0027] Table 1 Characteristic variables

[0028] In S2, the encoder embedding layer includes a Token-Embedding layer, a first Mamba module, and a multi-layer encoder structure connected in sequence. The multi-layer encoder structure includes several encoder layers connected in sequence. Each encoder layer includes a multi-head stream attention mechanism module, a feedforward network, a residual connection, and a layer normalization layer connected in sequence.

[0029] In this embodiment, the multi-layer encoder structure adopts a multi-head flow attention mechanism (Flow-Attention), replacing the traditional Softmax with a competition mechanism based on network flow theory.

[0030] In S2, the method for obtaining the encoding feature is as follows: S21. Input the cleaned cross-wind farm data into the Token-Embedding layer, project the input data into a unified feature space and add position encoding to generate intermediate codes. The intermediate codes are divided by channel dimension, and 50% of the intermediate codes are input into the first Mamba module to obtain time series feature vectors. The remaining intermediate codes are combined with the time series feature vectors to obtain the initial coding features. S22. Input the initial coding features into the multi-layer encoder structure to obtain coding features.

[0031] In this embodiment, the original data is projected into a unified feature space through the Token-Embedding layer, and the position encoding method of the position encoding module is used to add position encoding and inject absolute time information to solve the problem of time sequence disorder.

[0032] A portion of the feature vectors output by the Token-Embedding layer is directly input into the Mamba module, preserving the original wind farm characteristics. The remaining features, mapped by Token-Embedding, are combined with those that pass through the Mamba module, focusing on both local temporal dependencies and global trends to address the problem of data distribution offset across wind farms.

[0033] In S21, the initial coding features are obtained The specific expression is:

[0034] Where, is the input embedding module of the encoder, is the position encoding module, This is the cross-wind farm data after data cleaning. , For the i characteristic variables, Nis the total number of characteristic variables, which is 12 in this embodiment.

[0035] In S22, L is the total number of encoder layers, and the multi-layer encoder structure outputs The corresponding hidden state matrix is , d The number of feature dimensions corresponding to each feature variable, the first l The hidden state matrix calculated by the encoder layer The specific expression is:

[0036] Where, is the coding unit, , is the number of encoder layers.

[0037] In S3, the structure of the first Mamba module and the second Mamba module is the same, both including linear projection, convolution, activation function, custom S6 module and residual connection; The workflow of the Mamba module is as follows: The input data of the A1 and Mamba modules are linearly projected through two parallel linear layers to map the input data into different feature spaces and obtain the results after linear projection; A2: Input the linear projection result into the convolution layer through the first branch, perform convolution operation on the data to extract local features, and perform nonlinear transformation through activation function to obtain feature representation with nonlinear characteristics, and generate the output result of the convolution layer; A3. The result of linear projection is input into the custom S6 module through the second branch. The custom S6 module receives the hidden state of the previous moment and the current input, updates the state internally, and obtains the current hidden state, which is used as the output result of the custom S6 module. The output of the A4 convolutional layer is fused with the output of the custom S6 module. The fused result is then processed through a linear layer to produce the output of the Mamba module. Furthermore, a residual connection (a path directly connecting the input linear layer to the output linear layer) is implemented, allowing the Mamba module's input data to be directly added to its output. This helps alleviate the vanishing gradient problem and enhances model training stability.

[0038] Among them, the custom S6 module: receives the hidden state of the previous moment and the current input, updates the state using parameters such as the state transfer matrix A and the input matrix B, and discretizes the matrix according to the dynamically generated time step Δ dynamic parameter to obtain the hidden state at the current moment; The custom S6 module is a complex component in the Mamba architecture, responsible for processing the input sequence through a series of linear transformations and discretization processes. The dynamic parameters B, C and Δ are generated through linear transformation.

[0039] The specific expressions of the formula for generating dynamic parameters B and C are:

[0040] Where, For input The associated matrix, , used to generate dynamic parameters B and C; In this embodiment, by inputting Mapping to N-dimensional space to get B and C, B and C are both tensors of shape (B1, L, N), B1 is the batch size, B represents the input The associated matrix, C, is used to control how the state affects the output.

[0041] Through the activation function From the parameters and The dynamic parameter Δ is calculated in the , which is a tensor with a shape of (B1, L, D), where D is the feature dimension. The dynamic parameter Δ is used to control the selection of the time step, the activation function This enables the model to selectively remember or ignore information in the input.

[0042]

[0043]

[0044] Where, is the matrix associated with the input used to calculate the dynamic parameters, is the linear transformation function used to generate the matrix related to the dynamic parameters B and C, is a broadcast function used to perform dimension expansion or adaptation operations on the results of linear transformation. is the Δ activation function used to calculate the dynamic parameters.

[0045] The specific expression for discretization of the custom S6 module is:

[0046]

[0047] Where, Ais a matrix of shape (D, N), which represents a structured matrix. The parameter N represents the "state dimension" of the selective structure state space model, which determines the size of the hidden state space inside the SSM, which affects the model's ability to retain information in the time dimension. I is the unit matrix.

[0048] Compared with the conventional technology, the present invention has the effects shown in Table 2.

[0049] Table 2 Comparison of the effects of the present invention and traditional technology

[0050] In the description of the present invention, it should be understood that the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of such features.

Claims

1. A wind power prediction method that is adaptive across wind farms, characterized in that: The following steps are involved: S1. Acquire cross-wind farm data and pre-process the cross-wind farm data; S2. Input the cleaned cross-wind farm data into the encoder embedding layer to obtain the encoding features; S3. Input the encoded features into the second Mamba module to perform temporal information modeling and generate modeling features; S4. Input the modeling features into the fully connected layer to obtain the wind power prediction results.

2. The cross-wind farm adaptive wind power prediction method according to claim 1, characterized in that: In S1, the preprocessing method is as follows: the cross-wind farm data is cleaned by quartile method and K-means, the missing values ​​of the cross-wind farm data are filled by the interpolation model based on the attention mechanism, and the cross-wind farm data is Z-score normalized. The cross-wind farm data includes wind speed, wind direction and temperature.

3. The cross-wind farm adaptive wind power prediction method according to claim 1, characterized in that: In S2, the encoder embedding layer includes a Token-Embedding layer, a first Mamba module, and a multi-layer encoder structure connected in sequence. The multi-layer encoder structure includes several encoder layers connected in sequence. Each encoder layer includes a multi-head stream attention mechanism module, a feedforward network, a residual connection, and a layer normalization layer connected in sequence.

4. The cross-wind farm adaptive wind power prediction method according to claim 3, characterized in that: In S2, the method for obtaining the encoding feature is as follows: S21. Input the cleaned cross-wind farm data into the Token-Embedding layer, project the input data into a unified feature space and add position encoding to generate intermediate codes. The intermediate codes are divided by channel dimension, and 50% of the intermediate codes are input into the first Mamba module to obtain time series feature vectors. The remaining intermediate codes are combined with the time series feature vectors to obtain the initial coding features. S22. Input the initial coding features into the multi-layer encoder structure to obtain coding features.

5. The cross-wind farm adaptive wind power prediction method according to claim 4, characterized in that: In S21, the initial coding features are obtained The specific expression is: Where, is the input embedding module of the encoder, is the position encoding module, This is the cross-wind farm data after data cleaning; In S22, the first l The hidden state matrix calculated by the encoder layer The specific expression is: Where, is the coding unit, , is the number of encoder layers.

6. The cross-wind farm adaptive wind power prediction method according to claim 5, characterized in that: In S3, the structure of the first Mamba module and the second Mamba module is the same, both including linear projection, convolution, activation function, custom S6 module and residual connection; Among them, the custom S6 module processes the input sequence through linear transformation and discretization process. x Generate dynamic parameters B, C and Δ through linear transformation; The specific expressions of the formula for generating dynamic parameters B and C are: Where, For input The associated matrix, , used to generate dynamic parameters B and C; Through the activation function From the parameters and Calculate the dynamic parameter Δ; Where, is the matrix associated with the input used to calculate the dynamic parameters, is the linear transformation function used to generate the matrix related to the dynamic parameters B and C, is a broadcast function used to perform dimension expansion or adaptation operations on the results of linear transformation. is the Δ activation function used to calculate the dynamic parameters.

Images