A wind power prediction method across wind farms
By adopting an adaptive wind power prediction method across wind farms and utilizing the collaborative architecture of Token-Embedding and Mamba modules, the problems of large prediction errors and high computational complexity across wind farms are solved, achieving high-precision and low-cost wind power prediction, which is suitable for wind farm clusters.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTH CHINA UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2025-06-04
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies cannot adapt to wind speed differences in different geographical and climatic zones, resulting in large prediction errors across wind fields. Furthermore, existing models require structural reconstruction in cross-wind field scenarios, leading to high computational complexity and making it difficult to achieve joint training across multiple wind fields.
A cross-wind field adaptive wind power prediction method is adopted. Through the collaborative architecture of the Token-Embedding layer and the Mamba module, a unified spatial mapping of data from different wind fields is achieved. A dual-path fusion mechanism is used to combine the first and second Mamba modules for time series information modeling, thereby reducing computational complexity and improving prediction accuracy.
It significantly improves the accuracy of prediction results across wind farms, reduces deployment by 75%, lowers computational complexity, is suitable for large-scale applications in wind farm clusters, and has stronger adaptability and long-range dependency capture capabilities.
Smart Images
Figure CN120579672B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind field data processing technology, specifically relating to a cross-wind field adaptive wind power prediction method. Background Technology
[0002] Existing multi-wind field joint forecasting technologies are mainly divided into statistical models, traditional machine learning models, and deep learning models. However, applying existing technologies to cross-wind field forecasting still faces significant challenges, including:
[0003] ① Statistical models (such as ARIMA) rely on linear assumptions and can only model through linear combinations of historical sequences, failing to capture the nonlinear mapping relationship between wind speed and power. In addition, traditional ARIMA models are trained based on single wind field data and lack feature transfer mechanisms across wind fields, which significantly increases prediction errors when applied to different wind fields.
[0004] ② Traditional machine learning models (such as SVM) rely heavily on manual feature engineering, and their processing of time-series data is "independent and non-coherent". They cannot model dependencies across time steps. Furthermore, in cross-wind field scenarios, the feature correlation distribution of different wind fields is inconsistent, which means that the model needs to readjust the feature engineering logic when adding new wind fields, resulting in high migration costs.
[0005] ③ In deep learning models, LSTM-type models adopt a unidirectional recursive structure. In cross-wind field scenarios, the single-direction temporal modeling cannot adapt to the dynamic changes of different wind fields, leading to an explosion of errors in cross-field prediction. Transformer-type models are based on the self-attention mechanism, with a computational complexity of O(N²) (N is the sequence length). When inputting large-scale data from multiple wind fields, the memory usage and computational cost will increase, making it difficult to achieve joint training of multiple wind fields.
[0006] Therefore, the core defects of the existing technology can be summarized as follows: the model cannot adapt to the differences in wind speed in different geographical and climatic zones; the model structure needs to be reconstructed when the model is used for new wind fields. In view of this, this invention proposes a cross-wind field adaptive wind power prediction method. Summary of the Invention
[0007] To address the aforementioned shortcomings in existing technologies, this invention provides a cross-wind field adaptive wind power prediction method that solves the problem that traditional methods cannot adapt to wind speed differences in different geographical and climatic zones.
[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a cross-wind field adaptive wind power prediction method, comprising the following steps:
[0009] S1. Obtain cross-wind field data and preprocess the cross-wind field data;
[0010] S2. Input the cleaned cross-wind field data into the encoder embedding layer to obtain the encoded features;
[0011] S3. Input the encoded features into the second Mamba module to perform time series information modeling and generate modeling features;
[0012] S4. Input the modeling features into the fully connected layer to obtain the wind power prediction results.
[0013] Furthermore, in S1, the preprocessing method is as follows: the cross-wind field data is cleaned using the quartile method and K-means, the missing values of the cross-wind field data are filled using an interpolation model based on the attention mechanism, and the cross-wind field data is normalized by Z-score. The cross-wind field data includes wind speed, wind direction and temperature.
[0014] 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 flow attention mechanism module, a feedforward network, and a residual connection and layer normalization layer connected in sequence.
[0015] Furthermore, in S2, the method for obtaining the encoded features is as follows:
[0016] S21. Input the cleaned cross-wind field data into the Token-Embedding layer, project the input data into a unified feature space and add position encoding to generate intermediate encoding. The intermediate encoding is divided according to the channel dimension. 50% of the intermediate encoding is input into the first Mamba module to obtain the temporal feature vector. The remaining intermediate encoding is combined with the temporal feature vector to obtain the initial encoding features.
[0017] S22. Input the initial coding features into the multilayer encoder structure to obtain the coding features.
[0018] Further: In S21, the initial coding features are obtained. The specific expression is:
[0019]
[0020] In the formula, For the encoder's input embedding module, For the location encoding module, This refers to cross-wind field data after data cleaning;
[0021] In S22, the first layer of the multi-layer encoder structure l Hidden state matrix computed by the layer encoder layer The specific expression is:
[0022]
[0023] In the formula, For encoding units, , This represents the number of encoder layers.
[0024] 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;
[0025] The custom S6 module processes the input sequence through linear transformation and discretization. x Dynamic parameters B, C, and Δ are generated through linear transformation;
[0026] The specific expressions for the formulas that generate dynamic parameters B and C are as follows:
[0027]
[0028] In the formula, To be compatible with input Related matrices , used to generate dynamic parameters B and C;
[0029] Through activation function From parameters and Calculate the dynamic parameter Δ in the middle;
[0030]
[0031]
[0032] In the formula, The matrix associated with the input is used to compute the dynamic parameters. This is a linear transformation function used to generate the matrices related to the dynamic parameters B and C. This is a broadcast function used to perform dimensional expansion or adaptation operations on the result of a linear transformation. Δ is the activation function used to calculate dynamic parameters.
[0033] The beneficial effects of the above-mentioned further solutions are as follows: 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.
[0034] The beneficial effects of this invention are as follows:
[0035] (1) This invention provides a cross-wind field adaptive wind power prediction method, proposes a collaborative architecture of Token-Embedding and Mamba, realizes unified spatial mapping of different wind field data through encoder embedding layer, and adopts dual-path fusion. A part of the feature vector output by Token-Embedding layer is directly input into Mamba module to retain the original wind field characteristics; another part is combined with the feature after Token-Embedding mapping and the part through Mamba module, effectively eliminating the distribution offset between wind fields, and significantly improving the accuracy of prediction results of cross-wind field data.
[0036] (2) The present invention can learn historical information and future trends at the same time, and has stronger adaptability in complex wind power change modes. It adopts a dual mechanism combining the first Mamba module and the second Mamba module, which provides better long-range dependency capture capability, so that the prediction results can be closer to the real power change trend.
[0037] (3) The performance of the Transformer and Mamba models in the prior art is not as good as that of the model of this invention; the generalization ability of this invention across wind fields is outstanding, and a single model can cover 4 wind fields, reducing the deployment by 75%.
[0038] (4) The present invention adopts the Flow-Attention+Mamba architecture. The competition mechanism based on network flow theory will reduce the computational complexity from O(N²) to O(N). The joint training mode reduces the cost of manual operation and maintenance, and is suitable for large-scale application of wind farm clusters. Attached Figure Description
[0039] Figure 1 This is a flowchart of a cross-wind field adaptive wind power prediction method according to the present invention. Detailed Implementation
[0040] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0041] like Figure 1 As shown, in one embodiment of the present invention, a wind power prediction method that adapts across wind fields includes the following steps:
[0042] S1. Obtain cross-wind field data and preprocess the cross-wind field data;
[0043] S2. Input the cleaned cross-wind field data into the encoder embedding layer to obtain the encoded features;
[0044] S3. Input the encoded features into the second Mamba module to perform time series information modeling and generate modeling features;
[0045] S4. Input the modeling features into the fully connected layer to obtain the wind power prediction results.
[0046] In S1, the preprocessing method is as follows: the cross-wind field data is cleaned by the quartile method and K-means, the missing values of the cross-wind field data are filled by the interpolation model based on the attention mechanism, and the cross-wind field data is normalized by Z-score. The cross-wind field data includes wind speed, wind direction and temperature.
[0047] Table 1 Feature Variables
[0048]
[0049] 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 flow attention mechanism module, a feedforward network, and a residual connection and layer normalization layer connected in sequence.
[0050] In this embodiment, the multi-layer encoder structure adopts a multi-head flow-attention mechanism, replacing the traditional Softmax with a competitive mechanism based on network flow theory.
[0051] In S2, the specific method for obtaining the encoded features is as follows:
[0052] S21. Input the cleaned cross-wind field data into the Token-Embedding layer, project the input data into a unified feature space and add position encoding to generate intermediate encoding. The intermediate encoding is divided according to the channel dimension. 50% of the intermediate encoding is input into the first Mamba module to obtain the temporal feature vector. The remaining intermediate encoding is combined with the temporal feature vector to obtain the initial encoding features.
[0053] S22. Input the initial coding features into the multilayer encoder structure to obtain the coding features.
[0054] In this embodiment, the original data is projected onto a unified feature space through the Token-Embedding layer, and position encoding is added using the position encoding method of the position encoding module to inject absolute time information and solve the problem of temporal disorder.
[0055] A portion of the feature vector output from the Token-Embedding layer is directly input into the Mamba module, preserving the original wind field characteristics; another portion, after being mapped by Token-Embedding, is combined with the portion from the Mamba module, addressing both local temporal dependencies and global trends to resolve the issue of cross-wind field data distribution shift.
[0056] In S21, the initial coding features are obtained. The specific expression is:
[0057]
[0058] In the formula, For the encoder's input embedding module, For the location encoding module, This refers to the cross-wind field data after data cleaning. , For the first i Feature variables, N The total number of feature variables is 12 in this embodiment.
[0059] In S22, L The total number of encoder layers; output of a multi-layer encoder structure. The corresponding hidden state matrix is , d The number of feature dimensions corresponding to each feature variable, in the multi-layer encoder structure. l Hidden state matrix computed by the layer encoder layer The specific expression is:
[0060]
[0061] In the formula, For encoding units, , This represents the number of encoder layers.
[0062] 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;
[0063] The specific workflow of the Mamba module is as follows:
[0064] The input data of the A1 and Mamba modules are linearly projected through two parallel linear layers to map the input data to different feature spaces and obtain the results after linear projection.
[0065] A2. Input the linearly projected result into the convolutional layer through the first branch, perform convolution operation on the data to extract local features, and perform nonlinear transformation through the activation function to obtain a feature representation with nonlinear characteristics, and generate the output result of the convolutional layer.
[0066] A3. Input the result after linear projection into the custom S6 module through the second branch. The custom S6 module receives the hidden state of the previous time step and the current input, updates the state inside the module, obtains the hidden state of the current time step, and uses it as the output result of the custom S6 module.
[0067] A4. The output of the convolutional layer is fused with the output of the custom S6 module. The fused result is then further processed through a linear layer to obtain the output of the Mamba module. Simultaneously, residual connections (paths directly connecting the ingress and egress linear layers) exist, allowing the input data of the Mamba module to be directly added to its output. This helps alleviate the vanishing gradient problem and enhances the model's training stability.
[0068] The custom S6 module receives the hidden state from the previous time step and the current input, updates the state using parameters such as the state transition matrix A and the input matrix B, and discretizes the matrix according to the dynamically generated time step Δ to obtain the hidden state at the current time step.
[0069] The custom S6 module is a complex component in the Mamba architecture responsible for processing input sequences through a series of linear transformations and discretization processes. Specifically, the custom S6 module processes the input sequence through linear transformations and discretization processes. Dynamic parameters B, C, and Δ are generated through linear transformation.
[0070] The specific expressions for the formulas that generate dynamic parameters B and C are as follows:
[0071]
[0072] In the formula, To be compatible with input Related matrices , used to generate dynamic parameters B and C;
[0073] In this embodiment, by input Mapping to N-dimensional space yields B and C, where B and C are both tensors of shape (B1, L, N), where B1 is the batch size and B represents the relationship between the input and output. The associated matrix, C, is used to control how the state affects the output.
[0074] Through activation function From parameters and The dynamic parameter Δ is calculated in the following steps. Δ is a tensor of shape (B1, L, D), where D is the feature dimension. Δ controls the selection of the time step and the activation function. This allows the model to selectively remember or ignore information in the input.
[0075]
[0076]
[0077] In the formula, The matrix associated with the input is used to compute the dynamic parameters. This is a linear transformation function used to generate the matrices related to the dynamic parameters B and C. This is a broadcast function used to perform dimensional expansion or adaptation operations on the result of a linear transformation. Δ is the activation function used to calculate dynamic parameters.
[0078] The specific expression for discretization using the custom S6 module is as follows:
[0079]
[0080]
[0081] In the formula, A Let be a matrix of shape (D, N), representing the structured matrix. The parameter N represents the "state dimension" of the selective structured state-space model, which determines the size of the hidden state space inside the SSM. This affects the model's ability to retain information in the time dimension. I is the identity matrix.
[0082] Compared with traditional technologies, the present invention has the effects shown in Table 2.
[0083] Table 2 Comparison of the effects of the present invention and traditional technology
[0084]
[0085] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.
Claims
1. A wind power prediction method that adapts across wind fields, characterized in that, Includes the following steps: S1. Obtain cross-wind field data and preprocess the cross-wind field data; S2. Input the cleaned cross-wind field data into the encoder embedding layer to obtain the encoded features; 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 flow attention mechanism module, a feedforward network, and a residual connection and layer normalization layer connected in sequence. The specific method for obtaining encoded features is as follows: S21. Input the cleaned cross-wind field data into the Token-Embedding layer, project the input data into a unified feature space and add position encoding to generate intermediate encoding. The intermediate encoding is divided according to the channel dimension. 50% of the intermediate encoding is input into the first Mamba module to obtain the temporal feature vector. The remaining intermediate encoding is combined with the temporal feature vector to obtain the initial encoding features. S22. Input the initial coding features into the multilayer encoder structure to obtain the coding features; S3. Input the encoded features into the second Mamba module to perform time series 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 wind power prediction method with cross-wind field adaptive capability according to claim 1, characterized in that, In S1, the preprocessing method is as follows: the cross-wind field data is cleaned by the quartile method and K-means, the missing values of the cross-wind field data are filled by the interpolation model based on the attention mechanism, and the cross-wind field data is normalized by Z-score. The cross-wind field data includes wind speed, wind direction and temperature.
3. The cross-wind field adaptive wind power prediction method according to claim 1, characterized in that, In S21, the initial coding features are obtained. The specific expression is: In the formula, For the encoder's input embedding module, For the location encoding module, This refers to cross-wind field data after data cleaning; In S22, the first layer of the multi-layer encoder structure l Hidden state matrix computed by the layer encoder layer The specific expression is: In the formula, For encoding units, , This represents the number of encoder layers.
4. The wind power prediction method with cross-wind field adaptive capability according to claim 1, characterized in that, 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; The custom S6 module processes the input sequence through linear transformation and discretization. x Dynamic parameters B, C, and Δ are generated through linear transformation; The specific expressions for the formulas that generate dynamic parameters B and C are as follows: In the formula, To be compatible with input Related matrices , used to generate dynamic parameters B and C; Through activation function From parameters and Calculate the dynamic parameter Δ in the middle; In the formula, The matrix associated with the input is used to compute the dynamic parameters. This is a linear transformation function used to generate the matrices related to the dynamic parameters B and C. This is a broadcast function used to perform dimensional expansion or adaptation operations on the result of a linear transformation. This is the Δ activation function used to calculate dynamic parameters.