Fan power prediction method based on multistage decomposition and cross-domain augmentation

Abstract

The invention discloses a fan power prediction method based on multistage decomposition and cross-domain augmentation, and the method comprises the steps: firstly, obtaining a fan power data set and a local weather data set, screening related feature data, carrying out the preprocessing, and dividing the data into a training set and a test set; secondly, a TAform model used for fan power prediction is constructed; thirdly, the TAform model is trained by means of the power of a draught fan and historical environment data, and the TAform model after training is completed is obtained; and finally, obtaining to-be-tested data, and obtaining a prediction result of the to-be-tested data by using the trained TAform model. According to the method, the problem that the complex dependency relationship among multiple variables is difficult to comprehensively capture, especially the bottleneck problem of processing complex time modes and time information utilization in the existing method is solved.

Description

Technical Field

[0001] The present invention belongs to the field of wind power prediction, and in particular relates to a wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation. Background Art

[0002] Wind turbine power forecasting (WPF) technology is increasingly being used in the wind power sector, particularly to improve wind turbine operation and maintenance efficiency and grid dispatch accuracy. Traditionally, by analyzing historical wind turbine data, short-term forecasts of wind turbine output power can be achieved, providing a basis for real-time monitoring and response. However, with the increasing demand for wind power system optimization, relying solely on wind turbine data is clearly insufficient to support more complex decision-making needs. In particular, accurate analysis of external factors such as weather has become crucial in predictive maintenance of wind turbines and grid dispatch warnings. Meteorological conditions have a significant impact on wind power and wind turbine performance. Therefore, incorporating weather characteristics into prediction models can help improve the accuracy and reliability of predictions.

[0003] Although there are currently a variety of time series-based forecasting models, these models are primarily focused on single-step forecasting and are suitable for real-time monitoring and immediate dispatching. However, multi-step forecasting is particularly important for the long-term operation and management of wind farms, especially in planning maintenance strategies, optimizing operation and maintenance costs, and enhancing grid stability. Multi-step forecasting can predict wind turbine power output over a period of time in the future. This not only helps improve the cost-effective management of wind turbines, but also provides early warning of possible grid load fluctuations in the future, providing more basis for dispatching decisions. The biggest challenge in achieving accurate multi-step forecasting is how to maintain the accuracy and stability of the forecast over a long time span. Therefore, multi-step forecasting of wind turbine power is not only a technical challenge, but also the key to the wind power industry's efforts to improve intelligent operation and maintenance, dispatch optimization, and other aspects.

[0004] Existing wind turbine power prediction models typically process power time series data using physical models, statistics, and machine learning methods. These methods, such as artificial neural networks (ANNs), support vector machines (SVMs), and Gaussian processes, have limitations when processing long sequences of data. In recent years, deep learning has gained widespread application in the WPF field due to its superior memory capabilities and ability to extract hidden data structures. Examples include traditional multi-layer perceptrons (MLPs), long-short-term memory networks (LSTMs), and convolutional neural networks (CNNs). Furthermore, the widespread use of attention mechanisms enables models to process sequences in parallel and dynamically focus on the entire picture, simultaneously capturing both long- and short-term dependencies across time steps. However, existing methods are not ideal for capturing complex temporal patterns and long-term trends, particularly when processing long sequences and utilizing temporal information. Furthermore, the local receptive field of traditional models further limits the acquisition of sequence information, making it difficult to effectively support multi-step predictions. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation, comprising the following steps:

[0006] S1: Obtain wind turbine power dataset and local weather dataset, filter relevant feature data, pre-process and divide the data into training set and test set;

[0007] S2: Construct a TAformer model for wind turbine power prediction;

[0008] S3: Using the power and environmental history data of the wind turbine, the TAformer model is trained to obtain a trained TAformer model;

[0009] S4: Obtain the data to be tested, and use the trained TAformer model to obtain the prediction results of the data to be tested.

[0010] The beneficial effects produced by the present invention are:

[0011] By introducing a cross-domain augmentation module, an adaptive decomposition structure, and a linear attention module, the Taformer model is able to learn from multidimensional time series and capture the coupling relationships between features and both long-term and short-term dependencies. Furthermore, the Taformer model utilizes an adaptive decomposition structure to decompose the linear attention output, creating a multi-layered model of long-term trend components and seasonal residual components, thereby maximizing the utilization of cyclical information. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying figures of the present invention include a flowchart of the prediction method, the overall architecture of the TAformer model, detailed schematic diagrams of the cross-domain augmentation module, the linear attention module, and the adaptive decomposition structure. These figures help to intuitively understand the structure and working principle of the present invention.

[0013] Figure 1 This is a schematic diagram of a wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to an embodiment of the present application.

[0014] Figure 2 This is a schematic diagram of the structure of the TAformer model in the embodiment of this application. This figure shows how the Taformer model uses the cross-domain augmentation module, linear attention module, and adaptive decomposition structure to model long-term dependencies and ultimately output prediction results.

[0015] Figure 3 This is a schematic diagram of the cross-domain augmentation module of an embodiment of the present application. This diagram shows in detail how the input matrix is ​​processed, and how the module enhances the temporal pattern of the sequence, amplifies the local high-frequency information, and highlights the periodic trend.

[0016] Figure 4 This is a schematic diagram of the structure of the linear attention module of an embodiment of the present application. This figure shows in detail the calculation method of the attention score.

[0017] Figure 5 This is a diagram of the adaptive decomposition architecture of an embodiment of the present application. This diagram shows in detail how to decompose the input tensor to obtain the long-term trend component and the seasonal residual component. DETAILED DESCRIPTION

[0018] The specific embodiments of the present invention are described below in conjunction with the accompanying drawings to illustrate the present invention more clearly.

[0019] like Figure 1-Figure 5 As shown, the embodiment of the present application provides a wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation. This method can capture long-series time dependencies and implement multi-variable correlation modeling through a cross-domain augmentation module and an adaptive decomposition structure, capturing multi-granularity local high-frequency information and highlighting periodic trends. The specific steps are as follows:

[0020] S1: Obtain a wind turbine power data set, collect a local weather data set, and filter some feature-related data through PCA to obtain sequence data. The sequence data is normalized or standardized and divided into a training set and a test set. The input sequence includes C features (such as date, wind speed, wind direction, temperature, humidity, air pressure, air density, and power value, etc.). The prediction is based on L historical time points, and the target is the power data at O ​​time points in the future (C, L, and O are all set thresholds). Weather characteristics are selected as input variables, and the wind turbine output power at the current moment is the target variable.

[0021] Therefore, the power input matrix can be expressed as:

[0022]

[0023] in, Represents historical data at time i (such as time, environmental factors and wind turbine power).

[0024] The predicted value at this time It can be expressed as: and At this time, the input matrix can be expressed as:

[0025]

[0026] Where X is the input sequence; is the value of the Cth feature at time point L; L is the number of time steps; C is the number of input features.

[0027] In this embodiment, the power sequence data is normalized using a maximum-minimization process:

[0028]

[0029] Where X is the input sequence and X′ is the normalized sequence. min With X max Represent the maximum and minimum values ​​respectively.

[0030] In particular, for the obtained numerical data, due to the existence of outliers and missing values, it is necessary to perform data preprocessing and fill in missing or outliers through offline filling methods.

[0031] S2: Construct a TAformer model for wind turbine power forecasting, integrating the long-term trend component representation and the seasonal residual component representation to generate forecast results.

[0032] like Figure 1 As shown in the figure, the TAformer model includes a cross-domain augmentation module, a linear attention module and an adaptive decomposition structure.

[0033] All structures include several linear attention modules and adaptive decomposition structures.

[0034] like Figure 3 As shown in the figure, the cross-domain augmentation module is used to improve the model's ability to identify long-term dependencies in sequences. It uses FFT to eliminate the impact of high-frequency noise on wind turbine power trends. It also extracts local features of the data through convolution and LSTM, capturing multi-granular representations of data features, thereby helping the model better understand power trends.

[0035] Furthermore, the cross-domain augmentation module consists of an encoder component matrix and a decoder component matrix. The encoder component matrix uses a long short-term memory network to enhance the encoder's ability to extract temporal features and enhance the temporal information of the input matrix. The decoder component matrix uses a fast Fourier function to extract the trend component of the sequence, smoothing the data trend and highlighting the periodic trend in the power sequence.

[0036] like Figure 3 As shown in Figure 2, the specific working method of the cross-domain augmentation module is as follows:

[0037] The sequence is constructed into an encoder component matrix and a decoder component matrix, and position encoding is performed on each matrix to obtain and The cross-domain augmentation module processes the two sequences separately:

[0038] Through LSTM is used to extract long-term sequence trends, and convolution operations are used to expand the local high-frequency patterns of the sequence to capture multi-granularity representations between data features.

[0039] Through Perform fast Fourier transform operations to filter out high-frequency noise information in the sequence, smooth the data trend, highlight the periodic trend in the power sequence, and help the model better understand the trend of wind turbine power.

[0040] like Figure 4 As shown in Figure 2, the linear attention module is used to capture long-term dependencies between multiple variables, dynamically focus on global information, and process sequences in parallel. Its working method is as follows:

[0041] For each attention head i, the input matrix X in Projection is query Q i , key K i , value V i :

[0042] Q i =X in W Qi ,K i =Xin W Ki ,V i =X in W Vi

[0043] in, are the projection matrices of the i-th attention head respectively.

[0044] The attention calculation weight can be expressed as:

[0045]

[0046] in, is the transpose of the key, d k Indicates the dimension of the key.

[0047] The weight α calculated above i Input multi-head attention and merge the output:

[0048] M-LA(X in )=W O Concat(α1,…,α i )

[0049] Among them, W O is the projection matrix of multi-head attention.

[0050] like Figure 5 As shown in the figure, the adaptive decomposition (AED) structure can gradually extract long-term stationary trends from the predicted intermediate hidden variables, decomposing the linear attention output and hierarchically modeling the long-term trend component and seasonal residual component to maximize the use of periodic information. It extracts multiple trend components by weighting the value of the current time step and the value of the previous time step, and finally combines them into the final trend, which is then outputted through a residual connection module.

[0051]

[0052]

[0053]

[0054]

[0055] in, represents the output of the M-layer decoder, and S represents a set of adaptive factors.

[0056] like Figure 2 As shown in Figure 1, the TAformer model performs operations including temporal position encoding, cross-domain augmentation module, encoder channel and decoder channel. The specific working method is as follows:

[0057] Temporal Position Encoding:

[0058] The position encoding of time series information is introduced to obtain feature representation that introduces position information, which enables the model to adjust the weight according to the position of the data point, thereby processing variable-length sequence data more flexibly.

[0059] Furthermore, position encoding introduces learnable position information for data points at different positions, enabling the attention mechanism to distinguish elements at different positions, helping the model better capture sequence relationships and improving the parallelism of the model.

[0060]

[0061]

[0062] X emb =PE pso (x)+Tem(x,d m )+Linear(x,d m )

[0063] Among them, d m is the embedding dimension, X emb represents the encoded sequence, and X emb ∈R L×C ; pos is the data point position, i is the index of the dimension, and d m Represents the dimension of the input sequence; PE(·) is a position encoding function that uses sine and cosine functions to establish the relative position of the sequence to help the model learn the trend of the time series. m ) maps the input feature x to the corresponding input dimension, Tem(x,d m ) represents temporal feature embedding.

[0064] Cross-domain augmentation module:

[0065] The decoder and encoder input tensors are enhanced with time series information. CDAM uses FFT to eliminate the impact of high-frequency noise on wind turbine power trends. Convolution and LSTM are used to extract local features of the data, capturing multi-granular representations of data features, helping the model better understand power trends. The encoder extracts time series features from the input sequence using one-dimensional convolution and LSTM.

[0066] X enc =LSTM(Conv1D(X))

[0067] The decoder performs frequency domain processing by fast Fourier transform, performs Fourier frequency processing on the time series information of the last dimension of the input sequence X, and obtains the complex spectrum spectrum = [f1,…,fdomain ], where f i Represents the frequency domain information of different frequencies, f domain Indicates high-frequency change information.

[0068]

[0069]

[0070] X filter (f) = X(f)·(1-δ(ff max ))

[0071]

[0072] Where LSTM is a long short-term memory network, Conv1D is a convolution operation, x(t) is a time domain signal, X(f) is a complex number representation of a signal with a frequency of f. Re(X) is the real part of the frequency domain, Im(X) is the imaginary part of the frequency domain, and the max(·) function selects the high-frequency information of the frequency domain sequence. The frequency domain is selected by the Dirac function to filter out the high-frequency components in the wind turbine power data. δ represents the Dirac function, and f max is the main frequency function, X filter is the frequency information after filtering, x filter (t) is the inverse Fourier function. Through Fourier transformation, the periodic time domain data Xx is extracted filter and the residual value x filter .

[0073] Encoder channels:

[0074] Taking the encoder component matrix as the input matrix, a linear attention module is used to extract the global dependencies of the sequence and dynamically assign weights. At the same time, a stacked adaptive decomposition structure is used to introduce trend components to obtain the encoder output representation, which is then fed into the decoder as cross-information.

[0075] In this embodiment, the encoder component X en Input the linear attention module and input the weight matrix into the stacked adaptive decomposition structure to obtain the long-term trend component Features are mapped to decoder channels via a feed-forward neural network with layer normalization.

[0076] like Figure 2 As shown in Figure 2, the encoder channel uses linear multi-head attention, stacked adaptive decomposition structure, and feedforward neural network, and maps the results to the decoder channel, specifically:

[0077] Assume that the encoder structure is N layers, The encoder component matrix As input matrix, and D is a hidden layer. A linear attention module is used to extract global dependencies in the sequence and dynamically assign weights. A stacked adaptive decomposition structure is used to identify long-term trend components. The encoder output representation is obtained through an FCN (·), a fully connected feedforward neural network module consisting of a Relu (·) activation function and fully connected layers to address potential exploding and vanishing gradient issues. The encoder output is fed into the decoder as cross-information.

[0078] Decoder channel:

[0079] Using the decoder component matrix as the input matrix, the cross-domain augmentation module extracts multi-granular information from the input matrix, obtaining an input representation that incorporates position information. Based on this position information, the stacked linear attention module and adaptive decomposition structure are used to obtain representations of the trend component and residual component.

[0080] In this embodiment, the decoder component Input linear attention module and then combine it with encoder result through adaptive decomposition structure Input linear attention module, for each attention head, The projection is into query, key, and value matrices. The introduction of the weight matrix connects the query, key, and value matrices at each position to the entire input sequence through a linear transformation. The attention weight is calculated for each query, input into the multi-head attention mechanism, and the combined output is fed into the adaptive decomposition structure.

[0081] The linear attention module and adaptive decomposition structure can be expressed as:

[0082] MHead(Q,K,V)=W O Concat(h1,…,h i )

[0083] h i =LA(Q,K,V)

[0084]

[0085] Q=X·W Q ,K=X·W K ,V=X·W V

[0086]

[0087]

[0088] Among them, W Ois the projection matrix of the multi-head attention output. The function Concat(·) connects the matrix coefficients calculated by n attention heads into the attention weight matrix. The input matrix Q, K, V of the i-th attention head can be expressed as Q i , K i , i∈{1,…,n}. Q is the query matrix, K is the key matrix, V is the value matrix, and d k represents the dimension of K, and LA(·) represents the linear attention function. is a set of adaptive factors, is the data value of layer l at time t, express of and, Represents the final trend component.

[0089] In the stacked multi-head linear attention module, residual connections and layer normalization are used to stabilize the training process, prevent gradient vanishing and explosion, and promote convergence:

[0090] X l =FCN(N(x l ))+x l

[0091] x l =M-LA(N(X l-1 ))+X l-1

[0092] Among them, X l is the output feature of the linear attention module in the first layer; FCN(·) is the feedforward neural network; N(·) is the layer normalization operation; x l is the output of the l-th layer linear attention after multi-head attention; M-LA(·) is the linear multi-head attention, X l-1 is the input of the l-1th layer linear attention.

[0093] like Figure 2 As shown in the figure, the decoder channel adopts stacked linear multi-head attention, stacked adaptive decomposition structure and feedforward neural network. The decoder consists of two parts: seasonal residual component Y and long-term trend component P. The model can effectively process the periodicity of the sequence and capture the trend of power changes. The first linear attention module performs a zero mask operation to prevent the model from using information from future time steps when generating the output of the current time step, thereby ensuring that the model only relies on historical wind turbine power information when generating the sequence. The second linear attention module inputs the cross information of the encoder and the decoder. Component. The decoder channel combines the seasonal residual components of each adaptive decomposition structure through residual connections and finally combines them with the long-term trend component. Each seasonal residual component can be expressed as: The decoder output is represented as

[0094] In summary, the TAformer model integrates the long-term trend component representation and the seasonal residual component representation to generate the forecast results:

[0095]

[0096] Among them, ω de The output of the M-layer decoder is mapped to the target dimension and merged with the seasonal residual component.

[0097] S3: Use the wind turbine power and environmental historical data to train the TAformer model to obtain the trained TAformer model. The loss function uses the mean square error (MSE) to improve the model training effect, as shown in the following formula:

[0098]

[0099] Among them, l MSE is the mean square error; represents the predicted value of the i-th sample at time t; ‖·‖2 is the L2 norm.

[0100] By training the TAformer model, the predicted value at the next T time points is obtained based on the historical information of the previous L time points.

[0101] S4: Obtain the data to be tested, use the trained TAformer model to obtain the predicted results of the data to be tested, and compare the predicted results with the true values ​​through the error index.

[0102] S5: Based on the obtained predicted value, the error MSE and MAE are calculated between the predicted value and the true value to perform model evaluation.

[0103] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in this field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation is characterized by: The following steps are involved: S1: Obtain wind turbine power dataset and local weather dataset, filter relevant feature data, pre-process and divide the data into training set and test set; S2: Construct a TAformer model for wind turbine power prediction; S3: Using the power and environmental history data of the wind turbine, the TAformer model is trained to obtain a trained TAformer model; S4: Obtain the data to be tested, and use the trained TAformer model to obtain the prediction results of the data to be tested.

2. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 1 is characterized in that: Said S1 specifically includes: Obtain wind turbine datasets, collect local weather datasets, filter relevant feature data, and obtain sequence data; The sequence data is normalized or standardized and divided into a training set and a test set.

3. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 1 or 2, characterized in that: In S1, the obtained numerical data are filled with missing or abnormal values ​​by using the median interpolation method.

4. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 1 is characterized in that: The TAformer model includes: a cross-domain augmentation module, a linear attention module and an adaptive decomposition structure; The cross-domain augmentation module is used to improve the model's recognition of long-term dependencies in sequences; The linear attention module is used to capture long-order dependencies between multiple variables, dynamically focus on global information, and process sequences in parallel; The adaptive decomposition structure is used to decompose the linear attention output and hierarchically model the long-term trend component and the seasonal residual component to maximize the utilization of periodic information.

5. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 4 is characterized in that: The cross-domain augmentation module includes: The encoder component matrix is ​​used to improve the encoder's temporal feature extraction capability through the long short-term memory network and enhance the temporal information of the input matrix; The decoder component matrix is ​​used to extract the trend component of the sequence through the fast Fourier function, smooth the data trend, and highlight the periodic trend in the power sequence.

6. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 4 is characterized in that: The working method of the TAformer model is: Temporal position coding is introduced to position-code the preprocessed data sequence, dividing it into encoder input and decoder input; the cross-domain augmentation module processes the two input sequences separately; The sequence processed by the cross-domain augmentation module is input into the encoder channel and the decoder channel, and the encoded output is used as cross information, merged with the decoder through linear attention, and predicted through subsequent operations.

7. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 5 is characterized in that: The working method of the cross-domain augmentation module is: The sequence is constructed into an encoder component matrix and a decoder component matrix, and position encoding is performed on each matrix to obtain and The cross-domain augmentation module processes the two sequences separately: right LSTM is used to extract long-term sequence trends, and convolution operations are used to expand the local high-frequency patterns of the sequence to capture multi-granularity representations between data features; right Perform fast Fourier transform operations to filter out high-frequency noise information in the sequence and highlight the periodic trend in the power sequence, thereby helping the model better understand the trend of wind turbine power.

8. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 4 is characterized in that: The working method of the linear attention module is: For each attention head i, the input matrix X in Projection is query Q i , key K i , value V i : V i =X in ·W Vi ; in, are the projection matrices of the i-th attention head respectively; Calculate attention weights: The calculated weight α i Input multi-head attention and merge the output: M-LA(X in )=W O ·Concat(α1,…,α i )。 9. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 6 is characterized in that: The working method of the encoder channel and the decoder channel is: The encoder channel uses the encoder component matrix as the input matrix, uses the linear attention module to extract the global dependencies of the sequence, and dynamically assigns weight relationships; The stacked adaptive decomposition structure is used to introduce the trend component to obtain the encoder output representation, which is then input into the decoder as cross information. The decoder channel takes the decoder component matrix as the input matrix and uses the cross-domain augmentation module to extract the multi-granularity information of the input matrix to obtain the input representation with introduced position information; According to the position information, the stacked linear attention modules and adaptive decomposition structure are used to obtain the representation of trend component and residual component.

10. The wind turbine power prediction method based on multi-level decomposition and cross-domain augmentation according to claim 1, characterized in that: Also includes: S5: Based on the obtained predicted value, the error MSE and MAE are calculated between the predicted value and the true value to perform model evaluation.

Images