Wind power prediction method and device based on WTC-Informer algorithm, equipment and medium

Abstract

The invention discloses a wind power prediction method and device based on a WTC-Informer algorithm, equipment and a medium, and relates to the technical field of wind power prediction. The wind power prediction method comprises the following steps: acquiring operation data of a wind driven generator; and inputting the operation data into a pre-trained wind power prediction model which is trained in advance and is based on a WTC-Informer algorithm, and executing the following data processing steps to obtain a prediction sequence of the active power of the wind driven generator. According to the operation data, high-frequency components and low-frequency components are extracted through wavelet transform and spliced, and multi-scale features are obtained. Through an ECA channel attention mechanism, weights are dynamically distributed for each feature channel. And then, through an Informer encoder formed by stacking a plurality of sparse self-attention coding layers, the calculation complexity of attention is reduced. And finally, outputting a future wind power prediction sequence through a generative decoder by utilizing a sparse attention extraction time sequence mode and combining a connection feature map output by the cross-attention fusion encoder.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a wind power prediction method, device, equipment and medium based on a WTC-Informer algorithm. Background Art

[0002] Wind power forecasting is the process of predicting wind power at a certain moment or within a period of time in the future by analyzing and modeling the historical power data of wind farms. Accurate wind power forecasting can effectively alleviate the huge pressure and instability brought by large-scale wind power grid connection to the power system of wind power companies, and can also effectively reduce the operating costs of the power system and optimize the operation and maintenance plans of the entire wind farm, thereby improving the utilization rate of wind energy and its competitiveness and market value in the power market.

[0003] In order to make accurate wind power forecasts, a large amount of historical data is needed as support. However, in actual operations, data often have problems such as missing, abnormal or inconsistent data. These problematic data will have a negative impact on the accuracy of the forecasting model. In addition, wind power is usually affected by the superposition of long-term and short-term periodic changes. For example, hourly changes in wind speed and temperature may have a direct impact on short-term wind power output, while seasonal changes will affect the overall power generation capacity of the wind farm. In order to improve the prediction accuracy, the model needs to be able to handle these multi-level and multi-scale periodic characteristics at the same time. Therefore, the wind power forecasting model must have the ability to capture changes on different time scales, so as to more comprehensively reflect the fluctuation law of wind power.

[0004] Deep learning has been widely used in the field of wind power forecasting in recent years due to its powerful nonlinear modeling and feature extraction capabilities. However, existing deep learning models still have problems such as high computational complexity and insufficient multi-scale feature extraction in wind power forecasting. For example, the Transformer-based time series forecasting model introduces a multi-head self-attention mechanism, which enables the model to simultaneously focus on different parts of the input sequence and learn global features. It performs well in capturing global dependencies, but due to excessive focus on long-distance dependencies, it may ignore short-term or local dependency information, thus affecting the accuracy and efficiency of the prediction.

[0005] In addition, the computational complexity of its dot product attention is O(n²). As the length of the input sequence increases, the amount of computation and memory consumption will increase sharply, resulting in poor performance in long-term predictions. Summary of the invention

[0006] The present invention provides a wind power prediction method, device, equipment and medium based on the WTC-Informer algorithm to improve at least one of the above technical problems.

[0007] In a first aspect, the present invention provides a wind power prediction method based on the WTC-Informer algorithm, which comprises: Obtain operating data of wind turbines.

[0008] The operation data is input into a pre-trained wind power prediction model based on the WTC-Informer algorithm, and the following data processing steps are performed to obtain a prediction sequence of the active power of the wind turbine.

[0009] According to the operation data, high-frequency components and low-frequency components are extracted by wavelet transform and spliced ​​to obtain multi-scale features.

[0010] Through the ECA channel attention mechanism, weights are dynamically assigned to each feature channel to obtain weighted multi-scale features.

[0011] The Informer encoder is composed of multiple sparse self-attention encoding layers, which reduces the computational complexity of attention from the standard Transformer to Down to , get the connection feature graph; where, Indicates the time complexity, is the length of the input sequence.

[0012] Through the generative decoder, sparse attention is used to extract the temporal pattern, combined with the connection feature map output by the cross-attention fusion encoder to output the future wind power prediction sequence.

[0013] Preferably, obtaining the operating data of the wind turbine specifically includes: Obtain the original operating data of wind turbines.

[0014] The original operating data is preprocessed to obtain preprocessed operating data; wherein the preprocessing includes outlier screening, missing value filling, data standardization, and feature selection.

[0015] Preferably, according to the operation data, high-frequency components and low-frequency components are extracted by wavelet transform and spliced ​​to obtain a multi-scale feature tensor, which specifically includes: Initialize the filter bank containing the multilayer wavelet decomposition Wavelet transform module. Among them, represents a low-frequency wavelet filter, represents a high-frequency wavelet filter.

[0016] The operation data is input into the wavelet transform module, high-frequency components are extracted by the high-frequency wavelet filter of the wavelet decomposition filter group, low-frequency components are extracted by the low-frequency wavelet filter of the wavelet decomposition filter group, and the coefficients of the current layer are calculated by a convolution operation. The low-frequency components extracted by the wavelet decomposition filter group of the previous layer are further decomposed into high-frequency components and low-frequency components as the input of the wavelet decomposition filter group of the next layer.

[0017] According to the high-frequency components of all levels obtained by the wavelet transform module, noise suppression is performed on each high-frequency component through a depth-wise separable convolution and point-by-point convolution to capture local changes, and a residual connection is made with the original high-frequency component to obtain the high-frequency features after reducing the high-frequency noise.

[0018] The low-frequency components obtained by the last layer of wavelet transform are interpolated to make their size consistent with the original input sequence, and then the first-order difference is used to capture the change pattern of the sequence. Finally, the low-frequency components are feature enhanced through residual connection to obtain the enhanced low-frequency features.

[0019] The high-frequency features and low-frequency features of all levels are concatenated together to obtain the multi-scale feature tensor.

[0020] Preferably, through the ECA channel attention mechanism, weights are dynamically assigned to each feature channel to obtain weighted multi-scale features, specifically including: Adaptively average pooling is performed on each channel of the multi-scale feature tensor to compress the spatial dimension; wherein, ; In the formula, For the Adaptive average pooling results for channels, is the input sequence length, is the sequence number of the moment, For multi-scale features, The sequence number of the channel.

[0021] The local channel relationship is captured through one-dimensional convolution to obtain the weight. In the formula, For weight, represents convolution, is the multi-scale feature after adaptive average pooling, Indicates transpose.

[0022] Use the Sigmoid function to normalize the weights, then broadcast the weights to the original spatiotemporal dimensions, perform channel-level scaling, and obtain weighted multi-scale features. In the formula, is the weighted multi-scale feature tensor, For multi-scale features, represents the channel-by-channel product, represents the Sigmoid function, For weight, Indicates transpose.

[0023] Sin-cosine position encoding is used to convert the weighted multi-scale feature timestamps into position information.

[0024] Preferably, the encoder model is: .

[0025] in, For the Layer Samples of time, represents the maximum pooling, is the activation function, represents one-dimensional convolution, Indicates Layer Samples of time, Represents a multi-head sparse attention block.

[0026] Preferably, the generative decoder uses sparse attention to extract the temporal pattern, combines the connection feature map output by the cross-attention fusion encoder, and outputs the future wind power prediction sequence, specifically including: The data sequence is input into the sparse self-attention module, and the correlation features within the data sequence are captured through the multi-head sparse self-attention mechanism. The data sequence includes the running data sequence and the sequence to be predicted.

[0027] The connection feature map is input into the cross-attention module, the associated features and the connection feature map are interacted by cross-attention, the key features captured by the encoder are integrated into the prediction generation process of the decoder through the attention mechanism, and the decoding vector of the hidden representation is obtained.

[0028] The decoded vector is mapped to a wind power prediction result through a fully connected layer to obtain a target sequence.

[0029] Preferably, the wind power prediction model based on the WTC-Informer algorithm is trained by the following steps: The historical operation data of the wind turbine generator is obtained, wherein the historical operation data includes characteristic data of a plurality of wind turbine generators.

[0030] Preprocessing the historical operation data includes detecting the historical operation data according to preset abnormal value determination conditions, deleting all abnormal data, and filling the missing data with the value of the previous moment of the vacant moment.

[0031] The preprocessed historical running data is divided into training set, validation set and test set, and standardized.

[0032] A correlation analysis is performed on the various features of the standardized historical operation data to screen out features with a correlation with wind power higher than a preset value, and to obtain the final training set, validation set, and test set.

[0033] A wind power prediction model based on the WTC-Informer algorithm is constructed, and training is performed based on the final training set and the pre-set loss function to establish the mapping relationship between input features and wind power. The performance of the model is evaluated after each training cycle using the validation set, and the early stopping strategy is used. When the performance on the validation set does not improve in several consecutive training cycles, the training is terminated in advance to ensure the generalization ability of the model.

[0034] The trained model is evaluated by inputting the final test set. When the model meets the preset standards, the model parameters are saved to obtain the pre-trained wind power prediction model based on the WTC-Informer algorithm.

[0035] In a second aspect, the present invention provides a wind power prediction device based on the WTC-Informer algorithm, which is used to implement a wind power prediction method based on the WTC-Informer algorithm as described in any paragraph of the first aspect.

[0036] The wind power prediction device comprises: The operation data acquisition module is used to acquire the operation data of the wind turbine.

[0037] The prediction module is used to input the operating data into a pre-trained wind power prediction model based on the WTC-Informer algorithm, perform data processing through the following units, and obtain a prediction sequence of the active power of the wind turbine.

[0038] The wavelet transform subunit is used to extract high-frequency components and low-frequency components through wavelet transform according to the operation data and splice them to obtain multi-scale features.

[0039] The ECA channel attention subunit is used to dynamically assign weights to each feature channel through the ECA channel attention mechanism to obtain weighted multi-scale features.

[0040] The encoder unit is used to stack multiple sparse self-attention encoding layers into an Informer encoder, which reduces the computational complexity of attention from the standard Transformer Down to , get the connection feature graph; where, Indicates the time complexity, is the length of the input sequence.

[0041] The decoder unit is used to extract the temporal pattern using sparse attention through a generative decoder, and output the future wind power prediction sequence by combining the connection feature map output by the cross-attention fusion encoder.

[0042] In a third aspect, the present invention provides a wind power prediction device based on the WTC-Informer algorithm, comprising a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement a wind power prediction method based on the WTC-Informer algorithm as described in any paragraph of the first aspect.

[0043] In a fourth aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a wind power prediction method based on the WTC-Informer algorithm as described in any paragraph of the first aspect.

[0044] By adopting the above technical solution, the present invention can achieve the following technical effects: The wind power prediction method based on the WTC-Informer algorithm of the present invention performs multi-level wavelet decomposition on the input sequence through a multi-scale feature extraction module, and uses a wavelet filter group to decompose the time series data layer by layer into multi-level high-frequency components and deep low-frequency components. According to different frequency characteristics, the module uses hierarchical processing to perform deep separable convolution and then linear interpolation reconstruction on the high-frequency components, and differentially enhances the low-frequency components and performs residual connection with the original data. Finally, the multi-scale features are dynamically fused through the channel attention mechanism to achieve cross-scale feature interaction from short-term fluctuations to long-term trends, significantly improving the modeling ability of wind power sequences.

[0045] Using Informer as a predictor, through the sparse attention mechanism, only a small number of point pairs that contribute the most to the output are calculated, thereby reducing the complexity of attention calculation. The encoder reduces the sequence length by downsampling layer by layer, and retains the most important feature points at each layer, thereby further improving the calculation efficiency, and uses a multi-head sparse self-attention mechanism to capture the internal correlation features of the input sequence. In addition, Informer's cross-attention is used to interact the decoder input with the global context features of the encoder output, and integrate the key features captured by the encoder into the decoder's prediction generation process. The output of the decoder maps the hidden representation to the final prediction result through a fully connected layer to generate the target sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the specific implementation methods of the present invention. It should be understood that the following drawings only show certain specific implementation methods of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 It is a flow chart of the wind power prediction method.

[0048] Figure 2 It is a flowchart of the training process of the wind power prediction model.

[0049] Figure 3 This is the architecture diagram of the wind power prediction model based on the WTC-Informer algorithm. Figure 4 It is a schematic diagram of wavelet transform. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Example 1, please refer to Figures 1 to 4 The first embodiment of the present invention provides a wind power prediction method based on the WTC-Informer algorithm, which can be performed by a wind power prediction device based on the WTC-Informer algorithm (hereinafter referred to as: wind power prediction device). In particular, it is performed by one or more processors in the wind power prediction device to implement steps A1 to A2. It can be understood that the wind power prediction device can be an electronic device with computing performance such as a portable notebook computer, a desktop computer, a server, a smart phone or a tablet computer.

[0052] A1. Obtaining the operating data of the wind turbine. Preferably, step A1 specifically includes steps A11 to A12.

[0053] A11. Obtain the original operating data of the wind turbine.

[0054] A12. Preprocess the original operating data to obtain preprocessed operating data; wherein the preprocessing includes outlier screening, missing value filling, data standardization, and feature selection.

[0055] A2. Input the operating data into a pre-trained wind power prediction model based on the WTC-Informer algorithm, execute steps B1 to B4, and obtain a prediction sequence of the active power of the wind turbine for a period of time in the future.

[0056] Specifically, the wind power prediction model based on the WTC-Informer algorithm constructs a nonlinear mapping relationship from historical input features to future active power by integrating wavelet transform and Informer architecture. The model takes preprocessed time series data as input, and directly outputs the active power prediction sequence of wind turbines for a period of time in the future through multi-scale feature extraction and long sequence dependency modeling. This model can provide wind power operators with accurate decision-making basis, support them in optimizing wind farm power dispatch, grid stability control and spare capacity configuration, and promote the widespread application and sustainable development of wind power generation.

[0057] B1. According to the operation data, high-frequency components and low-frequency components are extracted by wavelet transform and then concatenated to obtain multi-scale features. Preferably, B1 specifically includes steps B11 to B15.

[0058] The process of wavelet transform in this embodiment can be understood as using a high-frequency wavelet filter and a low-frequency wavelet filter to perform convolution operations on the input sequence respectively to obtain a high-frequency component and a low-frequency component. Figure 4 As shown, the low-frequency components represent global characteristics and trends, and the high-frequency components represent rapidly changing detail characteristics.

[0059] B11. Initialize the filter bank containing multi-layer wavelet decomposition Wavelet transform module. Among them, represents a low-frequency wavelet filter, Represents a high-frequency wavelet filter. For the input sequence, the filter is used to perform low-frequency wavelet decomposition on each variable step by step.

[0060] B12, input the operation data into the wavelet transform module, extract the high-frequency component through the high-frequency wavelet filter of the wavelet decomposition filter bank, extract the low-frequency component through the low-frequency wavelet filter of the wavelet decomposition filter bank, and calculate the coefficient of the current layer through a convolution operation. Figure 3 As shown, the low-frequency components extracted by the previous wavelet decomposition filter bank are further decomposed into high-frequency components and low-frequency components as the input of the next wavelet decomposition filter bank.

[0061] Specifically, the input signal (i.e., running data) is decomposed layer by layer into low-frequency and high-frequency parts by a wavelet decomposition filter bank with multiple layers. For each layer of wavelet decomposition, a convolution operation is used to calculate the coefficients of the current layer, and the low-frequency part of the signal is used as the input of the next layer. The high-frequency part of each layer will be saved and used as the subsequent high-frequency features.

[0062] The calculation model of the wavelet transform module is: .

[0063] in, Indicates the level of wavelet transform, is the variable number, Indicates low frequency, Indicates high frequency, Indicates The input sequence of variables, Indicates The variable The low-frequency components extracted by wavelet transform, Indicates The variable The high frequency components extracted by wavelet transform, represents convolution, represents a low-frequency wavelet filter, represents a high-frequency wavelet filter, Indicates The variable The low-frequency components extracted by the wavelet transform. hour, Use the run data preprocessed in step A1.

[0064] B13. Based on the high-frequency components of all levels obtained by the wavelet transform module, noise suppression is performed on each high-frequency component through a depth-wise separable convolution and point-by-point convolution to capture local changes, and a residual connection is made with the original high-frequency component to obtain the high-frequency component after reducing the high-frequency noise, so as to reduce the impact of high-frequency noise on the prediction.

[0065] The processing model of high-frequency components is: .

[0066] in, Indicates The variable The high frequency components extracted by wavelet transform, represents the point-by-point convolution operation, Indicates splicing, Represents a depthwise separable convolution operation.

[0067] B14. Interpolate the low-frequency components obtained by the last layer of wavelet transform to make their size consistent with the original input sequence, then use the first-order difference to capture the change pattern of the sequence, and finally enhance the features of the low-frequency components through residual connection to obtain the enhanced low-frequency components.

[0068] Specifically, the low-frequency component obtained by the last layer of wavelet transform Represents the trend or long-term change in the input sequence. In this embodiment, the low-frequency component is first Interpolation is performed to ensure that the size is consistent with the original input sequence. The first-order difference is then used to capture the changing pattern of the sequence. Finally, the low-frequency components are enhanced through residual connections: The processing model of the low-frequency component obtained by the last layer of wavelet transform is: .

[0069] .

[0070] in, for The first-order difference of time, Indicates splicing, Indicates variables The low-frequency components extracted at all times, Indicates variables The low-frequency components extracted at all times, represents the low-frequency component obtained by the last layer of wavelet transform, Represents a point-wise convolution operation.

[0071] B15. All high-frequency features and low-frequency features at all levels are spliced ​​together to obtain the multi-scale features. The calculation model of the multi-scale feature tensor is: .in, For multi-scale features, Indicates splicing, Indicates The variable The high frequency components extracted by wavelet transform, Represents the low-frequency component obtained by the last layer of wavelet transform.

[0072] B2. Dynamically assign weights to each feature channel through the ECA channel attention mechanism to obtain weighted multi-scale features to further enhance the model's learning ability for features of different scales. Preferably, step B2 specifically includes steps B21 to B24.

[0073] B21. Perform adaptive average pooling on each channel of the multi-scale feature tensor to compress the spatial dimension.

[0074] .

[0075] in, For the Adaptive average pooling results for channels, is the input sequence length, is the sequence number of the moment, For multi-scale features, The sequence number of the channel. . is the total number of moments.

[0076] B22. Capture local channel relationships through one-dimensional convolution and obtain weights.

[0077] .

[0078] In the formula, For weight, represents convolution, is the multi-scale feature after adaptive average pooling, Indicates transpose.

[0079] B23. Use the Sigmoid function to normalize the weights, then broadcast the weights to the original spatiotemporal dimensions, perform channel-level scaling, and obtain weighted multi-scale features.

[0080] .

[0081] In the formula, is the weighted multi-scale feature tensor, For multi-scale features, represents the channel-by-channel product, represents the Sigmoid function, For weight, Indicates transpose.

[0082] B24. Use sine and cosine position encoding to convert the weighted multi-scale feature timestamps into position information.

[0083] Specifically, since the attention mechanism itself is insensitive to the order of sequence data, in order to enhance the model's ability to model the periodicity of the wind turbine operating status sequence, the present invention adopts sine and cosine position encoding to convert timestamps into position information, so that the model can distinguish different time steps and effectively retain the time sequence information, thereby improving the prediction accuracy.

[0084] The calculation model using sine and cosine position encoding is: .

[0085] .

[0086] in, For position coding, represents the position index of the current time step, Indicates the encoding dimension, is the dimension of the model, is a sine function, is the cosine function.

[0087] B3, through multiple sparse self-attention encoding layers stacked into an Informer encoder, the computational complexity of attention is reduced from the standard Transformer Down to , get the connection feature graph; where, Indicates the time complexity, is the length of the input sequence.

[0088] In this embodiment, the encoder is composed of multiple stacked encoding modules to extract the global dependencies and local features of the historical sequence and compress them into efficient context representations. The encoder reduces the sequence length by downsampling layer by layer. Each layer retains the most important feature points, thereby further improving efficiency.

[0089] The encoder model is: .

[0090] in, For the Layer Samples of time, represents the maximum pooling, is the activation function, represents one-dimensional convolution, Indicates Layer Samples of time, Represents a multi-head sparse attention block.

[0091] The encoding module uses a sparse attention mechanism to calculate only a small number of point pairs that contribute most to the output, reducing the complexity of attention calculation from the standard Transformer Down to In the sparse attention mechanism, the input is converted into three parts: query ,key ,value In all Random sampling is performed in indivual , whose distribution can represent all The distribution of Calculate each The attention score, The process can be expressed as: .

[0092] .

[0093] .

[0094] .

[0095] in, Represents the input of the encoding module, To generate the learnable matrix for the query, is the learnable matrix for generating keys, is a learnable matrix that generates values, The calculated The attention score of the query, Indicates queries, For query The index of Indicates keys, Key The index of Represents transposition, represents the dimension of the input vector, It is randomly selected The number of

[0096] The sparse self-attention calculation formula is: .

[0097] in, represents the output of the sparse self-attention mechanism, For query, For key, For value, represents the softmax function, represents the top selected according to the attention score "active" queries, Indicates transposition, Represents the dimension of the input vector.

[0098] B4, using a generative decoder, using sparse attention to extract the temporal pattern, combining the connection feature map output by the cross-attention fusion encoder, and outputting a future wind power prediction sequence. Preferably, step B4 specifically includes steps B41 to B43.

[0099] B41. Input a data sequence into the sparse self-attention module, and capture the internal correlation features of the data sequence through the multi-head sparse self-attention mechanism. The data sequence includes a running data sequence and a sequence to be predicted.

[0100] B42. Input the connection feature map into the cross-attention module, use cross-attention to interact the associated features with the connection feature map, integrate the key features captured by the encoder into the prediction generation process of the decoder through the attention mechanism, and obtain the decoding vector of the hidden representation.

[0101] B43. Map the decoded vector to a wind power prediction result through a fully connected layer to obtain a target sequence.

[0102] In this embodiment, the decoder is composed of a sparse self-attention module and a cross-attention module. The decoder generates a prediction sequence of the wind turbine active power for a period of time in the future based on the context vector output by the encoder. The historical input sequence and length are The sequence to be predicted is filled with 0 to generate the prediction result.

[0103] Specifically, the decoder uses a multi-head sparse self-attention mechanism to capture the internal correlation features of the decoder input sequence. Cross-attention is used to interact the decoder input with the global context features of the encoder output, and the key features captured by the encoder are integrated into the decoder's prediction generation process through the attention mechanism. Finally, the decoder output maps the hidden representation to the final prediction result through a fully connected layer to generate the target sequence.

[0104] The wind power prediction method based on the WTC-Informer algorithm of the present invention integrates wavelet transform and deep neural network architecture to realize cross-scale feature learning of multivariate time series data. The model performs multi-level wavelet transform on multi-channel input through a wavelet decomposition module, wherein the high-frequency component uses deep separable convolution and residual gating to achieve noise robustness processing, and the low-frequency component is fused with the original signal feature through the first-order difference to enhance the trend characterization capability. Through dynamic weighted fusion of the channel attention network, a multi-scale feature tensor is formed. This architecture breaks through the limitations of traditional single-scale convolution and improves the modeling accuracy of non-stationary features and cross-scale dependencies in wind power sequences.

[0105] The sparse self-attention mechanism based on the Informer algorithm realizes efficient attention calculation and fully captures the correlation features within the input sequence. Based on the context vector output by the encoder, the decoder directly generates the prediction results of the future sequence, thereby achieving accurate and efficient wind power prediction. Power dispatchers can use the prediction results as a basis for power demand dispatch planning, effectively alleviating the huge pressure and instability brought by large-scale wind power grid connection to the power system, while also effectively reducing the operating cost of the power system and optimizing the operation and maintenance plan of the entire wind farm.

[0106] The pre-training step of the pre-trained wind power prediction model based on the WTC-Informer algorithm includes steps S1 to S6.

[0107] S1. Obtain historical operation data of wind turbines.

[0108] The historical operation data of wind turbines are collected through sensors distributed on wind turbines. The historical operation data includes: including timestamps , wind speed , the angle between the wind direction and the position of the turbine nacelle , Ambient temperature , the temperature in the turbine nacelle , Nacelle direction (yaw angle) 、Blade 1 pitch angle , blade 2 pitch angle 、Blade 3 pitch angle , reactive power , Active Power Among them, the active power It serves as both input feature and predicted target value.

[0109] Specifically, the fan The running state at each moment is represented by the feature vector: . Define the target task as using the past time step data , to predict the future Active power in time steps The learning goal is to find a mapping relationship , so that the predicted value As close to the true value as possible .in The wind power prediction model based on the WTC-Informer algorithm proposed in the present invention (such as Figure 3 ), is a trainable parameter, .

[0110] S2. Preprocess the historical operation data. The preprocessing includes detecting the historical operation data according to a preset abnormal value judgment condition, deleting all abnormal data, and filling the missing data with the value of the previous moment of the vacant moment to ensure the integrity of the data.

[0111] Specifically, the abnormal value determination condition is: When the active power is less than 0, it is considered abnormal.

[0112] When the wind speed is less than 1 m / s and the active power is greater than 10 kilowatts, it is considered abnormal.

[0113] When the wind speed is less than 2 m / s and the active power is greater than 100 kilowatts, it is considered abnormal.

[0114] When the wind speed is less than 3 m / s and the active power is greater than 200 kilowatts, it is considered abnormal.

[0115] When the wind speed is greater than 2.5 m / s and the active power is equal to 0, it is considered abnormal.

[0116] When the wind speed is equal to 0, the angle between the wind direction and the turbine nacelle position is equal to 0, and the ambient temperature is equal to 0, it is considered abnormal.

[0117] When the ambient temperature is less than -21 degrees Celsius or greater than 60 degrees Celsius, it is considered abnormal.

[0118] When the temperature in the turbine compartment is less than -21 degrees Celsius or greater than 70 degrees Celsius, it is considered abnormal.

[0119] When the angle between the wind direction and the turbine nacelle position is greater than 180 degrees or less than -180 degrees, it is considered abnormal.

[0120] When the cabin direction is greater than 720 degrees or less than -720 degrees, it is considered abnormal.

[0121] When any one of blade pitch angle 1, blade pitch angle 2 or blade pitch angle 3 is greater than 89 degrees, it is considered abnormal.

[0122] After removing the outliers, check the missing values ​​in the data set. If the missing value appears in the first data, delete the record at that moment. Otherwise, fill the missing value with the value of the previous moment.

[0123] S3. Divide the preprocessed historical operation data into a training set, a validation set, and a test set, and perform standardization.

[0124] In this embodiment, the historical operation data is divided into a training set, a validation set, and a test set according to a certain ratio. The training set is used for model training, the validation set is used to adjust hyperparameters and monitor model performance during the training process, and the test set is used to finally evaluate the generalization ability of the model. Then, the historical data is standardized based on the statistics of the training set to eliminate the impact of the dimensions between different features.

[0125] Specifically, the operation data of the fan is divided into a training set, a validation set, and a test set according to a certain ratio in chronological order. The mean and standard deviation of all features and active power except timestamp in the training set, validation set and test set The features outside the dataset are Z-score standardized and mapped to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0126] Furthermore, the fan is Features , its standardized formula can be expressed as:

[0127] in, For the standardized data, is the original value, For the Features The mean value on the training set, Features The standard deviation on the training set.

[0128] and The calculation method is:

[0129]

[0130] in, Represents the total number of time steps in the training set, that is, the number of original data items contained in the training set.

[0131] Fan at the moment Active power The standardized formula can be expressed as:

[0132] in, is the normalized active power, is the original value of active power, is the mean of active power in the training set, is the standard deviation of active power in the training set.

[0133] After standardization, the fan Operation data at the moment It is expressed as: .

[0134] S4. Perform correlation analysis on various features of the standardized historical operation data, and screen out features whose correlation with wind power is higher than a preset value.

[0135] Specifically, we conduct correlation analysis on each feature in the historical data to evaluate its impact on wind power. We screen out features that are highly correlated with wind power and construct a new data set as input for model training to reduce redundant features, reduce the risk of overfitting, and improve the training efficiency and calculation speed of the model.

[0136] In this embodiment, the Pearson correlation analysis is performed using the training set data after standardization to obtain the Pearson correlation matrix. Since the data has been Z-score standardized, Pearson correlation coefficient The calculation formula can be expressed as:

[0137] in, represents the total number of time steps in the training set, For the standardized data, is the normalized active power.

[0138] The Pearson correlation coefficient can be used to measure the correlation between two variables, and its value range is [−1,1]. indicates a completely positive correlation. indicates a completely negative correlation. Indicates that there is no linear correlation between the two. By calculating the Pearson correlation matrix, the N features with the highest correlation with active power are screened out, and a new data set is constructed to be used as the input of the model to improve the performance and computational efficiency of the model.

[0139] The data after feature selection is represented as: Training set ,in Indicates that the nth feature is between t=1 and t= sequence on.

[0140] Validation set ,in Indicates that the nth feature at t= +1 to t= sequence on; Validation set ,in Indicates that the nth feature at t= +1 to t= sequence on.

[0141] For the new training set, validation set, and test set, the sliding window technique is used to generate the input matrix of the final model, where the sliding window size is .

[0142] S5. Build a wind power prediction model based on the WTC-Informer algorithm and train it according to the final training set. The architecture of the wind power prediction model is as follows: Figure 3 As shown in FIG. 1 , the model as a whole is composed of a multi-scale feature extraction unit, an encoder unit, and a decoder unit. The multi-scale feature extraction unit includes a wavelet transform subunit and an ECA channel attention subunit.

[0143] The multi-scale feature extraction unit uses wavelet transform, deep separable convolution and channel attention to extract multi-scale features from the input sequence. In addition, the position encoding module of the embedding layer introduces position information for the time series data and maps the features to high dimensions to make up for the defect that the self-attention mechanism cannot directly perceive the sequence order. The encoder unit is composed of multiple encoding modules stacked together to extract the global dependencies and local features of the historical sequence and compress them into an efficient context representation. The decoder unit generates a prediction sequence for the active power of the wind turbine in the future based on the context vector output by the encoder.

[0144] The training set is used to train the wind power prediction model based on the WTC-Informer algorithm to establish the mapping relationship between input features and wind power. Through the pre-set loss function, back propagation is performed to gradually adjust the model parameters, so that the model can more accurately predict wind power.

[0145] Use the validation set to evaluate the performance of the model after each training cycle. Use the early stopping strategy to terminate the training early when the performance on the validation set has not improved in several consecutive training cycles to ensure the generalization ability of the model.

[0146] In this embodiment, the model parameters are continuously optimized using the data in the training set through a pre-set loss function. The present invention uses the mean square error (MSE) as the loss function. During the model training process, an early stopping strategy based on the validation set monitoring is adopted to balance the model performance and training efficiency. After each training cycle, the pre-set evaluation index MSE is calculated based on the validation set to evaluate the generalization ability of the model, and the current optimal model parameters are dynamically saved.

[0147] If the performance of the validation set does not improve for several consecutive training cycles or reaches the maximum iteration limit, the training is terminated. After the training is completed, the model parameters with the best performance on the validation set are rolled back to ensure that the final model has the best generalization ability and can be used for test set evaluation or actual wind farm power prediction, realizing the mapping from historical features to future power.

[0148] S6. The trained model is evaluated by inputting the final test set. When the model meets the preset standards, the model parameters are saved to obtain the pre-trained wind power prediction model based on the WTC-Informer algorithm. Specifically, the test set data is input into the trained prediction model to obtain the prediction results. By comparing with the real data, the model evaluation index is calculated to further evaluate the prediction effect of the model. By analyzing the deviation between the prediction results and the actual situation, the model design is adjusted and trained again to improve its accuracy and robustness.

[0149] The test set is input into the trained wind power prediction model, and forward reasoning is performed to obtain the future The wind turbine active power prediction sequence of time steps is calculated by comparing with the real data and calculating the mean square error (MSE), mean absolute error (MAE) and other evaluation indicators.

[0150] The final generalization ability of the model is evaluated based on the evaluation indicators. If the main indicators meet the preset standards, the model is judged to be qualified, and the model parameters are saved synchronously to form a deployable pre-trained prediction model. If it does not meet expectations, the model is retrained.

[0151] Analyze the evaluation results, check the performance of the model in different scenarios, and understand its strengths and limitations. Based on the evaluation results, the model may need to be adjusted and optimized to improve its performance and obtain the final wind power prediction model based on the WTC-Informer algorithm.

[0152] Save the best model and load it directly for future use, avoiding the time and computing resource overhead of retraining.

[0153] In order to verify the effectiveness of the model and solution method proposed in this invention, the operating data of a wind turbine in a wind farm was selected as the research object. This data set covers 245 consecutive days of operation records, providing sufficient data to ensure that the experimental results are statistically significant and reliable.

[0154] In the experiment, the proposed model was used to predict the active power of wind turbines at different time lengths in the future, including the output active power prediction for the next 2 hours, 4 hours, 8 hours and 12 hours. For each prediction period, two main evaluation indicators, mean square error (MSE) and mean absolute error (MAE), were calculated to quantify the accuracy of the prediction.

[0155] In order to comprehensively evaluate the performance of the model proposed in the present invention, this application selects four types of representative time series prediction architectures as benchmark models: the basic Informer model, the classic Transformer model, and the LightTS model.

[0156] The Transformer model uses the original self-attention mechanism as a core component, and its multi-head attention layer models the global dependencies of sequences through parallel computing. The Informer model optimizes computational complexity through the ProbSparse self-attention mechanism and designs distillation operations to achieve hierarchical feature extraction. The LightTS model builds a lightweight temporal network based on deep separable convolutions, and its temporal convolution module captures multi-scale temporal patterns through dilated causal convolutions.

[0157] Experimental results show that the WTC-Informer model has certain advantages in the multi-period forecasting task of wind power: within the 2-12 hour forecast range, its MSE and MAE indicators are significantly better than the Transformer, Informer and other comparison models, especially in the 12-hour long-term forecast, the MSE is 4.8% lower than the suboptimal model, and the error growth rate is the slowest. Through the cross-scale fusion mechanism of wavelet processor and deep network, the model accurately captures the transient fluctuations and slow-changing trends of wind power in the joint time-frequency domain, effectively solves the error accumulation problem of traditional methods in long-term forecasting, verifies the high adaptability of the multi-scale feature learning architecture to complex time series patterns, and provides more reliable forecasting support for wind power grid-connected scheduling.

[0158] Table 1 Comparative experimental results

[0159] In summary, the embodiment of the present invention introduces the idea of ​​wavelet transform into the processing of wind power and its related variable data, and constructs a multi-scale feature extraction module. Through the layer-by-layer low-frequency decomposition of wavelet transform, differential feature enhancement and convolution are performed on the low-frequency and high-frequency results of different levels, respectively, to efficiently capture the local and global features in the time series, enhance the modeling ability of multi-scale information, and avoid the problem of a sharp increase in the number of parameters caused by directly using large convolution kernels to extract features from the original sequence. Through the sparse attention mechanism designed specifically for long time series modeling, only a small number of point pairs that contribute the most to the output are calculated, achieving efficient attention calculation.

[0160] By combining wavelet convolution and Informer algorithm, the present invention provides a more powerful wind power prediction model. The model has a powerful multi-scale feature extraction capability and attention mechanism, which can make full use of the long-term trend and short-term fluctuation of variables, establish an accurate nonlinear mapping relationship between historical features and future power generation, and thus provide a new and effective method for wind power prediction.

[0161] Embodiment 2: A wind power prediction device based on the WTC-Informer algorithm, which is used to implement a wind power prediction method based on the WTC-Informer algorithm as described in any paragraph of Embodiment 1. The wind power prediction device includes: an operation data acquisition module, a prediction module, a wavelet transform subunit, an ECA channel attention subunit, an encoder unit and a decoder unit.

[0162] The operation data acquisition module is used to acquire the operation data of the wind turbine.

[0163] The prediction module is used to input the operating data into a pre-trained wind power prediction model based on the WTC-Informer algorithm, perform data processing through the following units, and obtain a prediction sequence of the active power of the wind turbine.

[0164] The wavelet transform subunit is used to extract high-frequency components and low-frequency components through wavelet transform according to the operation data and splice them to obtain multi-scale features.

[0165] The ECA channel attention subunit is used to dynamically assign weights to each feature channel through the ECA channel attention mechanism to obtain weighted multi-scale features.

[0166] The encoder unit is used to stack multiple sparse self-attention encoding layers into an Informer encoder, which reduces the computational complexity of attention from the standard Transformer Down to , get the connection feature graph; where, Indicates the time complexity, is the length of the input sequence.

[0167] The decoder unit is used to extract the temporal pattern using sparse attention through a generative decoder, and output the future wind power prediction sequence by combining the connection feature map output by the cross-attention fusion encoder.

[0168] Embodiment 3: A wind power prediction device based on the WTC-Informer algorithm, comprising a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement a wind power prediction method based on the WTC-Informer algorithm as described in any paragraph of Embodiment 1.

[0169] Embodiment 4: A computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute a wind power prediction method based on the WTC-Informer algorithm as described in any paragraph of Embodiment 1.

[0170] In several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed apparatus and method can also be implemented in other ways. The apparatus and method embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the apparatus, method and computer program product according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, program segment or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0171] In addition, the functional modules in the various embodiments of the present invention may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0172] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, electronic device, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk. It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements inherent to such a process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0173] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0174] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0175] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0176] The "first\second" mentioned in the embodiments is only to distinguish similar objects, and does not represent a specific order for the objects. It is understandable that the "first\second" can be interchanged with the specific order or sequence where permitted. It should be understood that the objects distinguished by "first\second" can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0177] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wind power prediction method based on WTC-Informer algorithm, characterized in that: Include: Obtaining wind turbine operating data; Input the operating data into a pre-trained wind power prediction model based on the WTC-Informer algorithm, and perform the following data processing steps to obtain a prediction sequence of the active power of the wind turbine; According to the operation data, high-frequency components and low-frequency components are extracted by wavelet transform and spliced ​​to obtain multi-scale features; Through the ECA channel attention mechanism, weights are dynamically assigned to each feature channel to obtain weighted multi-scale features; The Informer encoder is stacked with multiple sparse self-attention encoding layers, which reduces the computational complexity of attention from the standard Transformer to Down to , get the connection feature graph; where, Indicates the time complexity, is the length of the input sequence; Through the generative decoder, sparse attention is used to extract the temporal pattern, combined with the connection feature map output by the cross-attention fusion encoder to output the future wind power prediction sequence.

2. The wind power prediction method based on the WTC-Informer algorithm according to claim 1 is characterized in that: Obtain the operating data of wind turbines, including: Obtain the original operating data of wind turbines; The original operating data is preprocessed to obtain preprocessed operating data; wherein the preprocessing includes outlier screening, missing value filling, data standardization, and feature selection.

3. The wind power prediction method based on the WTC-Informer algorithm according to claim 1 is characterized in that: According to the operation data, high-frequency components and low-frequency components are extracted and spliced ​​by wavelet transform to obtain multi-scale feature tensors, which specifically includes: Initialize the filter bank containing the multilayer wavelet decomposition The wavelet transform module of represents a low-frequency wavelet filter, represents a high-frequency wavelet filter; Input the operating data into the wavelet transform module, extract high-frequency components through the high-frequency wavelet filter of the wavelet decomposition filter group, extract low-frequency components through the low-frequency wavelet filter of the wavelet decomposition filter group, and calculate the coefficients of the current layer through a convolution operation; wherein the low-frequency components extracted by the wavelet decomposition filter group of the previous layer are further decomposed into high-frequency components and low-frequency components as the input of the wavelet decomposition filter group of the next layer; According to the high-frequency components of all levels obtained by the wavelet transform module, noise suppression is performed on each high-frequency component through a depth-wise separable convolution and point-by-point convolution to capture local changes, and residual connections are made with the original high-frequency components to obtain high-frequency features after reducing high-frequency noise; The low-frequency components obtained by the last layer of wavelet transform are interpolated to make their size consistent with the original input sequence, and then the first-order difference is used to capture the change pattern of the sequence. Finally, the low-frequency components are enhanced through residual connection to obtain the enhanced low-frequency features; The high-frequency features and low-frequency features of all levels are concatenated together to obtain the multi-scale feature tensor.

4. The method for wind power prediction based on WTC-Informer algorithm according to claim 1, characterized in that: Through the ECA channel attention mechanism, weights are dynamically assigned to each feature channel to obtain weighted multi-scale features, including: Adaptively average pooling is performed on each channel of the multi-scale feature tensor to compress the spatial dimension; wherein, ; In the formula, For the Adaptive average pooling results for channels, is the input sequence length, is the sequence number of the moment, For multi-scale features, is the sequence number of the channel; The local channel relationship is captured through one-dimensional convolution to obtain the weight; where ; In the formula, For weight, represents convolution, is the multi-scale feature after adaptive average pooling, represents transpose; The weights are normalized using the Sigmoid function, and then broadcast to the original spatiotemporal dimensions, channel-level scaling is performed, and weighted multi-scale features are obtained; where: ; In the formula, is the weighted multi-scale feature tensor, For multi-scale features, represents the channel-by-channel product, represents the Sigmoid function, For weight, represents transpose; Sin-cosine position encoding is used to convert the weighted multi-scale feature timestamps into position information.

5. The method for wind power prediction based on WTC-Informer algorithm according to claim 1, characterized in that: The encoder model is: ； in, For the Layer Samples of time, represents the maximum pooling, is the activation function, represents one-dimensional convolution, Indicates Layer Samples of time, Represents a multi-head sparse attention block.

6. The method for wind power prediction based on WTC-Informer algorithm according to claim 1, characterized in that: Through the generative decoder, sparse attention is used to extract the temporal pattern, combined with the connection feature map output by the cross-attention fusion encoder, and the future wind power prediction sequence is output, including: Input a data sequence into the sparse self-attention module, and capture the internal correlation features of the data sequence through a multi-head sparse self-attention mechanism; wherein the data sequence includes a running data sequence and a sequence to be predicted; Inputting the connection feature map into the cross-attention module, using cross-attention to interact the associated features with the connection feature map, integrating the key features captured by the encoder into the prediction generation process of the decoder through the attention mechanism, and obtaining a decoding vector of the hidden representation; The decoded vector is mapped to a wind power prediction result through a fully connected layer to obtain a target sequence.

7. The method for wind power prediction based on WTC-Informer algorithm according to claim 1, characterized in that: The wind power prediction model based on the WTC-Informer algorithm is trained through the following steps: Acquire historical operation data of the wind turbine generator; wherein the historical operation data includes characteristic data of a plurality of wind turbine generators; Preprocessing the historical operation data; wherein the preprocessing includes detecting the historical operation data according to a preset abnormal value determination condition, deleting all abnormal data, and filling the missing data with the value of the previous moment of the vacant moment; Divide the preprocessed historical operation data into training set, validation set and test set, and perform standardization; Perform correlation analysis on the various features of the standardized historical operation data, screen out features with a correlation with wind power higher than a preset value, and obtain the final training set, validation set, and test set; Construct a wind power prediction model based on the WTC-Informer algorithm, and train it according to the final training set and the pre-set loss function to establish the mapping relationship between input features and wind power; and use the validation set to evaluate the performance of the model after each training cycle. Use the early stopping strategy to terminate the training early when the performance on the validation set does not improve in several consecutive training cycles to ensure the generalization ability of the model; The trained model is evaluated by inputting the final test set. When the model meets the preset standards, the model parameters are saved to obtain the pre-trained wind power prediction model based on the WTC-Informer algorithm.

8. A wind power prediction device based on WTC-Informer algorithm, characterized in that: Used to implement a wind power prediction method based on the WTC-Informer algorithm as described in any one of claims 1 to 7; The wind power prediction device comprises: An operation data acquisition module, used to acquire the operation data of the wind turbine; A prediction module, used for inputting the operation data into a pre-trained wind power prediction model based on the WTC-Informer algorithm, performing data processing through the following units, and obtaining a prediction sequence of the active power of the wind turbine; A wavelet transform subunit is used to extract high-frequency components and low-frequency components through wavelet transform according to the operation data and splice them to obtain multi-scale features; The ECA channel attention subunit is used to dynamically assign weights to each feature channel through the ECA channel attention mechanism to obtain weighted multi-scale features; The encoder unit is used to stack multiple sparse self-attention encoding layers into an Informer encoder, which reduces the computational complexity of attention from the standard Transformer Down to , get the connection feature graph; where, Indicates the time complexity, is the length of the input sequence; The decoder unit is used to extract the temporal pattern using sparse attention through a generative decoder, and output the future wind power prediction sequence by combining the connection feature map output by the cross-attention fusion encoder.

9. A wind power prediction device based on WTC-Informer algorithm, characterized in that: It comprises a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement a wind power prediction method based on the WTC-Informer algorithm as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a wind power prediction method based on the WTC-Informer algorithm as described in any one of claims 1 to 6.

Images