Wind speed ultra-short-term prediction method based on transfer learning

Abstract

The invention belongs to the field of wind power generation, and discloses a wind speed ultra-short-term prediction method based on transfer learning, which comprises the following steps of: 1, performing correlation analysis on high-dimensional SCADA (Supervisory Control and Data Acquisition) data through a data preprocessing and feature selection module, and selecting features related to wind speed; step 2, capturing transferable time scale features through a feature extractor and a discriminator; step 3, extracting a key dynamic state of time feature transfer through a self-attention mechanism and a maximum mean value difference; and step 4, performing ultra-short-term prediction on the wind speed through the prediction network. By adopting transfer learning, learning of the target field can be helped by utilizing knowledge of the source field, so that the generalization ability of the model in the new field is improved. The method aims at overcoming the limitation of a traditional method in the aspects of data imbalance, domain transfer and the like, more accurate and stable wind speed prediction is achieved, and reliable technical support is provided for safe operation of a wind power plant and stable supply of a power system.

Description

Technical Field

[0001] The present invention relates to the field of wind power generation, and particularly to a very short-term wind speed prediction method based on transfer learning. Background Art

[0002] In the field of wind power generation, accurate prediction of wind speed is crucial for the safe operation of wind farms and the stability of power systems. Traditional wind speed prediction methods are usually based on mathematical models or statistical methods, such as ARIMA, SARIMA, etc. These methods have limited prediction accuracy in the short term, especially under sudden meteorological changes or complex terrain conditions, where large prediction errors are likely to occur. In recent years, with the rapid development of machine learning and artificial intelligence technologies, data-driven wind speed prediction methods have gradually become a research hotspot. These methods use a large amount of historical wind speed data and related meteorological parameters, such as air pressure, temperature, humidity, etc., combined with machine learning algorithms for modeling and prediction to improve prediction accuracy and stability. Common data-driven methods include support vector machine (SVM), artificial neural network (ANN), decision tree (Decision Tree), etc. However, traditional data-driven methods are often limited by the specific domain or region of the training data and are difficult to meet the wind speed prediction requirements in different environments. Especially when the target scenario data is insufficient or domain transfer occurs, the generalization ability of existing methods is poor, easily leading to a problem of decreased prediction performance.

[0003] Therefore, it is of great significance to develop a very short-term wind speed prediction method based on transfer learning. Transfer learning can help the learning of the target domain by using the knowledge of the source domain, thereby improving the generalization ability of the model in the new domain. This method aims to overcome the limitations of traditional methods in data imbalance, domain transfer, etc., and achieve more accurate and stable wind speed prediction, providing reliable technical support for the safe operation of wind farms and the stable supply of power systems. Summary of the Invention

[0004] The purpose of the present invention is to provide a very short-term wind speed prediction method based on transfer learning to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A very short-term wind speed prediction method based on transfer learning, comprising:

[0007] Step 1, perform a correlation analysis on high-dimensional SCADA data through a data preprocessing and feature selection module to select features related to wind speed;

[0008] Step 2, capture transferable time-scale features through a feature extractor and a discriminator;

[0009] Step 3: Extract the key dynamics of time feature transfer through the self-attention mechanism and the maximum mean discrepancy.

[0010] Step 4: Perform a very short-term prediction of the wind speed through the prediction network.

[0011] Further, the specific steps of Step 1 include:

[0012] Step 1.1: Select all features with an absolute correlation coefficient h with the wind speed greater than a1, where the range of a1 is 0.65 - 0.75, as candidate features.

[0013] Step 1.2: Calculate the correlation coefficients in pairs for the selected features to perform redundancy analysis, and retain one of the features with stronger correlation.

[0014] Step 1.3: According to practical experience, retain one of the two features with a correlation coefficient u greater than a2 as the selected feature, and delete the other as a redundant feature, where the range of a2 is 0.88 - 0.92.

[0015] Further, a1 is 0.7.

[0016] Further, a2 is 0.9.

[0017] Further, the working process of the feature extractor and the discriminator in Step 2 includes:

[0018] Step 2.1: Input the data X s and X t of the source domain and the target domain into the same temporal convolutional network Gse respectively, so as to obtain the source domain spatio-temporal feature Z s and the target domain spatio-temporal feature Z t , and the expressions are as follows:

[0019] Z s , Z t = Gse(X s ), Gse(X t )

[0020] Perform domain labeling on the data samples of the source domain and the target domain respectively, where the source domain data is labeled as 1 and the target domain data is labeled as 0.

[0021] Step 2.2: Introduce a feature discriminator G sd , and use the source domain spatio-temporal feature Z s and the target domain spatio-temporal feature Z t with domain labels as inputs; through the method of adversarial training, optimize the feature discriminator G sd so that it can effectively classify the input features accurately into the corresponding domains; the optimization direction of the temporal convolutional network feature extractor Gse is opposite to that of the feature discriminator Gsd Conversely, to reduce the discriminability of features between domains and promote the feature extractor to generate more domain-invariant features, the expression is as follows:

[0022]

[0023] where i is the i-th sample in the source domain and j is the j-th sample in the target domain, and L domain is the loss of domain adversarial, N s is the number of samples in the source domain, and N t is the number of samples in the target domain.

[0024] Furthermore, the working process of the self-attention mechanism and the maximum mean discrepancy in step 3 includes:

[0025] Step 3.1, dynamically capture domain weights through the self-attention mechanism, map the source domain data to a query vector Q with dimension m s , a key vector K with dimension m s and a value vector V with dimension m s , map the target domain data to a query vector Q with dimension m t , a key vector K with dimension m t and a value vector V with dimension m t , the self-attention mechanism emphasizes those more critical time features when processing the data of the source domain and the target domain by assigning different weights, thereby improving the understanding and adaptation ability of the temporal convolutional network feature extractor Gse to the time-dependent characteristics. The expression is as follows:

[0026]

[0027] where C t and C s are the features of the source domain and the target domain extracted after the self-attention mechanism respectively, W Q , W K W K and b Q , b K b V are the weight matrix and bias matrix corresponding to the multi-head self-attention mechanism; Q, K, V are the query matrix, key matrix and value matrix, the subscripts s and t represent the source domain and the target domain respectively, T is the transpose of this matrix, and softmax(*) is the activation function;

[0028] Step 3.2, the maximum mean discrepancy method calculates the distribution difference between the source domain data and the target domain dataset in the feature space; the feature distribution distance between the source domain and the target domain is constrained by the maximum mean discrepancy method loss, and the specific formula is:

[0029]

[0030] Among them, \(H\) represents the Reproducing Kernel Hilbert Space (RKHS), which provides a more abstract and effective representation space for data. \(\varPhi()\) represents the non-linear mapping function that maps data samples to the Reproducing Kernel Hilbert Space. \(n\) and \(p\) are the numbers of samples in the source domain and the target domain respectively. and are the sample features of the source domain and the target domain respectively. \(L\) MMD is the finally obtained alignment loss function between the source domain and the target domain.

[0031] Furthermore, the specific steps of step 4 include:

[0032] The prediction network includes a multi-layer perceptron. After capturing the spatio-temporal features of transferability, these features are input into the multi-layer perceptron for the final wind speed prediction, and the MSE is used as the prediction loss function. \(y\) is the true value, is the predicted value.

[0033] The overall technical means of the present invention have significant technical effects in the ultra-short-term wind speed prediction of wind turbines, overcoming multiple problems in the current technical status, including but not limited to:

[0034] 1) Data scarcity: Traditional wind speed prediction methods usually rely on a large amount of historical data for modeling, and sufficient historical data may not be available in a new wind farm. The method based on transfer learning overcomes the data scarcity by utilizing data from other wind farms.

[0035] 2) High modeling cost: Traditional deep learning models usually require a large amount of computing resources and time for training, while the transfer learning method can reduce the cost and time of model training by utilizing pre-trained models or data from other wind farms.

[0036] 3) Instability of new wind farm data: In a new wind farm, the distribution and characteristics of data may be different from those of existing wind farms, and it is difficult for traditional models to adapt to this data instability. The method based on transfer learning improves the adaptability of the model to new wind farm data by transferring knowledge between different wind farms.

[0037] 4) Insufficient consideration of time dynamics: Traditional methods often lack consideration of time dynamics and cannot capture the changing patterns of data over time. The method based on transfer learning combines the transfer of time series features, enhancing the model's understanding and prediction ability of time dynamics.

[0038] 5) Insufficient sensitivity to dynamic features: Some traditional methods are not sensitive enough to the dynamic features in wind field data and cannot accurately predict the instantaneous changes in wind speed. The method based on transfer learning improves the model's perception ability of dynamic features by dynamically transferring knowledge, thus achieving more accurate ultra-short-term wind speed prediction. Description of the Drawings

[0039] Figure 1 This is the overall flowchart of the present invention.

[0040] Figure 2 This is the framework diagram of feature selection. Detailed Embodiments

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0042] Please refer to Figure 1 and Figure 2 , an ultra-short-term wind speed prediction method based on transfer learning, including:

[0043] Step 1, perform correlation analysis on high-dimensional SCADA data through a data preprocessing and feature selection module to select features related to wind speed. Specifically, it includes:

[0044] Step 1.1, select all features with an absolute correlation coefficient h with wind speed greater than a1 as candidate features, where the range of a1 is 0.65 - 0.75, preferably 0.7;

[0045] Step 1.2, calculate the correlation coefficients of the selected features in pairs for redundancy analysis, and retain one of the features with stronger correlation;

[0046] Step 1.3, according to practical experience, retain one of the two features with a correlation coefficient u greater than a2 as the selected feature, and delete the other as a redundant feature, where the range of a2 is 0.88 - 0.92, preferably 0.9.

[0047] Step 2, capture transferable time-scale features through a feature adversarial adaptation module. The working process of the feature adversarial adaptation module includes:

[0048] Step 2.1, the data X s and X tThey are respectively input into the same temporal convolutional network Gse to obtain the source domain spatio-temporal feature Z s and the target domain spatio-temporal feature Z t , and the expression is as follows:

[0049] Z s , Z t = Gse(X s ), Gse(X t )

[0050] Domain labels are respectively given to the data samples of the source domain and the target domain, where the source domain data is labeled as 1 and the target domain data is labeled as 0;

[0051] Step 2.2, introduce a feature discriminator G sd , and use the source domain spatio-temporal feature Z s with domain labels and the target domain spatio-temporal feature Z t as inputs; through the way of adversarial training, optimize the feature discriminator G sd so that it can effectively and accurately classify the input features into the corresponding domains; the optimization direction of the temporal convolutional network feature extractor Gse is opposite to that of the feature discriminator G sd , to reduce the discriminability of features between domains and promote the feature extractor to generate more domain-invariant features. The expression is as follows:

[0052]

[0053] where L dis is the loss of domain adversarial, N s is the number of source domain samples, N t is the number of target domain samples, where i is the i-th sample in the source domain and j is the j-th sample in the target domain.

[0054] Step 3, extract the key dynamics of time feature transfer through the time attention adaptation module. The working process of the time attention adaptation module includes:

[0055] Step 3.1, dynamically capture domain weights through the self-attention mechanism, which enables the present invention to quantify the influence of each local time feature in the source domain and the target domain. The self-attention mechanism emphasizes the more critical time features when processing the data of the source domain and the target domain by assigning different weights, so as to improve the understanding and adaptation ability of time-dependent characteristics. The expression is as follows:

[0056]

[0057] C t and C s are respectively the features of the source domain and the target domain extracted after passing through the self-attention mechanism, and W and b are the parameters corresponding to the multi-head self-attention mechanism; where C tand C s are the features of the source domain and the target domain extracted after the self-attention mechanism respectively, and W Q , W K W K and b Q , b K b V are the weight matrix and bias matrix corresponding to the multi-head self-attention mechanism; Q, K, and V are the query matrix, key matrix, and value matrix respectively. The subscripts s and t represent the source domain and the target domain respectively, T represents the transpose of this matrix, and softmax(*) is the activation function.

[0058] Step 3.2, the maximum mean discrepancy (MMD) is a method for measuring the difference between two distributions and is widely used in domain adaptation problems. In this step, the feature distribution distance between the source domain and the target domain is constrained by the maximum mean discrepancy method loss, and the specific formula is:

[0059]

[0060] where φ(*) represents mapping the data to the reproducing kernel Hilbert space (RKHS), n and m are the number of samples in the source domain and the target domain respectively, and are the sample features of the source domain and the target domain respectively. Among them, H represents the reproducing kernel Hilbert space (RKHS), which provides a more abstract and effective representation space for the data. Φ() represents the non-linear mapping function that maps the data samples to the RKHS space. n and p are the number of samples in the source domain and the target domain respectively, and are the sample features of the source domain and the target domain respectively, and L MMD is the finally obtained alignment loss function between the source domain and the target domain

[0061] Step 4, the wind speed is predicted in the very short term through the prediction network. Specifically, it includes:

[0062] The prediction network includes a multi-layer perceptron. After capturing the transferable spatio-temporal features, these features are input into a multi-layer perceptron (MLP) for the final wind speed prediction.

[0063] The method proposed by the present invention consists of four components: the first part is the data preprocessing and feature selection module, whose function is to perform correlation analysis on the high-dimensional SCADA data and select the variables related to the wind speed; the second part is the feature adversarial adaptation module, which uses the domain adversarial method to capture the transferable time-scale features; the third part is the time attention adaptation module, which focuses on the key dynamics of time feature transfer; the fourth part is the prediction network, which aims to learn the task-driven transferable knowledge. Each part will be introduced in detail below.

[0064] Data preprocessing and feature selection module: Feature selection is based on historical data and is an offline behavior. The selected features are closely related to the input variables. The present invention uses the quadratic correlation coefficient method to select feature variables. The specific process is as Figure 1 shown.

[0065] For the SCADA data of the target wind turbine, calculate the Pearson correlation coefficient between each feature and the wind speed feature. First, select all features with an absolute correlation coefficient h greater than 0.7 with the wind speed as candidate features. Then, calculate the correlation coefficients of the selected features in pairs for redundancy analysis, and retain one of the features with stronger correlation. According to practical experience, retain one of the two features with a correlation coefficient u greater than 0.9, and delete the other as a redundant feature.

[0066] Feature adversarial adaptation module: The present invention uses the temporal convolutional network (TCN) extraction technique to select the spatio-temporal features of the SCADA high-dimensional data related to the wind speed. The temporal convolutional network (TCN) is a deep learning architecture specifically used for the modeling and prediction of time series data. Structurally, the TCN network uses causal convolution and dilated convolution as internal standard convolutional layers, encapsulates two convolutional layers and the same mapping layer into a residual module, and then stacks it into the TCN network. Through the TCN network, the present invention obtains the spatio-temporal features Z s , Z t . According to the obtained Z s , Z t , the present invention further labels these samples as 1 or 0, where 1 indicates that the source domain data is real data, and 0 indicates that the target domain data is fake data. Then, the present invention introduces the feature discriminator G sd . Specifically, the present invention mixes the labeled Z s , Z t as the input of Gsd. Through the adversarial objective, the discriminator G sd is optimized to classify the input spatial features into different domains, while the feature extractor Gse is optimized in the opposite direction.

[0067] Temporal Attention Adaptation Module: The temporal attention adaptation module is designed for temporal perception feature transfer because domain differences are reflected in the temporal dimension. Specifically, first, a self-attention mechanism is utilized to capture domain weights, which represent the influence of local temporal features in the source domain and the target domain. Then, a difference-based domain adaptation method, namely MMD, is adopted to capture the global temporal features in the source domain and the target domain. Based on the local temporal features, the present invention can calculate the global temporal features by summing up each local temporal feature. However, not all local temporal features contribute equally to the global temporal features, so the internal correlation between local temporal features is important. The present invention introduces a self-attention layer to consider both local and global aspects. Specifically, for feature adaptation between the source domain and the target domain, the present invention designs a domain weight to capture those important local temporal features while discarding those local temporal features that may have a negative impact on domain transfer.

[0068] Prediction Network: After extracting the transferable spatio-temporal features, the present invention uses MLP as the final wind speed prediction network module.

[0069] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A very short-term wind speed prediction method based on transfer learning, characterized in that, Including: Step 1: Perform correlation analysis on high-dimensional SCADA data through the data preprocessing and feature selection module to select features related to wind speed. Step 2: Capture transferable time-scale features through the feature extractor and discriminator. Step 3: Extract the key dynamics of time feature transfer through the self-attention mechanism and maximum mean discrepancy. Step 4: Perform ultra-short-term prediction of wind speed through the prediction network.

2. The short-term wind speed prediction method based on transfer learning according to claim 1, characterized in that The specific content of Step 1 includes: Step 1.1: Select all features with an absolute correlation coefficient h with wind speed greater than a1 as candidate features, where the range of a1 is 0.65 - 0.

75. Step 1.2: Calculate the correlation coefficients of the selected features in pairs for redundancy analysis, and retain one of the features with stronger correlation. Step 1.3: According to practical experience, retain one of the two features with a correlation coefficient u greater than a2 as the selected feature, and delete the other as a redundant feature, where the range of a2 is 0.88 - 0.

92.

3. The short-term wind speed prediction method based on transfer learning according to claim 2, wherein The value of a1 is 0.

7.

4. A short-term wind speed prediction method based on transfer learning according to claim 2, characterized in that The value of a2 is 0.

9.

5. A short-term wind speed prediction method based on transfer learning according to claim 1, characterized in that, The working process of the feature extractor and discriminator in Step 2 includes: Step 2.1, input the data X of the source domain and the target domain s and X t into the same temporal convolutional network Gse respectively, so as to obtain the source domain spatio-temporal feature Z s and the target domain spatio-temporal feature Z t , and the expression is as follows: Z s ,Z t = Gse(X s ), Gse(X t ) Perform domain marking on the data samples of the source domain and target domain respectively, where the source domain data is marked as 1 and the target domain data is marked as 0. Step 2.2, introduce a feature discriminator G sd , and use the source domain spatio-temporal feature Z with domain labels s and the target domain spatio-temporal feature Z t as inputs; through adversarial training, optimize the feature discriminator G sd so that it can effectively and accurately classify the input features into the corresponding domains; the optimization direction of the temporal convolutional network feature extractor Gse is opposite to that of the feature discriminator G sd to reduce the discriminability of features between domains and promote the feature extractor to generate more domain-invariant features. The expression is as follows: where \(i\) is the \(i\)-th sample in the source domain and \(j\) is the \(j\)-th sample in the target domain, \(L\) domain is the loss of domain adversarial, \(N\) s is the number of source domain samples, \(N\) t is the number of target domain samples.

6. The short-term wind speed prediction method based on transfer learning according to claim 1, characterized in that The working process of the self-attention mechanism and maximum mean discrepancy in Step 3 includes: Step 3.1, dynamically capture domain weights through the self-attention mechanism, and map the source domain data to the query vector Q with dimension m s , the key vector K with dimension m s and the value vector V with dimension m s , map the target domain data to the query vector Q with dimension m t , the key vector K with dimension m t and the value vector V with dimension m t , the self-attention mechanism emphasizes those more critical time features when processing the data of the source domain and the target domain by assigning different weights, so as to improve the understanding and adaptation ability of the temporal convolutional network feature extractor Gse to the time dependence characteristics. The expression is as follows: Among them, C t and C s are the features of the source domain and the target domain extracted after the self-attention mechanism respectively. W Q , W K W K and b Q , b K b V are the weight matrix and bias matrix corresponding to the multi-head self-attention mechanism; Q, K, V are the query matrix, key matrix and value matrix. The subscripts s and t represent the source domain and the target domain respectively. T represents the transpose of the matrix, and softmax(*) is the activation function; Step 3.2: The maximum mean discrepancy method calculates the distribution difference between the source domain data and the target domain dataset in the feature space; the feature distribution distance between the source domain and the target domain is constrained by the maximum mean discrepancy method loss, and the specific formula is: Among them, H represents the reproducing kernel Hilbert space, Φ() represents the non-linear mapping function that maps data samples to the reproducing kernel Hilbert space, n and p are the numbers of samples in the source domain and the target domain respectively, C si and C tj are the sample features of the source domain and the target domain respectively, L MMD is the finally obtained alignment loss function between the source domain and the target domain.

7. A very short-term wind speed prediction method based on transfer learning according to claim 1, characterized in that The specific content of Step 4 includes: The prediction network includes a multi-layer perceptron. After capturing the spatio-temporal features with transferability, these features are input into the multi-layer perceptron for the final wind speed prediction, and the MSE is used as the prediction loss function. y is the true value, is the predicted value.

Images