CN117893362A - A multi-temporal and spatial scale offshore wind power feature screening and enhanced power prediction method - Google Patents

Abstract

The invention discloses a multi-time space-scale offshore wind power characteristic screening and enhanced power prediction method, which comprises the steps of firstly acquiring a multi-source data set of a target offshore wind power plant, screening a motif spectrum cluster expansion optimal sensitive factor according to acquired ocean current data, and reducing the dimension of the data; the characteristics of the offshore wind power are enhanced by taking the internal and external relations of the characteristics into consideration through a multidimensional attention mechanism, a multi-attention depth fusion mechanism is developed from different dimensions, deep coupling relations on time, space and characteristics are mined, and corresponding weights are given to different characteristics in a self-adaptive mode. According to different influence degrees of geographic positions on the output of the offshore wind turbine, the implicit correlation of different areas is excavated, different weights are self-adaptively given by using the graph annotation force, a new feature matrix is finally obtained, a prediction model is transmitted, prediction is carried out from a plurality of time scales, and the power time sequence of the corresponding offshore wind turbine at different time scales is obtained. The method effectively improves the precision of offshore wind power prediction.

Description

A multi-temporal and spatial scale offshore wind power feature screening and enhanced power prediction method

Technical Field

The present invention relates to the technical field of multi-temporal and spatial scale prediction of offshore wind power, and more specifically, to a multi-temporal and spatial scale offshore wind power feature screening and enhanced power prediction method.

Background technique

Large-scale development, deep-sea floating high-power offshore equipment, and intelligent operation and maintenance are injecting strong impetus into the development of offshore wind power. Since wind energy itself is random, it directly leads to the instability of offshore wind power. Therefore, it is necessary to accurately predict offshore wind power in advance, which can provide a basis for the reliable operation of the power grid and the power reserve plan. Accurate offshore wind power prediction is of great significance to the power system.

As a time series with strong randomness and volatility, the prediction accuracy of offshore wind power is closely related to the quality of meteorological characteristic data and the temporal and spatial scales. Since the influencing factors and meteorological characteristics of offshore wind power are more complex and contain a large number of complex data features, effective feature screening and enhancement are necessary. So far, the existing offshore wind power correlation prediction method has failed to fully consider the processing method under a large number of data features, resulting in low prediction accuracy of offshore wind power.

In order to screen and enhance the features of offshore wind farms with a large number of complex data features, an ultra-short-term power prediction model considering the spatiotemporal characteristics of multiple offshore wind power units is proposed in the prior art. First, the dynamic time warping distance algorithm is used and the DTW algorithm is improved by adding the idea of abstraction and de-abstraction. At the same time, the bus and geographic information are considered to cluster the units to form a unit group to quantify and measure the time series similarity between units; then, the Transformer model based on the attention mechanism is used and the probabilistic improvement is made in the attention module to reduce the power prediction time; finally, the sequence that comprehensively considers the spatiotemporal characteristics and location information is predicted and analyzed. However, since the influencing factors and meteorological characteristics of offshore wind power are more complex, regional division is required when considering multiple spatiotemporal scales, and multiple attentions are integrated from different dimensions for effective feature screening and enhancement to improve the accuracy of offshore wind farm power prediction.

Summary of the invention

In order to overcome the defects in the above-mentioned prior art, the present invention provides a power prediction method and system for offshore wind power feature screening and enhancement at multiple time and space scales, so as to realize feature screening and enhancement at different spatial scales, and predict at different time scales, so as to improve the prediction accuracy of offshore wind power.

In order to solve the above technical problems, the technical solution adopted by the present invention is:

A multi-temporal and spatial scale offshore wind power feature screening and enhanced power prediction method comprises the following steps:

S1. Obtain multi-source data sets of the target and adjacent offshore wind farms, including the power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus, and ocean current data of the target offshore wind farm, and perform preliminary processing;

S2. Divide the pre-processed historical meteorological record data of offshore wind farms into training samples and test samples;

S3. Perform motif spectral clustering based on the acquired ocean current data, use Laplace eigenmap to screen the best sensitive factors, reduce the data dimension, and divide the feature data into three areas: nearshore, shallow sea, and deep sea; the processed offshore wind farm power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus, and ocean current data form a feature matrix ,/> ,/> ;

S4. Construct a prediction model of multi-dimensional attention fusion mechanism and bidirectional gated recurrent unit network;

S5. The feature matrix ,/> ,/> Input into the prediction model, consider the internal and external relationship of the features through the multi-dimensional attention mechanism in the prediction model, carry out offshore wind power feature enhancement from different dimensions with the multi-attention deep fusion mechanism, explore the deep coupling relationship in time, space and features, and adaptively give corresponding weights to different features;

S6. The three feature matrices of nearshore, shallow sea and deep sea are further mined according to the different degrees of influence of geographical location on the output of offshore wind turbines. Different weights are adaptively assigned using graph attention, and finally a new feature matrix is obtained. ;

S7. The feature matrix after adding weights and enhancing features The data is transmitted to a bidirectional gated recurrent unit network, which mines the implicit relationship in the feature matrix.

S8. Use the trained prediction model to predict the power of the target offshore wind farm from multiple time scales, obtain the power time series of the corresponding offshore wind farm at different time scales, and realize offshore wind power prediction at multiple time and space scales.

According to the above technical means, the present invention provides an offshore wind power prediction method based on motif spectrum clustering and multi-dimensional attention mechanism. First, a multi-source data set of the target offshore wind farm is obtained. After preliminary processing, the motif spectrum clustering is performed according to the acquired ocean current data to screen the best sensitive factors, reduce the dimension of the data, and divide the feature data into three regions: nearshore, shallow sea, and deep sea. The internal and external relationships of the features are considered through the multi-dimensional attention mechanism, and the offshore wind power feature enhancement of the multiple attention deep fusion mechanism is carried out from different dimensions, and the deep coupling relationship in time, space, and features is excavated, and the corresponding weights of different features are adaptively given. The three feature matrices of nearshore, shallow sea, and deep sea are further excavated according to the different degrees of influence of geographical location on the output of offshore wind turbines, and different weights are adaptively given using graph attention. Finally, a new feature matrix is obtained and transmitted to a bidirectional gated recurrent unit network, and a trained prediction model is used to predict from multiple time scales to obtain the power time series of the corresponding offshore wind farm at different time scales, so as to realize the offshore wind power prediction at multiple time and space scales. The present invention can effectively improve the accuracy of offshore wind power prediction.

Furthermore, in step S1, the specific process of preliminary data processing is as follows: based on the measured meteorological data near the offshore wind farm area, the background field systems such as the Global Forecasting System (GFS) and the Weather Research Forecast (WRF) are corrected to obtain NWP data with a horizontal resolution of 3km×3km grid. The missing data in the historical meteorological record data of the offshore wind farm are filled to obtain complete historical meteorological record data of the wind farm; the power sequence, wind speed sequence, temperature sequence, pressure sequence, cumulus sequence and ocean current sequence are processed by min-max normalization to obtain the processed power sequence P, wind speed sequence WS, temperature sequence T, humidity sequence H, precipitation sequence PRECIP, pressure sequence PA, cumulus sequence CU and ocean current sequence OC, while the wind direction sequence is processed by sine and cosine to obtain the wind direction sine WDS and wind direction cosine WDC.

Furthermore, in step S2, the pre-processed historical meteorological record data of the offshore wind farm is divided into training samples and test samples in a ratio of 8:2.

Furthermore, in step S3, motif spectrum clustering is used for the processed sample features, and Laplace eigenmap is used to screen the best sensitive factors and reduce the dimension of the data. The specific steps are as follows:

S31. Constructing similarity matrix , according to the distance between data samples, calculate the similarity between samples;

S32. Constructing the adjacency matrix , constructed using the Gaussian distance method:

In the formula, 、/> is the i-th point and the j-th point in n samples (i, j = 1, .., n), s is the standard deviation, and e is the exponential function;

S33. Calculate the order moment D and Laplace matrix L:

In the formula, is the i-row and j-column moment matrix,/> is the Laplace matrix of i rows and j columns;

S34. Perform eigendecomposition on the Laplace matrix to obtain eigenvectors and eigenvalues; select the first k eigenvectors as new data representations according to the size of the eigenvalues; and generate a low-dimensional eigenmatrix;

S35. Perform K-Means clustering in the new feature space after dimensionality reduction and divide it into three categories: nearshore, shallow sea, and deep sea, and obtain the feature matrix ,/> ,/> ;

=［M ₁ ,M ₂ ,...,M _n ］,/> , where _Mn represents the matrix composed of the characteristics of the nth wind farm from time t -1 to time t- m.

Furthermore, in step S4, the specific steps of constructing the multi-dimensional attention fusion mechanism are as follows:

The temporal attention module and the feature attention module are connected in series, and the input is a H×W×C matrix. The feature maps obtained after processing by the two attention sub-modules are subjected to element-wise addition operation to fuse the two features together. The output result is processed by the ReLU activation function and used as the input of the channel attention module and the spatial attention module. The two modules are connected together using a series structure. In these two attention modules, average pooling and maximum pooling operations are used simultaneously to obtain different feature information from different perspectives.

Furthermore, the temporal attention module includes a convolution layer, an FC layer, a softmax layer, a sigmoid activation function and a global average pooling layer; in the temporal attention module, the global spatial average pooling GAP is applied to the feature matrix to ensure that the temporal attention module has a low computational cost; then multiple 1D convolutions with nonlinearity are used over the entire time domain to generate a position-sensitive importance map to enhance frame-by-frame features; then, after the FC layer, a channel-adaptive kernel is generated based on the global temporal information in each channel, and the temporal attention module weight coefficient Mt is obtained after the softmax layer; finally, the weight coefficient and the feature matrix are used to generate the temporal attention module weight coefficient Mt. Multiplying them will give you the scaled new feature matrix;

In the formula, GAP is the global spatial average pooling, Conv1D is the one-dimensional convolution, X is the input sample, is the sigmoid activation function, FC is the fully connected layer, Mt is the weight coefficient of the temporal attention module, /> is the new feature matrix; δ represents the attention weight, /> is the vector product.

Furthermore, the feature attention module includes a convolution layer, a Relu activation function, a sigmoid activation function and a pooling layer; in the feature attention module, the input feature is first passed through the pooling layer, and then through two convolution layers, the activation function is Relu, and then the obtained feature is fed back through a Sigmoid activation function to obtain the adaptive weight coefficient of the feature in different environments, and the weight coefficient is obtained. ; Finally, use the weight coefficient and the original feature matrix/> Multiplying them will give you the scaled new feature matrix;

In the formula, Conv is convolution, is the Relu activation function, /> is the new feature matrix.

Furthermore, the spatial attention module includes a convolution layer, a maximum pooling layer, an average pooling layer, and a sigmoid activation function. In the spatial attention module, firstly, average pooling and maximum pooling of a channel dimension are performed to obtain two H×W×1 channels and spliced together. Then, a C×C convolution layer is passed, and the activation function is Sigmoid to obtain a weight coefficient Ms. Finally, the weight coefficient and the feature matrix are used to obtain the weight coefficient Ms. Multiplying them will give you the scaled new feature matrix;

In the formula, is average pooling,/> For maximum pooling, is the new feature matrix.

Furthermore, the channel attention module includes a convolution layer, a maximum pooling layer, an average pooling layer, and a sigmoid activation function; in the channel attention module, firstly, global average pooling and maximum pooling are performed in space to obtain two 1×1×C channels, and then they are respectively sent to a two-layer shared neural network, the number of neurons in the first layer is C/r, the activation function is Relu, and the number of neurons in the second layer is C; then the two features are added and then passed through a Sigmoid activation function to obtain a weight coefficient Mc; finally, the weight coefficient and the original feature matrix are used to obtain the weight coefficient Mc. Multiplying them will give you the scaled new feature matrix;

In the formula, MLP is a multi-layer perceptron, is the new feature matrix.

Further, in step S5, the feature matrix ,/> ,/> Input into the prediction model, consider the internal and external relationship of the features through the multi-dimensional attention mechanism in the prediction model, carry out offshore wind power feature enhancement from different dimensions with multiple attention deep fusion mechanism, explore the deep coupling relationship in time, space and features, and adaptively give corresponding weights to different features; based on time attention/> , Feature Attention/> , spatial attention/> , channel attention , carry out feature enhancement in depth;

In the formula, The weight matrix calculated for attention,/> is the neural network corresponding to attention,/> is the eigenvalue corresponding to the calculated dimension;

The obtained weights are multiplied by the characteristics of the corresponding offshore wind farms, and the weighted feature matrix is obtained: ,/> ,/> ;

.

Furthermore, step S6 specifically includes:

S61. Define the weight matrix W and convert the feature matrix into adjacent nodes: ,/> is the jth input sample (h=h1,…hn);

S62. Concatenate adjacent nodes i and j and map them into scalars, Calculate the attention function: /> , where / > The original attention contribution calculated for a node is used for normalization in the next step;

S63. Pass the adjacent node matrix through the leakyRelu layer and then the softmax layer, calculate the contribution of each neighbor node j of node i to i and normalize the contribution of each neighbor node j;

S64. After calculating the contribution of each neighboring node of node i, the feature sum of all neighboring nodes of node i is updated according to the weight; as the final output of node i, ;

Let the number of output neurons be 3, corresponding to the weights of the three regions of nearshore, shallow sea and deep sea, and finally get the new feature matrix ;

in, ;

.

Further, in step S7, the bidirectional gated recurrent unit network is constructed as follows: with the feature matrix As input, a multi-layer cascade bidirectional gated recurrent unit network is built, which consists of a forward GRU and a backward GRU, and the activation function is tanh;

In the formula, 、/> 、/> 、/> 、/> 、/> is the weight parameter matrix, /> 、/> 、/> is the bias parameter matrix, /> is matrix multiplication, /> is the Sigmoid function, /> To reset the gate, /> To update the gate, /> is the candidate state of the hidden layer at the current moment, /> is the current implicit state, /> is the implicit state of the previous moment, /> is the input status at the current moment, /> Calculate for the reverse pass/> and/> is the hidden state of the forward GRU and the backward GRU, and F is the output merging method in both directions.

Further, in step S8, the time scale is n, where n is an integer, and n 2.

Assume that the time scale of the current forecasting system is i and the input is ,/> is the running data of the previous moment of time scale i ; the output of the prediction system is/> ,/> The prediction data for the current time scale; use the trained prediction model to predict from multiple time scales, predict the power of the target offshore wind farm, obtain the power time series of the corresponding offshore wind farm at different time scales, complete the ultra-short-term prediction and short-term prediction of offshore wind power respectively, and realize the offshore wind power prediction at multiple time and space scales;

In the formula, is the predicted output load result at the i -th scale,/> is the input matrix for the time scales i - n -1 to i -1.

The present invention also provides a multi-temporal and spatial scale offshore wind power feature screening and enhanced power prediction system, comprising:

A data acquisition unit, used to obtain multi-source data sets of the target and adjacent offshore wind farms, including power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus cloud cover, and ocean current data of the target offshore wind farm;

A preprocessing unit is used to preprocess the feature sequence obtained by the data acquisition unit to obtain a feature vector, and use the offshore wind power time series as a prediction target sample, and divide the feature vector and the offshore wind power time series into a training set and a verification set respectively;

The multi-spatial-scale feature screening unit is used to perform motif spectral clustering based on the ocean current data acquired by the data acquisition unit, use Laplace eigenmap to screen the best sensitive factors, reduce the dimension of the data, and divide the feature data into three areas: nearshore, shallow sea, and deep sea, forming three feature matrices respectively;

The feature enhancement unit considers the internal and external relationships of features through the multi-dimensional attention mechanism in the prediction model, and performs offshore wind power feature enhancement with a multi-attention deep fusion mechanism from different dimensions, digs into the deep coupling relationship in time, space, and features, and adaptively gives corresponding weights to different features; based on time attention, feature attention, space attention, and channel attention, it deeply performs feature enhancement, multiplies the attention weight with the input feature map, obtains the weighted feature matrix, and sends it to the multi-time scale prediction unit;

The multi-time scale prediction unit is used to input the weighted feature matrix into the bidirectional gated recurrent unit network, use the trained prediction model to make predictions from multiple time scales, predict the power of the target offshore wind farm, obtain the power time series of the corresponding offshore wind farm at different time scales, and complete the offshore wind power prediction at multiple time and space scales.

Compared with the prior art, the beneficial effects are as follows: the method for power prediction of offshore wind power feature screening and enhancement at multiple spatiotemporal scales provided by the present invention is implemented based on motif spectrum clustering and multi-dimensional attention mechanism, wherein motif spectrum clustering carries out optimal sensitive factor screening, reduces the dimension of data, and effectively enhances deep coupling spatiotemporal features from the spatial scale; and the multi-dimensional attention mechanism considers the internal and external relationship of features, and carries out offshore wind power feature enhancement of multiple attention deep fusion mechanisms from different dimensions, digs deep coupling relationships in time, space, and features, and uses graph attention to the three feature matrices of nearshore, shallow sea, and deep sea according to the different degrees of influence of geographical location on the output of offshore wind turbines, further digs implicit correlations in different regions, which is helpful for improving the accuracy of offshore wind power prediction; prediction is carried out from multiple time scales to realize offshore wind power prediction at multiple spatiotemporal scales, which can capture future long-term trends and improve reliability. It has certain practical significance for offshore wind power prediction. The present invention can effectively improve the accuracy of offshore wind power prediction.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of the method for offshore wind power feature screening and enhanced power prediction at multiple time and space scales of the present invention.

FIG2 is a schematic diagram of the prediction model.

Figure 3 shows the weights assigned to wind farms in different regions by the multiple attention mechanisms.

FIG4 is a diagram showing the effect of offshore wind power feature screening and enhanced power prediction at multiple time and space scales according to the present invention.

Figure 5 is a schematic diagram of the process flow of the offshore wind power feature screening and enhanced power prediction system at multiple temporal and spatial scales.

Detailed ways

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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. The present invention is described in one of the embodiments in combination with the specific implementation methods. Among them, the drawings are only used for exemplary descriptions, and only schematic diagrams are shown, not physical drawings, and cannot be understood as limitations on this patent; in order to better illustrate the embodiments of the present invention, some parts of the drawings will be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

In the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right" and the like indicate an orientation or positional relationship based on the orientation or positional relationship shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only used for exemplary explanations and cannot be understood as limitations on this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances. In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In addition, the meaning of "and/or" appearing in the full text is to include three parallel schemes. Taking "A and/or B" as an example, it includes scheme A, or scheme B, or schemes that satisfy both A and B.

Embodiment 1:

As shown in FIG1 , a method for screening and enhancing offshore wind power characteristics at multiple time and space scales includes the following steps:

S1. Obtain multi-source data sets of the target and adjacent offshore wind farms, including the power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus, and ocean current data of the target offshore wind farm, and perform preliminary processing;

S2. Divide the pre-processed historical meteorological record data of offshore wind farms into training samples and test samples;

S3. Perform motif spectral clustering based on the acquired ocean current data, use Laplace eigenmap to screen the best sensitive factors, reduce the data dimension, and divide the feature data into three areas: nearshore, shallow sea, and deep sea; the processed offshore wind farm power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus, and ocean current data form a feature matrix ,/> ,/> ;

S4. Construct a prediction model of multi-dimensional attention fusion mechanism and bidirectional gated recurrent unit network;

S5. The feature matrix ,/> ,/> Input into the prediction model, consider the internal and external relationship of the features through the multi-dimensional attention mechanism in the prediction model, carry out offshore wind power feature enhancement from different dimensions with the multi-attention deep fusion mechanism, explore the deep coupling relationship in time, space and features, and adaptively give corresponding weights to different features;

S6. The three feature matrices of nearshore, shallow sea and deep sea are further mined according to the different degrees of influence of geographical location on the output of offshore wind turbines. Different weights are adaptively assigned using graph attention, and finally a new feature matrix is obtained. ;

S7. The feature matrix after adding weights and enhancing features The data is transmitted to a bidirectional gated recurrent unit network, which mines the implicit relationship in the feature matrix.

S8. Use the trained prediction model to predict the power of the target offshore wind farm from multiple time scales, obtain the power time series of the corresponding offshore wind farm at different time scales, and realize offshore wind power prediction at multiple time and space scales.

According to the above technical means, the present invention provides an offshore wind power prediction method based on motif spectrum clustering and multi-dimensional attention mechanism. First, a multi-source data set of the target offshore wind farm is obtained. After preliminary processing, the motif spectrum clustering is performed according to the acquired ocean current data to screen the best sensitive factors, reduce the dimension of the data, and divide the feature data into three regions: nearshore, shallow sea, and deep sea. The internal and external relationships of the features are considered through the multi-dimensional attention mechanism, and the offshore wind power feature enhancement of the multiple attention deep fusion mechanism is carried out from different dimensions, and the deep coupling relationship in time, space, and features is excavated, and the corresponding weights of different features are adaptively given. The three feature matrices of nearshore, shallow sea, and deep sea are further excavated according to the different degrees of influence of geographical location on the output of offshore wind turbines, and different weights are adaptively given using graph attention. Finally, a new feature matrix is obtained and transmitted to a bidirectional gated recurrent unit network, and a trained prediction model is used to predict from multiple time scales to obtain the power time series of the corresponding offshore wind farm at different time scales, so as to realize the offshore wind power prediction at multiple time and space scales. The present invention can effectively improve the accuracy of offshore wind power prediction.

Example 2

In this embodiment, the specific process of preliminary data processing in step S1 is as follows: based on the measured meteorological data near the offshore wind farm area, the background field systems such as the Global Forecasting System (GFS) and the Weather Research Forecast (WRF) are corrected to obtain NWP data with a horizontal resolution of 3km×3km grid. The missing data in the historical meteorological record data of the offshore wind farm are filled to obtain complete historical meteorological record data of the wind farm; the power sequence, wind speed sequence, temperature sequence, pressure sequence, cumulus sequence and ocean current sequence are processed by min-max normalization to obtain the processed power sequence P, wind speed sequence WS, temperature sequence T, humidity sequence H, precipitation sequence PRECIP, pressure sequence PA, cumulus sequence CU and ocean current sequence OC, and the wind direction sequence is processed by sine and cosine to obtain the wind direction sine WDS and wind direction cosine WDC.

Furthermore, in step S2, the pre-processed historical meteorological record data of the offshore wind farm is divided into training samples and test samples in a ratio of 8:2.

In this embodiment, motif spectral clustering is used for the processed sample features, and Laplace eigenmap is used to screen the best sensitive factors and reduce the dimension of the data. The specific steps are as follows:

S31. Constructing similarity matrix , according to the distance between data samples, calculate the similarity between samples;

S32. Constructing the adjacency matrix , constructed using the Gaussian distance method:

In the formula, 、/> is the i -th point and the j-th point in n samples ( i , j =1,..,n), s is the standard deviation, and e is the exponential function;

S33. Calculate the order moment D and Laplace matrix L:

In the formula, is the i- row and j -column moment matrix,/> is the Laplace matrix of i rows and j columns;

S34. Perform eigendecomposition on the Laplace matrix to obtain eigenvectors and eigenvalues; select the first k eigenvectors as new data representations according to the size of the eigenvalues; and generate a low-dimensional eigenmatrix;

S35. Perform K-Means clustering in the new feature space after dimensionality reduction and divide it into three categories: nearshore, shallow sea, and deep sea, and obtain the feature matrix ,/> ,/> ;

=［M ₁ ,M ₂ ,...,M _n ］,/> , where _Mn represents the matrix composed of the characteristics of the nth wind farm from time t -1 to time t- m.

In this embodiment, the specific steps of constructing a multi-dimensional attention fusion mechanism are as follows:

The temporal attention module and the feature attention module are connected in series, and the input is a H×W×C matrix. The feature maps obtained after processing by the two attention sub-modules are subjected to element-wise addition operation to fuse the two features together. The output result is processed by the ReLU activation function and used as the input of the channel attention module and the spatial attention module. The two modules are connected together using a series structure. In these two attention modules, average pooling and maximum pooling operations are used simultaneously to obtain different feature information from different perspectives.

The temporal attention module includes a convolution layer, an FC layer, a softmax layer, a sigmoid activation function and a global average pooling layer; in the temporal attention module, the global spatial average pooling GAP is applied to the feature matrix to ensure that the temporal attention module has a low computational cost; then multiple 1D convolutions with nonlinearity are used on the entire time domain to generate a position-sensitive importance map to enhance the frame-by-frame features; then, after the FC layer, a channel-adaptive kernel is generated based on the global temporal information in each channel, and the temporal attention module weight coefficient Mt is obtained after the softmax layer; finally, the weight coefficient and the feature matrix are used to generate the temporal attention module weight coefficient Mt. Multiplying them will give you the scaled new feature matrix;

In the formula, GAP is the global spatial average pooling, Conv1D is the one-dimensional convolution, X is the input sample, is the sigmoid activation function, FC is the fully connected layer, Mt is the weight coefficient of the temporal attention module, /> is the new feature matrix; δ represents the attention weight, /> is the vector product.

The feature attention module includes a convolution layer, a Relu activation function, a sigmoid activation function and a pooling layer; in the feature attention module, the input feature is first passed through the pooling layer, and then through two convolution layers, the activation function is Relu, and then the obtained feature is fed back through a Sigmoid activation function to obtain the adaptive weight coefficient of the feature in different environments, and the weight coefficient is obtained. ; Finally, use the weight coefficient and the original feature matrix/> Multiplying them will give you the scaled new feature matrix;

In the formula, Conv is convolution, is the Relu activation function, /> is the new feature matrix.

The spatial attention module includes a convolution layer, a maximum pooling layer, an average pooling layer, and a sigmoid activation function. In the spatial attention module, firstly, average pooling and maximum pooling of a channel dimension are performed to obtain two H×W×1 channels and spliced together; then, a C×C convolution layer is passed, and the activation function is Sigmoid to obtain a weight coefficient Ms; finally, the weight coefficient and the feature matrix are used to obtain the weight coefficient Ms. Multiplying them will give you the scaled new feature matrix;

In the formula, is average pooling,/> For maximum pooling, is the new feature matrix.

The channel attention module includes a convolution layer, a maximum pooling layer, an average pooling layer, and a sigmoid activation function. In the channel attention module, firstly, global average pooling and maximum pooling are performed in space to obtain two 1×1×C channels, and then they are respectively sent to a two-layer shared neural network, the number of neurons in the first layer is C/r, the activation function is Relu, and the number of neurons in the second layer is C; then the two features are added and then passed through a Sigmoid activation function to obtain a weight coefficient Mc; finally, the weight coefficient and the original feature matrix are used to obtain the weight coefficient Mc. Multiplying them will give you the scaled new feature matrix;

In the formula, MLP is a multi-layer perceptron, is the new feature matrix.

In this embodiment, the feature matrix ,/> ,/> Input into the prediction model, consider the internal and external relationship of the features through the multi-dimensional attention mechanism in the prediction model, carry out offshore wind power feature enhancement from different dimensions with multiple attention deep fusion mechanism, explore the deep coupling relationship in time, space and features, and adaptively give corresponding weights to different features; based on time attention/> , Feature Attention/> , spatial attention/> , channel attention/> , carry out feature enhancement in depth;

In the formula, The weight matrix calculated for attention,/> is the neural network corresponding to attention,/> is the eigenvalue corresponding to the calculated dimension;

The obtained weights are multiplied by the characteristics of the corresponding offshore wind farms, and the weighted feature matrix is obtained: ,/> ,/> ;

.

In this embodiment, the three feature matrices of nearshore, shallow sea and deep sea are further mined according to the different degrees of influence of geographical location on the output of offshore wind turbines, and different weights are adaptively assigned using graph attention. Finally, a new feature matrix is obtained. ;Specific steps are as follows:

S61. Define the weight matrix W and convert the feature matrix into adjacent nodes: ,/> is the jth input sample (h=h1,…hn);

S62. Concatenate adjacent nodes i and j and map them into scalars, Calculate the attention function: /> , where / > The original attention contribution calculated for a node is used for normalization in the next step;

S63. Pass the adjacent node matrix through the leakyRelu layer and then the softmax layer, calculate the contribution of each neighbor node j of node i to i and normalize the contribution of each neighbor node j ;

S64. After calculating the contribution of each neighboring node of node i , the feature sum of all neighboring nodes of node i is updated according to the weight; as the final output of node i , ;

Let the number of output neurons be 3, corresponding to the weights of the three regions of nearshore, shallow sea and deep sea, and finally get the new feature matrix ;

in, ;

.

In this embodiment, the bidirectional gated recurrent unit network is constructed as follows: As input, a multi-layer cascade bidirectional gated recurrent unit network is built, which consists of a forward GRU and a backward GRU, and the activation function is tanh;

In the formula, 、/> 、/> 、/> 、/> 、/> is the weight parameter matrix, /> 、/> 、/> is the bias parameter matrix, /> is matrix multiplication, /> is the Sigmoid function, /> To reset the gate, /> To update the gate, /> is the candidate state of the hidden layer at the current moment, /> is the current implicit state, /> is the implicit state of the previous moment, /> is the input status at the current moment, /> Calculate for the reverse pass/> and/> is the hidden state of the forward GRU and the backward GRU, and F is the output merging method in both directions.

In this embodiment, the number of time scales is n, where n is an integer and n 2. Use data from four time scales, 15 minutes, one hour, one day, and three days, to predict the next four time scales, thereby achieving ultra-short-term and short-term predictions.

Assume that the time scale of the current forecasting system is i and the input is ,/> is the running data of the previous moment of time scale i ; the output of the prediction system is/> ,/> The prediction data for the current time scale; use the trained prediction model to predict from multiple time scales, predict the power of the target offshore wind farm, obtain the power time series of the corresponding offshore wind farm at different time scales, complete the ultra-short-term prediction and short-term prediction of offshore wind power respectively, and realize the offshore wind power prediction at multiple time and space scales;

In the formula, is the predicted output load result at the i -th scale,/> is the input matrix for the time scale from i -n-1 to i -1.

To further verify the effectiveness of the offshore wind power interval prediction method in Example 1 of the present invention, in this embodiment:

In step S1, the power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus and ocean current data of an offshore wind farm in a certain area are obtained;

In step S6, to ensure that the variance of the output is always positive, the number of neurons in the input layer is 64 and the number of neurons in the output layer is 3;

The above data is substituted into the offshore wind power interval prediction method in Example 1 for prediction, with 64 training batches and 256 training times.

Finally, the offshore wind power interval prediction effect is obtained as shown in Figure 4. From the prediction effect in Figure 4, it can be seen that the offshore wind power interval prediction method of the present invention can effectively improve the accuracy of offshore wind power interval prediction.

Example 3

This embodiment provides a multi-temporal and spatial scale offshore wind power feature screening and enhanced power prediction system, as shown in FIG5 , including:

A data acquisition unit, used to obtain multi-source data sets of the target and adjacent offshore wind farms, including power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cumulus cloud cover, and ocean current data of the target offshore wind farm;

A preprocessing unit is used to preprocess the feature sequence obtained by the data acquisition unit to obtain a feature vector, and use the offshore wind power time series as a prediction target sample, and divide the feature vector and the offshore wind power time series into a training set and a verification set respectively;

The multi-spatial-scale feature screening unit is used to perform motif spectral clustering based on the ocean current data acquired by the data acquisition unit, use Laplace eigenmap to screen the best sensitive factors, reduce the dimension of the data, and divide the feature data into three areas: nearshore, shallow sea, and deep sea, forming three feature matrices respectively;

The feature enhancement unit considers the internal and external relationships of features through the multi-dimensional attention mechanism in the prediction model, and performs offshore wind power feature enhancement with a multi-attention deep fusion mechanism from different dimensions, digs into the deep coupling relationship in time, space, and features, and adaptively gives corresponding weights to different features; based on time attention, feature attention, space attention, and channel attention, it deeply performs feature enhancement, multiplies the attention weight with the input feature map, obtains the weighted feature matrix, and sends it to the multi-time scale prediction unit;

The multi-time scale prediction unit is used to input the weighted feature matrix into the bidirectional gated recurrent unit network, use the trained prediction model to make predictions from multiple time scales, predict the power of the target offshore wind farm, obtain the power time series of the corresponding offshore wind farm at different time scales, and complete the offshore wind power prediction at multiple time and space scales.

In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims (10)

1. The multi-time space-scale offshore wind power characteristic screening and enhanced power prediction method is characterized by comprising the following steps of:

S1, acquiring a multi-source data set of a target and an adjacent offshore wind farm, wherein the multi-source data set comprises power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cloud storage and ocean current data of the target offshore wind farm, and performing primary processing;

S2, dividing the preprocessed historical meteorological record data of the offshore wind farm to obtain a training sample and a testing sample;

S3, carrying out motif spectrum clustering according to the acquired ocean current data, utilizing Laplace feature mapping to develop optimal sensitivity factor screening, carrying out dimension reduction on the data, and dividing the feature data into three areas of offshore, shallow and deep sea; the processed power, wind speed, wind direction, temperature, precipitation, humidity, air pressure, cloud storage quantity and ocean current data of the offshore wind farm form a feature matrix ,/>,;

S4, constructing a multidimensional attention fusion mechanism and a prediction model of a bidirectional gating circulation unit network;

S5, inputting a feature matrix ,/>,/> into a prediction model, taking the internal and external relations of the features into consideration through a multidimensional attention mechanism in the prediction model, developing the offshore wind power feature enhancement of a multi-attention depth fusion mechanism from different dimensions, excavating deep coupling relations on time, space and features, and giving corresponding weights to different features in a self-adaptive manner;

s6, further excavating implicit correlations of different areas according to different degrees of influence of geographical positions on the offshore wind turbine output on three feature matrixes of offshore, shallow sea and deep sea, giving different weights by using self-adaption of the schematic injection force, and finally obtaining a new feature matrix ;

S7, transmitting the feature matrix with the given weight and the enhanced features to a bidirectional gating circulation unit network, and mining hidden relations in the feature matrix by the bidirectional gating circulation unit network;

And S8, predicting the power of the target offshore wind power plant from a plurality of time scales by using the trained prediction model, and obtaining a power time sequence of the corresponding offshore wind power plant under different time scales so as to realize offshore wind power prediction under a plurality of time scales.

2. The multi-time space-scale offshore wind power feature screening and enhanced power prediction method according to claim 1, wherein in step S1, the specific process of performing preliminary processing on data is as follows: correcting a global forecasting system and a weather forecasting mode system according to actual measurement meteorological data near an offshore wind farm area to obtain NWP data with the horizontal resolution of Nkm multiplied by Nkm gridding; filling up the missing data in the historical meteorological record data of the offshore wind farm to obtain complete historical meteorological record data of the wind farm; carrying out min-max normalization processing on the power sequence, the wind speed sequence, the temperature sequence, the air pressure sequence, the cloud amount sequence and the ocean current sequence to obtain a processed power sequence P, a processed wind speed sequence WS, a processed temperature sequence T, a processed humidity sequence H, a processed precipitation amount sequence PRECIP, a processed air pressure sequence PA, a processed cloud amount sequence CU and a processed ocean current sequence OC, wherein the wind direction sequence adopts sine and cosine processing to obtain a wind direction sine WDS and a wind direction cosine WDC; in step S2, the preprocessed historical meteorological record data of the offshore wind farm is divided according to 8:2 and obtaining training samples and test samples.

3. The multi-time-space scale offshore wind power feature screening and enhanced power prediction method according to claim 1, wherein in step S3, motif spectral clustering is used for the processed sample features, laplace feature mapping is utilized to develop optimal sensitivity factor screening, and the data is subjected to dimension reduction, which comprises the following specific steps:

S31, constructing a similarity matrix , and calculating the similarity between samples according to the distance between the data samples;

s32, constructing an adjacent matrix , and constructing by using a Gaussian distance method:

Wherein 、/> is the i-th and j-th points (i, j=1,., n) in the n samples, s is the standard deviation, and e is an exponential function;

S33, calculating an order moment D and a Laplacian matrix L:

wherein is an order moment matrix of i rows and j columns, and/() is a Laplacian matrix of i rows and j columns;

S34, carrying out feature decomposition on the Laplace matrix to obtain feature vectors and feature values; selecting the first k eigenvectors as new data representations according to the magnitude of the eigenvalues; generating a feature matrix under a low dimension;

S35, carrying out K-Means clustering in the new feature space after dimension reduction, and dividing the new feature space into offshore, shallow and deep sea types to obtain a feature matrix ,/>,/>;

＝［M₁,M₂,...,M_n］,/> Wherein M _n represents a matrix of features at times t-1 to t-M of the nth wind farm.

4. The multi-time-space scale offshore wind power feature screening and enhanced power prediction method according to claim 3, wherein in step S4, the specific steps of constructing a multi-dimensional attention fusion mechanism are as follows:

The time attention module and the feature attention module are connected in series, input into a matrix of H multiplied by W multiplied by C, the feature diagram obtained after the processing of the two attention sub-modules is subjected to element addition operation, so that the two features are fused together, and output results are processed through a ReLU activation function and used as the input of the channel attention module and the space attention module, and the two modules are connected together by adopting a serial structure; in both attention modules, different feature information is acquired from different perspectives using both the average pooling and the maximum pooling operations.

5. The multi-time-space scale offshore wind feature screening and enhanced power prediction method of claim 4, wherein the time attention module comprises a convolution layer, an FC layer, a softmax layer, a sigmoid activation function and a global average pooling layer; in the time attention module, global spatial average pooling GAP is applied to the feature matrix to ensure that the time attention module has low computational cost; then generating a position-sensitive importance map using a plurality of 1D convolutions with non-linearities over the entire time domain to enhance the frame-by-frame feature; generating a channel self-adaptive kernel based on global time information in each channel through the FC layer, and obtaining a time attention module weight coefficient Mt through the softmax layer; finally, multiplying the characteristic matrix by a weight coefficient to obtain a new zoomed characteristic matrix;

In the formula, GAP is global space average pooling, conv1D is one-dimensional convolution, X is an input sample, is a sigmoid activation function, FC is a full-connection layer, mt is a time attention module weight coefficient, and/() is a new feature matrix; delta represents the attention weight, is the vector product;

The feature attention module comprises a convolution layer, relu activation functions, sigmoid activation functions and a pooling layer; in the feature attention module, features are input first, pass through a pooling layer, then pass through two convolution layers, an activation function is Relu, and then the obtained features are fed back to self-adaptive weight coefficients of the features in different environments through a Sigmoid activation function to obtain weight coefficients ; finally, multiplying the original characteristic matrix/> by a weight coefficient to obtain a new zoomed characteristic matrix;

wherein Conv is convolution, is Relu activation function, and/> is new feature matrix;

The spatial attention module comprises a convolution layer, a maximum pooling layer, an average pooling layer and a sigmoid activation function; in the space attention module, firstly, carrying out average pooling and maximum pooling of one channel dimension respectively to obtain two H multiplied by W multiplied by 1 channels, and splicing the channels together; then, through a C×C convolution layer, the activation function is Sigmoid, and a weight coefficient is obtained; finally, multiplying the characteristic matrix/> by a weight coefficient to obtain a new zoomed characteristic matrix;

wherein is average pooling,/> is maximum pooling, and/> is a new feature matrix;

The channel attention module comprises a convolution layer, a maximum pooling layer, an average pooling layer and a sigmoid activation function; in the channel attention module, firstly, carrying out global average pooling and maximum pooling on space to obtain two channels of 1 multiplied by C, respectively sending the channels into a two-layer shared neural network, wherein the number of neurons in the first layer is C/r, the activation function is Relu, and the number of neurons in the second layer is C; adding the obtained two features, and then obtaining a weight coefficient Mc through a Sigmoid activation function; finally, multiplying the original feature matrix by the weight coefficient to obtain a new zoomed feature matrix;

Wherein, MLP is a multi-layer perceptron and is a new feature matrix.

6. The multi-time space-scale offshore wind power feature screening and enhancing power prediction method according to claim 5, wherein in step S5, feature matrix ,/>,/> is input into a prediction model, the internal and external relationships of features are considered through a multi-dimensional attention mechanism in the prediction model, the offshore wind power feature enhancement of a multi-attention depth fusion mechanism is carried out from different dimensions, the deep coupling relationship in time, space and features is mined, and the corresponding weights are given to the different features in a self-adaptive manner; based on time attention/> , feature attention/> , spatial attention/> , channel attention/> , deep-development feature enhancement;

Wherein is a weight matrix calculated by attention, wherein/() is a neural network corresponding to the attention, and/() is a characteristic value corresponding to the calculated dimension;

Multiplying the obtained weights with the characteristics of the corresponding offshore wind farm respectively to obtain a weighted characteristic matrix ,/>,;

。

7. The multi-time-space scale offshore wind power feature screening and enhanced power prediction method according to claim 6, wherein step S6 specifically comprises:

S61, defining a weight matrix W, and converting the feature matrix into adjacent nodes: ,/> For the j-th input sample (h=h1, … hn);

S62, splicing and mapping adjacent nodes i and j into scalar quantities, wherein is an attention calculating function: in the formula,/> , wherein/> is the original attention contribution degree obtained by calculating a certain node and is used for carrying out the normalization of the next step;

s63, passing the adjacent node matrix through leakyRelu layers and then through a softmax layer, calculating the contribution degree of each adjacent node j of the node i to the i and normalizing the contribution degree of each adjacent node j;

s64, after the contribution degree of each adjacent node of the i node is calculated, carrying out feature summation update on all the adjacent nodes of the i node according to the weight; as the final output of the i-node, is obtained;

the number of the output neurons is 3, the weights of the offshore, shallow and deep sea areas are respectively corresponding to the output neurons, and finally a new feature matrix is obtained;

Wherein ;

。

8. The multi-time-space scale offshore wind power feature screening and enhanced power prediction method according to claim 7, wherein in step S7, the bidirectional gating cyclic unit network is constructed as follows: taking a feature matrix as input, constructing a multi-layer cascade bidirectional gating circulation unit network, wherein the multi-layer cascade bidirectional gating circulation unit network consists of a forward GRU and a backward GRU, and the activation function is tanh;

Wherein 、/>、/>、/>、/>、/> is a weight parameter matrix,/> 、/>、/> is a bias parameter matrix,/> is a matrix multiplication,/> is a Sigmoid function,/> is a reset gate,/> is an update gate,/> is a candidate state of an hidden layer at the current moment, is a current hidden state,/> is a hidden state at the previous moment,/> is an input state at the current moment,/> is a reverse transfer calculation/> and/> are hidden states of a forward GRU and a backward GRU, and F is an output merging method of two directions.

9. The multi-time-space scale offshore wind power feature screening and enhancement power prediction method of claim 8, wherein in step S8, the time scale is n, where n is an integer and n 2;

setting the time scale of the current prediction system as i, and inputting the running data of the previous moment with ,/> as the time scale i; the output of the prediction system is/> ,/> which is the prediction data of the current time scale; predicting power of a target offshore wind farm from a plurality of time scales by using a trained prediction model, obtaining power time sequences of the corresponding offshore wind farms under different time scales, respectively completing ultra-short-term prediction and short-term prediction of the offshore wind farm, and realizing offshore wind power prediction under the plurality of time scales;

Where is the predicted output load result on the i-th scale and/() is the input matrix for the i-n-1 to i-1 time scale.

10. The multi-time-space scale offshore wind power feature screening and enhancement power prediction method according to claim 9, wherein data of 15 minutes, one hour, one day, three days and four time scales are used for predicting four time scales in the future respectively to realize ultra-short-term and short-term prediction.