A wind farm power prediction method based on big data

Abstract

The present invention belongs to the technical field of wind power generation, and specifically relates to a method for predicting wind farm power generation based on big data; the method comprises: acquiring wind farm data and preprocessing the data to obtain preprocessed wind farm data; wherein the wind farm data comprises historical meteorological data, future meteorological data and power generation output time series data; embedding discrete features in the wind farm data to obtain embedded features; splicing the embedded features and continuous features in the wind farm data to obtain spliced ​​features; performing block processing on the spliced ​​features to obtain wind power features; using a trained wind farm power generation prediction model to process the wind power features to obtain a wind farm power generation prediction result; the prediction result of the present invention has high accuracy and good application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and in particular relates to a method for predicting wind farm power generation based on big data. Background Art

[0002] As the global demand for renewable energy continues to increase, wind power, as one of the important representatives of clean energy, is developing rapidly. However, due to the uncertainty and variability of natural factors such as wind speed, the prediction of wind farm power generation has always been a challenging problem. Traditional prediction methods based on physical models are often limited by the complexity of the model and the accuracy of the parameters, and it is difficult to fully consider the complex meteorological environment and the operating status of the wind farm. At the same time, with the rapid development of big data and deep learning technology, people realize that the use of massive data and deep learning models may bring new breakthroughs in wind farm power generation prediction.

[0003] The emergence of big data technology has provided strong support for wind farm power generation prediction. A large amount of data is generated during the operation of wind farms, including meteorological data such as wind speed, temperature, humidity, and operating data such as wind turbine status and power generation. These data are huge and diverse, with high temporal and spatial resolution, providing rich information for power generation prediction. Using big data technology, these data can be effectively collected, stored, processed and analyzed, and the potential laws and characteristics can be mined to provide a more reliable basis for power generation prediction.

[0004] On the other hand, the rapid development of deep learning technology has also brought new opportunities for wind farm power generation prediction. Deep learning models have powerful nonlinear fitting and representation learning capabilities, and can automatically learn the feature representation of data from big data and capture the complex relationships between data. Compared with traditional prediction methods based on physical models, deep learning models do not require too much prior knowledge, can better adapt to complex and changeable meteorological environments and wind farm operating conditions, and have higher prediction accuracy and generalization capabilities.

[0005] Therefore, the wind farm power prediction method based on big data and deep learning technology has important research value and broad application prospects. By making full use of big data and deep learning technology, it is possible to accurately predict the power generation of wind farms, provide a scientific basis for the operation and dispatch of wind farms, and promote the intelligent and efficient development of the wind power industry. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes a wind farm power prediction method based on big data, which includes:

[0007] S1: Acquire wind farm data and preprocess it to obtain preprocessed wind farm data; wherein the wind farm data includes historical meteorological data, future meteorological data and power generation output time series data;

[0008] S2: embedding the discrete features in the wind farm data to obtain embedded features; concatenating the embedded features with the continuous features in the wind farm data to obtain concatenated features;

[0009] S3: Block processing is performed on the spliced ​​features to obtain wind power features;

[0010] S4: Use the trained wind farm power generation prediction model to process wind power characteristics to obtain the wind farm power generation prediction result.

[0011] Preferably, in step S1, the historical meteorological data include historical wind speed, historical wind direction, historical temperature, historical air pressure, historical humidity, historical precipitation, historical cloud cover and historical solar radiation; the future meteorological data include forecast wind speed, forecast wind direction, forecast temperature, forecast air pressure, forecast humidity, forecast precipitation and forecast cloud cover; the power generation output timing data include power generation power, turbine status, load demand, turbine efficiency, fault record data, maintenance record data, grid dispatch instructions and turbine speed.

[0012] Preferably, in step S4, the process of the wind farm power generation prediction model processing wind power characteristics includes:

[0013] S41: using the SCTConv module to process the wind power characteristics and obtain an SCTConv module output feature map;

[0014] S42: Use the PTransFormer module to extract temporal features from the feature map output by the SCTConv module to obtain a comprehensive feature map;

[0015] S43: inputting the comprehensive feature map into the feature decoding module for processing to obtain a decoding feature;

[0016] S44: using the Enhance module to process the decoding feature to obtain an enhanced decoding feature;

[0017] S45: Input the enhanced decoding features into the DNN prediction module for processing to obtain the wind farm power generation prediction result.

[0018] Furthermore, the process of SCTConv module processing wind power characteristics includes:

[0019] Performing group normalization and weighting processing on wind power characteristics to obtain a first characteristic map;

[0020] Dividing the first feature map into an information part and a non-information part according to a preset gated threshold;

[0021] Reconstruct the information part and the non-information part to obtain two reconstructed information parts and two reconstructed non-information parts;

[0022] The two reconstructed information parts are added and passed through a linear layer to obtain a new reconstructed information part; the two reconstructed non-information parts are added and passed through a linear layer to obtain a new reconstructed non-information part;

[0023] The two newly reconstructed information parts are concatenated and added to the result of the concatenation of the two newly reconstructed non-information parts; the added feature map is processed through a linear layer to obtain a second feature map;

[0024] Perform global average pooling and convolution operations on the second feature map to obtain a channel attention weight vector; process the second feature map according to the channel attention weight vector to obtain an SCTConv module output feature map.

[0025] Furthermore, the formula for reconstructing the information part and the non-information part is:

[0026]

[0027]

[0028] Among them, x 11 and x 12 Represent the first and second reconstruction information parts respectively, x 21 and x 22 represents the first and second reconstructed non-information parts, x1 represents the information part, x2 represents the non-information part, size(x) represents the number of channels of x, and dim=1 represents the first dimension.

[0029] Furthermore, the process of extracting temporal features from the feature map output by the SCTConv module by the PTransFormer module includes:

[0030] Calculate the Query value, Key value and Value value based on the feature map output by the SCTConv module; Calculate the attention score matrix based on the Query value and Key value;

[0031] Perform sparse processing on the attention score matrix to obtain a sparse attention score matrix;

[0032] The comprehensive feature map is calculated based on the output feature map, Value value, Query value and sparse attention score matrix of the SCTConv module.

[0033] Furthermore, the formula for calculating the attention score matrix is:

[0034]

[0035]

[0036]

[0037] Among them, AttScore represents the attention score matrix, x c represents the output feature map of the SCTConv module, W K Represents the weight of the queried vector, W Q represents the query vector weight, d represents the data dimension, Linear() represents the linear layer, Q represents the Query value, and K represents the Key value.

[0038] Furthermore, the formula for calculating the comprehensive feature map is:

[0039]

[0040] Among them, x d represents the comprehensive feature map, x c Represents the output feature map of the SCTConv module, SparseAttScore represents the sparse attention score matrix, V represents the Value value, AttScore represents the attention score matrix, Q represents the Query value, and ReLu() represents the ReLu activation function.

[0041] Furthermore, the feature decoding module processes the comprehensive feature map including:

[0042] Perform a convolution operation on the comprehensive feature map to obtain the first spatial feature, and perform two convolution operations on the comprehensive feature map to obtain the second spatial feature;

[0043] The first spatial feature and the second spatial feature are convolved and then spliced, and the average value and the range value of the splicing result are calculated respectively;

[0044] Calculate the attention weight according to the average and range of the splicing results;

[0045] The attention weight is multiplied by the first spatial feature and the second spatial feature respectively, and the multiplication result is subjected to a convolution operation to obtain the decoding feature.

[0046] Furthermore, the formula for processing the decoding features by the Enhance module is:

[0047] x f =ReLu(Conv(ReLu(Conv(x e ))))

[0048] z1=MaxPool(xf , 2)

[0049] z2=MaxPool(x f , 4)

[0050] z1′=UpSample(ReLu(Conv(z1)))

[0051] z2′=UpSample(ReLu(Conv(z2)))

[0052] x f ′=Concat(z1′,z2′,x f )

[0053] x g =LeakyReLu(ReLu(Conv(x f ′))+x e )

[0054] Among them, x e Represents the decoding feature, Conv() represents the convolution operation, MaxPool() represents the maximum pooling operation, ReLu() represents the ReLu activation function, UpSample() represents the upsampling operation, and Concat() represents the concatenation operation; x f , z1, z2, z1′, z2′ and x f ′ represents the first, second, third, fourth, fifth and sixth intermediate parameters respectively; LeakyReLu() represents the LeakyReLu activation function, x g Represents enhanced decoding features.

[0055] The beneficial effects of the present invention are as follows: the present invention performs feature extraction through the SCTConv module, and combines the PTransFormer module to extract features from time series data, effectively capturing the spatiotemporal information and time series relationship in the data, and providing strong support for subsequent prediction tasks; the present invention adopts the SpatialAttentionDecoder module to decode the data, and further enhances the features through the Enhance module to enhance the model's ability to represent and understand the data; the present invention combines the attention mechanism and feature enhancement technology to effectively improve the prediction accuracy, while enhancing the generalization ability and computational efficiency of the model, solving the deficiencies of the prior art in capturing spatiotemporal information and model representation capabilities, and has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of the wind farm power generation prediction method based on big data in the present invention. DETAILED DESCRIPTION

[0057] 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. 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.

[0058] The present invention proposes a wind farm power generation prediction method based on big data, such as Figure 1 As shown, the method includes the following contents:

[0059] S1: Acquire wind farm data and preprocess it to obtain preprocessed wind farm data; wherein the wind farm data includes historical meteorological data, future meteorological data and power generation output time series data.

[0060] Obtain wind farm data, including historical meteorological data, future meteorological data and power generation output time series data. Preferably, the historical meteorological data is the meteorological data of each day in the past period of time, including historical wind speed (average wind speed or wind speed range), historical wind direction (wind direction or wind direction distribution), historical temperature (average temperature), historical air pressure (average air pressure), historical humidity average humidity, historical precipitation (rainfall or snowfall), historical cloud cover (cloud coverage) and historical solar radiation (solar radiation intensity or sunshine duration). Future meteorological data is the meteorological data of each day in the future period of time predicted by the weather forecast, including forecast wind speed, forecast wind direction, forecast temperature, forecast air pressure, forecast humidity, forecast precipitation and forecast cloud cover. The power generation output time series data is the power generation output time series data of each day in the past period of time, including power generation power (during training, the label is the real power generation power), turbine status (in operation or shut down for maintenance), load demand, turbine efficiency (turbine power generation efficiency), fault record data (whether a fault has occurred), maintenance record data (whether maintenance has been performed), power grid dispatching instructions (instructions issued by the power grid dispatching center, such as power adjustment, start-stop control, etc.) and turbine speed.

[0061] The wind farm data is preprocessed to improve the data quality. Specifically, the WNMI method is used to fill the missing values ​​of the data. The calculation formula of the WNMI method is as follows:

[0062]

[0063] Among them, α is the prior parameter of the mean, preferably 0.3; μ0 is the mean of m values ​​near the missing value, and w i is the weight parameter of the i-th value, y i is the value of the i-th sample near the missing value.

[0064] S2: Embed the discrete features in the wind farm data to obtain embedded features; and concatenate the embedded features with the continuous features in the wind farm data to obtain concatenated features.

[0065] The discrete features in the wind farm data include historical wind direction, forecast wind direction, turbine status, fault record data, and maintenance record data, and the rest are continuous features;

[0066] The discrete features in the wind farm data are embedded, that is, mapped through Embedding to obtain embedded features:

[0067] h i =Embedding(x i ), for i∈C

[0068] Among them, C represents the index set of discrete categories, x i Represents the feature value of the i-th discrete category.

[0069] The embedded features and the continuous features in the wind farm data are concatenated to obtain the concatenated features.

[0070] S3: Block the spliced ​​features to obtain wind power features.

[0071] Overlapping Patch technology is used to divide the splicing features into blocks. The formula is:

[0072] chunks=[data[i:i+k]for i in range(0,n-k+1,ko)]

[0073] Among them, the data length is n, and it is to be divided into m blocks, each block length is k, and there is overlap between blocks, the overlap length is o, chunks is the result after chunking, by controlling k and o, the overlap between chunks can be achieved so that the model can better capture the continuity and changes of the data.

[0074] Overlapping Patch technology helps to extract richer feature information and reduce the dimension of data. Wind power features are obtained by dividing the spliced ​​features into blocks.

[0075] S4: Use the trained wind farm power generation prediction model to process wind power characteristics to obtain the wind farm power generation prediction result.

[0076] S41: Use the SCTConv module to process the wind power characteristics and obtain the SCTConv module output feature map.

[0077] The SCTConv module includes two modules: SR and CALayer. In the SR module, the wind power feature xa After group normalization, we get gn x , according to the weight ω γ About gn x Re-weighted processing is performed to obtain the first feature map:

[0078]

[0079] reweights = sigmoid(gn x ×ω γ )

[0080] Among them, reweights represents the first feature map, g n .weight represents the learnable weight of the nth block.

[0081] The first feature map is divided into an information part x1 and a non-information part x2 according to the preset gate threshold gate_threshold; the information part contains elements whose weight is greater than or equal to the gate threshold, and the non-information part contains elements whose weight is less than the gate threshold; the division formula is expressed as:

[0082] x1=reweights≥gate_threshold

[0083] x2=reweights<gate_threshold

[0084] The information part and the non-information part are reconstructed to obtain two reconstructed information parts and two reconstructed non-information parts, which are expressed as:

[0085]

[0086]

[0087] Among them, x 11 and x 12 Represent the first and second reconstruction information parts respectively, x 21 and x 22 represents the first and second reconstructed non-information parts, x1 represents the information part, x2 represents the non-information part, size(x) represents the number of channels of x, and dim=1 represents the first dimension.

[0088] The two reconstructed information parts are added and passed through the linear layer to obtain the new reconstructed information part; the two reconstructed non-information parts are added and passed through the linear layer to obtain the new reconstructed non-information part; it can be expressed as:

[0089] x′ 11 , x′ 12 =Linear(x 11 +x12 ), Linear(x 11 +x 12 )

[0090] x′ 21 , x′ 22 =Linear(x 21 +x 22 ), Linear(x 21 +x 22 )

[0091] Concatenate the two newly reconstructed information parts and add the result of concatenating the two newly reconstructed non-information parts:

[0092] x′=concatenate(x′ 11 , x′ 12 )+concatenate(x′ 21 , x′ 22 ), dim=1

[0093] The added feature map is processed through a linear layer to obtain the second feature map:

[0094] x b =ReLu(W·x′+b)

[0095] Among them, Linear is the linear layer, concatenate is the concatenation operation, W is the weight of the fully connected layer, and b is the bias term.

[0096] The main idea of ​​the CALayer model is to perform global average pooling on the input feature map, then generate a channel attention weight vector through a series of convolution operations, and finally apply the weight vector to the input feature map.

[0097] The input of the CALayer model is the second feature map x b , first perform global average pooling on the second feature map, and then perform convolution operation to obtain the channel attention weight vector; specifically: the number of channels of the CALayer model is C, first perform global average pooling on the input feature map to obtain y, whose dimension is 1×1×C:

[0098] y=AvgPool(x)

[0099] Then, a series of convolution operations are performed to generate the channel attention weight vector ca_weights, where the ReLu function ensures that the weight is non-negative, and the Sigmoid function ensures that the weight is in the range of [0, 1]:

[0100] ca_weights=Sigmoid(Conv(ReLu(Conv(y))))

[0101] The second feature map is processed according to the channel attention weight vector to obtain the SCTConv module output feature map x c :

[0102] x c =x b +(x b ⊙ca_weights)

[0103] S42: Use the PTransFormer module to extract temporal features from the feature map output by the SCTConv module to obtain a comprehensive feature map.

[0104] Calculate the Q value, K value, and V value based on the output feature map of the SCTConv module:

[0105]

[0106] Among them, W K Represents the weight of the queried vector, W Q represents the query vector weight, W V Represents the mapping vector weight, which are all learnable parameters; Q represents the Query value, K represents the Key value, and V represents the Value value.

[0107] Calculate the attention score matrix based on the Q value and K value:

[0108]

[0109] Among them, AttScore represents the attention score matrix and d represents the data dimension.

[0110] The attention score matrix is ​​sparsely processed to obtain a sparse attention score matrix:

[0111]

[0112] Among them, AttScore ij is the attention score of the jth block to the ith block in the original attention score matrix, SparseAttScore ij is the attention score after sparse regularization, and threshold is the threshold. This operation can filter out positions with low attention scores, thereby achieving sparse regularization.

[0113] The comprehensive feature map is calculated based on the SCTConv module output feature map, V value, Q value and sparse attention score matrix:

[0114]

[0115] Among them, x d Represents the comprehensive feature map, and ReLu() represents the ReLu activation function.

[0116] The weighted summation of values ​​is performed by sparse attention scores to improve computational efficiency and reduce overfitting, the expressiveness of the model is introduced through nonlinear transformation, the original input information is retained through residual connections, and the dimension mapping is performed through projection layers to maintain the consistency and stability of the model. These steps together promote the model's effective learning and understanding of the characteristics of the input data.

[0117] S43: Input the comprehensive feature map into the feature decoding module for processing to obtain decoding features.

[0118] The feature decoding module SpatialAttentionDecoder performs the comprehensive feature map x d The processing process includes: performing a convolution operation on the comprehensive feature map to obtain the first spatial feature attn1, and performing two convolution operations on the comprehensive feature map to obtain the second spatial feature attn2:

[0119] attn1=LeakyReLu(Conv(x)), attn2=Conv spatial (attn1)

[0120] Among them, Conv spatial The spatial attention mechanism is extracted for attn1, using a larger convolution kernel, larger padding, and larger dilation rate to capture a wider range of spatial information.

[0121] After convolution operation is performed on the first spatial feature and the second spatial feature, they are concatenated and the average value and range value of the concatenated results are calculated respectively:

[0122] attn1=conv1(attn1), attn2=Conu2(attn2)

[0123] attn = concat(attn1, attn2)

[0124] avg attn =mean(attn)

[0125] diff attn =max(attn)-min(attn)

[0126] Among them, avg attn Indicates the average value of the splicing results, diff attnIndicates the extreme value of the splicing result, max(attn) indicates the maximum value of the splicing result, min(attn) indicates the minimum value of the splicing result, and attn indicates the splicing result.

[0127] The attention weight is calculated according to the average value and the range value of the splicing result, that is, through a convolution operation Conu_squeeze, the average value and the range value are compressed, and then the attention weight agg is obtained through the sigmoid activation function:

[0128] agg=Conv squeeze ([avg attn , diff attn ]), sig = sigmoid (agg)

[0129] Multiply the attention weights with the first spatial feature and the second spatial feature respectively, and pass the multiplication result through a convolution operation to obtain the decoding feature:

[0130] attn′=attn1⊙sig[0]+attn2⊙sig[1]

[0131] x e =Conv(attn′)

[0132] S44: Use the Enhance module to process the decoding features to obtain enhanced decoding features.

[0133] The Enhance module further enhances the features, expressed as:

[0134] x f =ReLu(Conv(ReLu(Conv(x e ))))

[0135] z1=MaxPool(x f , 2)

[0136] z2=MaxPool(x f , 4)

[0137] z1′=UpSample(ReLu(Conv(z1)))

[0138] z2′=UpSample(ReLu(Conv(z2)))

[0139] x f ′=Concat(z1′,z2′,x f )

[0140] x g =LeakyReLu(ReLu(Conv(xf ′))+x e )

[0141] Among them, x e Represents the decoding feature, Conv() represents the convolution operation, MaxPool() represents the maximum pooling operation, ReLu() represents the ReLu activation function, UpSample() represents the upsampling operation, and Concat() represents the concatenation operation; x f , z1, z2, z1′, z2′ and x f ′ represents the first, second, third, fourth, fifth and sixth intermediate parameters respectively; LeakyReLu() represents the LeakyReLu activation function, x g Represents enhanced decoding features.

[0142] The enhancement module further enhances the features and then combines them with the original features x f The LeakyReLu activation function is used to introduce nonlinear mapping, and finally the enhanced features are combined with the original features through feature fusion and residual connection to improve the representation and generalization capabilities of the model while retaining more original information, thereby effectively improving the performance and robustness of the model.

[0143] S45: Input the enhanced decoding features into the DNN prediction module for processing to obtain the wind farm power generation prediction result.

[0144] The DNN is defined as follows:

[0145] DNN=Linear(ReLu(Conv(x)))

[0146] After the enhanced decoding features are processed by the DNN prediction module, the wind farm power generation prediction results are output.

[0147] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting wind farm power generation based on big data, characterized in that: include: S1: Acquire wind farm data and preprocess it to obtain preprocessed wind farm data; wherein the wind farm data includes historical meteorological data, future meteorological data and power generation output time series data; S2: Embed the discrete features in the wind farm data, i.e., map them through Embedding to obtain embedded features; concatenate the embedded features with the continuous features in the wind farm data to obtain concatenated features; S3: Use Overlapping Patch technology to block the splicing features to obtain wind power features; S4: Using the trained wind farm power prediction model to process wind power characteristics, and obtain wind farm power prediction results; the process of using the wind farm power prediction model to process wind power characteristics includes: S41: Using the SCTConv module to process the wind power characteristics, and obtaining the SCTConv module output feature map; the process of the SCTConv module processing the wind power characteristics includes: Performing group normalization and weighting processing on wind power characteristics to obtain a first characteristic map; Dividing the first feature map into an information part and a non-information part according to a preset gated threshold; Reconstruct the information part and the non-information part to obtain two reconstructed information parts and two reconstructed non-information parts; The two reconstructed information parts are added and passed through a linear layer to obtain a new reconstructed information part; the two reconstructed non-information parts are added and passed through a linear layer to obtain a new reconstructed non-information part; The two newly reconstructed information parts are concatenated and added to the result of the concatenation of the two newly reconstructed non-information parts; the added feature map is processed through a linear layer to obtain a second feature map; Perform global average pooling and convolution operations on the second feature map to obtain a channel attention weight vector; process the second feature map according to the channel attention weight vector to obtain an SCTConv module output feature map; S42: Using the PTransFormer module to extract the temporal features of the feature map output by the SCTConv module to obtain a comprehensive feature map; the process of the PTransFormer module extracting the temporal features of the feature map output by the SCTConv module includes: Calculate the Query value, Key value and Value value based on the feature map output by the SCTConv module; Calculate the attention score matrix based on the Query value and Key value; Perform sparse processing on the attention score matrix to obtain a sparse attention score matrix; The comprehensive feature map is calculated based on the output feature map, Value value, Query value and sparse attention score matrix of the SCTConv module; S43: Input the comprehensive feature map into the feature decoding module for processing to obtain decoding features; the feature decoding module processes the comprehensive feature map including: Perform a convolution operation on the comprehensive feature map to obtain the first spatial feature, and perform two convolution operations on the comprehensive feature map to obtain the second spatial feature; The first spatial feature and the second spatial feature are convolved and then spliced, and the average value and the range value of the splicing result are calculated respectively; Calculate the attention weight according to the average and range of the splicing results; Multiply the attention weight with the first spatial feature and the second spatial feature respectively, and perform a convolution operation on the multiplied result to obtain the decoding feature; S44: using the Enhance module to process the decoding feature to obtain an enhanced decoding feature; S45: Input the enhanced decoding features into the DNN prediction module for processing to obtain the wind farm power generation prediction result.

2. The method for predicting wind farm power generation based on big data according to claim 1, characterized in that: In step S1, the historical meteorological data include historical wind speed, historical wind direction, historical temperature, historical air pressure, historical humidity, historical precipitation, historical cloud cover and historical solar radiation; the future meteorological data include forecast wind speed, forecast wind direction, forecast temperature, forecast air pressure, forecast humidity, forecast precipitation and forecast cloud cover; the power generation output timing data include power generation power, turbine status, load demand, turbine efficiency, fault record data, maintenance record data, power grid dispatch instructions and turbine speed.

3. The method for predicting wind farm power generation based on big data according to claim 1, characterized in that: The formula for reconstructing the information part and the non-information part is: Among them, x 11 and x 12 Represent the first and second reconstruction information parts respectively, x 21 and x 22 represents the first and second reconstructed non-information parts, x1 represents the information part, x2 represents the non-information part, size(x) represents the number of channels of x, and dim=1 represents the first dimension.

4. The method for predicting wind farm power generation based on big data according to claim 1, characterized in that: The formula for calculating the attention score matrix is: Among them, AttScore represents the attention score matrix, x c represents the output feature map of the SCTConv module, W K Represents the weight of the queried vector, W Q represents the query vector weight, d represents the data dimension, Linear() represents the linear layer, Q represents the Query value, and K represents the Key value.

5. The method for predicting wind farm power generation based on big data according to claim 1, characterized in that: The formula for calculating the comprehensive feature map is: Among them, x d represents the comprehensive feature map, x c Represents the output feature map of the SCTConv module, SparseAttScore represents the sparse attention score matrix, V represents the Value value, AttScore represents the attention score matrix, Q represents the Query value, and ReLu() represents the ReLu activation function.

6. The method for predicting wind farm power generation based on big data according to claim 1, characterized in that: The formula for processing the decoding features by the Enhance module is: x f =ReLu(Conv(ReLu(Conv(x e )))) Z1=MaxPool(x f ,2) z2=MaxPool(x f ,4) z1′=UpSample(ReLu(Conv(z1))) z2′=UpSample(ReLu(Conv(z2))) x f ′=Concat(z1′,z2′,x f ) x g =LeakyReLu(ReLu(Conv(x f ′))+x e ) Among them, x e Represents the decoding feature, Conv() represents the convolution operation, MaxPool() represents the maximum pooling operation, ReLu() represents the ReLu activation function, UpSample() represents the upsampling operation, and Concat() represents the concatenation operation; x f , z1, z2, z1′, z2′ and x f ′ represents the first, second, third, fourth, fifth and sixth intermediate parameters respectively; LeakyRelu() represents the LeakyReLu activation function, x g Represents enhanced decoding features.

Images