Single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition

Abstract

The invention discloses a univariate ultra-short-term wind power prediction method based on two-stage trend decomposition, and the method comprises the steps: carrying out the multiple times of trend decomposition of historical wind power time series data, and generating a macroscopic trend component, a mesoscale trend component and a residual component; performing exponential distribution initialization causal convolution kernel extraction on the decomposed macroscopic trend component and the mesoscale trend component to extract multi-scale trend characteristics, and keeping time sequence causality by adopting a proportional normalized adaptive weight; residual modeling is enhanced by adopting a loop reconstruction attention mechanism, and residual component dynamic features are obtained through sequence splicing and double residual connection; and performing linear processing on each component and performing result fusion to generate an ultra-short-term wind power prediction value. And precise prediction of ultra-short-term wind power can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power prediction, and relates to a univariate ultra-short-term wind power prediction method based on two-stage trend decomposition. Background Art

[0002] The intermittent and volatile characteristics of wind power are likely to cause impact power disturbances to the main power grid, and even affect the operation safety and stability of the local power grid. Existing methods, such as ARIMA time series analysis and LSTM multivariate fusion models, still have problems such as relying on the collaborative modeling of multivariate meteorological data and low efficiency in extracting the time series characteristics of ultra-short-term prediction.

[0003] The existing patent with the application number CN2023101478202 discloses a short-term power generation prediction method based on ensemble learning. By using MSTL to fully extract the trend term, periodic term, and residual term of the original sequence, and predicting different components separately. The residual term decomposed by MSTL is further decomposed by using the VMD algorithm, and the components are predicted separately. Although this method considers the periodicity of the historical power data, it does not consider the time series causality, and there are certain limitations in the univariate method. Summary of the Invention

[0004] The purpose of the present invention is to provide a univariate ultra-short-term wind power prediction method based on two-stage trend decomposition, which performs two-stage trend decomposition on the historical wind power time series data, extracts multi-scale trend features, constructs a cyclic reconstruction attention mechanism to model the dynamic characteristics of the residual component, and fuses the prediction results of multiple components to achieve accurate prediction of ultra-short-term wind power.

[0005] The technical solution to achieve the purpose of the present invention is as follows: A univariate ultra-short-term wind power prediction method based on two-stage trend decomposition, comprising the following steps: S01: Perform multiple trend decompositions on the historical wind power time series data to generate a macro trend component, a meso-scale trend component, and a residual component; S02: Respectively use the exponential distribution to initialize the causal convolution kernel for the decomposed macro trend component and meso-scale trend component to extract multi-scale trend features, and use the proportion-normalized adaptive weight to maintain the time series causality; S03: Use the cyclic reconstruction attention mechanism to strengthen the residual modeling, and obtain the dynamic characteristics of the residual component through sequence splicing and double residual connection; S04: Perform linear processing on each component respectively and fuse the results to generate an ultra-short-term wind power prediction value.

[0006] In the preferred technical solution, before step S01, it further includes: The univariate historical power data set of the wind farm is split and processed in batches, including: Based on the Z-score standardization method, the mean is calculated based on the training data set μ and standard deviation σ , perform standardization on the entire dataset; The original time series data is divided into training set, validation set and test set, where a forward overlapping window is set at the boundary between the training set and the validation set, and the overlapping length is set to the input sequence length to ensure time series continuity; Convert the processed tensor data into a time series format suitable for the prediction model input ; in, represents the power data at each time step of the input sequence, L seq Represents the time step length of the input sequence, and the tensor structure of the input sequence is ; in, B Indicates the batch size, L pred represents the prediction step length, N Indicates the number of channels and is fixed to 1 to match the needs of univariate prediction.

[0007] In the preferred technical solution, step S01 specifically includes: The macro trend component is extracted using the first causal convolutional layer, which uses a forward one-sided zero padding strategy to maintain the sequence length: ;

[0008] in, T Macro Represents the macro trend component, DTD represents the two-level trend decomposition operation, Cat represents the sequence splicing operation, Padding represents the sequence zero padding operation, K Macro Represents the convolution kernel size used to extract the macro trend component, P is the historical wind power time series data; Obtain temporary intermediate state components : ;

[0009] The temporal intermediate state sequence is processed using a second causal convolutional layer to extract the mesoscale trend component: ;

[0010] in, T Meso represents the mesoscale trend component, KMeso Indicates the convolution kernel size for extracting the mesoscale trend component; Remove the mesoscale trend component from the temporary intermediate state sequence to obtain the residual component R: .

[0011] In the preferred technical solution, in order to achieve differential modeling of different time scales, the convolution kernel sizes of each causal convolution layer satisfy the following conditions: ;

[0012] Wherein, L seq Indicates the time step length of the input sequence.

[0013] In the preferred technical solution, the initialization strategy of the convolution weights in step S02 adopts exponential distribution initialization: ;

[0014] Wherein, λ is the decay rate parameter, K c is the convolution kernel size, i is the time step index where the weight is located, w i represents the convolution weight at the i th time step, and exp represents the exponential operation; Perform proportional normalization on the convolution weights along the sequence dimension: ;

[0015] Wherein, w ’ i represents the proportional normalization convolution weight at the i th time step.

[0016] In the preferred technical solution, step S03 specifically includes: Adopt a cyclic reconstruction attention mechanism to strengthen the residual modeling. Segment the residual component input sequence according to a specific cycle period, and cyclically shift each sequence segment to the end of the sequence in turn to generate a multi-version spliced sequence: ;

[0017] Wherein, R cat represents the spliced residual component sequence, Cat represents the sequence splicing operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H represents the cycle period; Generate the three elements of attention through linear transformation​​Q , K and V : ;

[0018] Among them: W Q , W K and W V represent weight matrices, b Q , b K and b V represent bias terms; Introduce a learnable scaling factor s , and use the Softmax function to perform scaled dot-product calculations on Q and K to obtain attention scores Scores Attention : ;

[0019] Among them, d k represents K the sequence dimension size of; Use the attention scores and V to perform weighted calculations to generate the attention output Output Attention : ;

[0020] Adopt two residual connections and a feed-forward neural network to enhance the training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization: ;

[0021] Among them, Residual 1st 、Residual 2nd respectively represent the outputs of the two residual connections, BatchNorm represents the batch normalization operation, LayerNorm represents the layer normalization operation, FFN represents the feed-forward neural network, Output FFN represents the output of the feed-forward neural network, represents the residual component after being enhanced by the cyclic reconstruction attention mechanism.

[0022] In the preferred technical solution, step S04 for generating the ultra-short-term wind power prediction value includes: Establish independent linear mapping layers for the macro trend component, mesoscale trend component, and residual component generated by the two-level trend decomposition respectively; Perform unbiased summation calculation using the linear mapping outputs of each component, and use the component mapping fusion result as the final prediction output: ;

[0023] Among them, T Macro represents the macroscopic trend component, T Meso represents the mesoscale trend component, represents the residual component after being strengthened and modeled by the cyclic reconstruction attention mechanism, W macro , W meso and W Res represent the mapping weight matrices applied to the three components respectively, P pred represents the prediction output.

[0024] The present invention also discloses a univariate ultra-short-term wind power prediction system based on two-stage trend decomposition, including: A two-stage trend decomposition module that performs multiple trend decompositions on historical wind power time series data to generate a macroscopic trend component, a mesoscale trend component, and a residual component; A trend modeling module that respectively uses an exponentially distributed initialization causal convolution kernel to extract multi-scale trend features for the decomposed macroscopic trend component and mesoscale trend component, and uses a proportionally normalized adaptive weight to maintain time series causality; A residual modeling module that strengthens residual modeling using a cyclic reconstruction attention mechanism, and obtains the dynamic features of the residual component through sequence splicing and double residual connection; A mapping prediction module that linearly processes each component and fuses the results to generate an ultra-short-term wind power prediction value.

[0025] The present invention also discloses a computer storage medium, on which a computer program is stored, and when the computer program is executed, it implements the above-mentioned univariate ultra-short-term wind power prediction method based on two-stage trend decomposition.

[0026] Compared with the prior art, the present invention has the following remarkable advantages: First, based on the cascaded causal convolution and two-stage trend decomposition technologies, the original power time series data is decoupled into a macroscopic trend component, a mesoscale trend component, and a residual component. Through the differential receptive field design of the two-stage convolution kernel, the spatio-temporal feature separation across time scales is realized, breaking through the limitations of traditional univariate methods for coupling and modeling multi-scale features.

[0027] Secondly, the proposed double weight constraint strategy of exponential distribution initialization and ratio normalization constructs a temporal causal barrier from the source of convolutional kernel parameter initialization, strengthens the dynamic perception ability of features at the current moment through the exponentially decaying weight distribution, and eliminates the parameter sensitivity in long-range dependence modeling by the ratio normalization operation along the time dimension, avoiding gradient vanishing or gradient explosion during the training process.

[0028] Finally, the component-independent mapping fusion mechanism is adopted to perform hierarchical feature decoupling and directional mapping fusion, construct an independently optimized representation space for features at different time scales, and use the component mapping fusion result as the final prediction output. It provides a solution with both mechanism interpretability and univariate data adaptability for univariate ultra-short-term wind power prediction. Description of the Drawings

[0029] Figure 1 is the flow chart of the univariate ultra-short-term wind power prediction method based on two-stage trend decomposition in this embodiment; Figure 2 is the comparison chart of the prediction result and the true value under the condition of DM (hidden layer dimension) = 16 and LR (learning rate) = 0.0001; Figure 3 is the comparison chart of the prediction result and the true value under the condition of DM = 32 and LR = 0.0001; Figure 4 is the comparison chart of the prediction result and the true value under the condition of DM = 64 and LR = 0.0001; Figure 5 is the comparison chart of the prediction result and the true value under the condition of DM = 32 and LR = 0.0005; Figure 6 is the comparison chart of the prediction result and the true value under the condition of DM = 32 and LR = 0.00005. Detailed Implementation Manner

[0030] The principle of the present invention is: perform multiple trend decompositions on historical wind power time series data to generate a macro trend component, a mesoscale trend component, and a residual component; extract multi-scale trend features by initializing causal convolutional kernels with an exponential distribution, and use ratio normalization of adaptive weights to maintain temporal causality; construct a recurrent reconstruction attention mechanism to model the dynamic characteristics of the residual component; adopt a component-independent mapping fusion mechanism to perform hierarchical feature decoupling and directional mapping fusion, construct an independently optimized representation space for features at different time scales, and use the component mapping fusion result as the final prediction output. The present invention provides a solution with both mechanism interpretability and univariate data adaptability for univariate ultra-short-term wind power prediction.

[0031] Example 1:

[0032] As Figure 1As shown in the figure, a univariate ultra-short-term wind power prediction method based on two-stage trend decomposition includes the following steps: S01: Perform multiple trend decompositions on the historical wind power time series data to generate a macro trend component, a mesoscale trend component, and a residual component; S02: Respectively use the exponential distribution to initialize the causal convolution kernel for the decomposed macro trend component and mesoscale trend component to extract multi-scale trend features, and use the proportion-normalized adaptive weight to maintain the time series causality; S03: Use the cyclic reconstruction attention mechanism to strengthen the residual modeling, and obtain the dynamic features of the residual component through sequence splicing and double residual connection; S04: Perform linear processing on each component respectively and perform result fusion to generate the ultra-short-term wind power prediction value.

[0033] Specifically, before step S01, it also includes: Segment the univariate historical power data set of the wind farm and process it in batches, including: Based on the Z-score standardization method, calculate the mean with the training data set as the benchmark μ and the standard deviation σ , and perform standardization processing on the full data set x; its calculation formula is: ;

[0034] Among them, xstd represents the standardized time series, and the time series data is divided into a training set, a validation set, and a test set according to the ratio of 7:1:2. The boundary between the training set and the validation set is set with a forward overlapping window, and the overlapping length is set to the input sequence length to ensure the time series continuity; Convert the processed tensor data into a time series adapted to the input format of the prediction model ; Among them, represents the power data of each time step of the input sequence, L seq represents the time step length of the input sequence, and the tensor structure of the input sequence is ; Among them, B represents the batch size, L pred represents the prediction step length, N represents the number of channels and is fixed at 1 to match the univariate prediction requirement.

[0035] In a preferred embodiment, step S01 specifically includes: Use the first causal convolution layer to extract the macro trend component, and this causal convolution layer adopts the forward unilateral zero-padding strategy to maintain the sequence length: ;

[0036] Among them, T Macro represents the macro trend component, DTD represents the two-level trend decomposition operation, Cat represents the sequence concatenation operation, and Padding represents the sequence zero-padding operation. K Macro represents the convolutional kernel size for extracting the macro trend component. P is the historical wind power time series data; Obtain the temporary intermediate state component : ;

[0037] Process the temporary intermediate state sequence using the second causal convolutional layer to extract the mesoscale trend component: ;

[0038] Among them, T Meso represents the mesoscale trend component, K Meso represents the convolutional kernel size for extracting the mesoscale trend component; Remove the mesoscale trend component from the temporary intermediate state sequence to obtain the residual component R : .

[0039] Preferably, in order to achieve differential modeling at different time scales, the convolutional kernel sizes of each causal convolutional layer satisfy the following conditions: ;

[0040] Among them, L seq represents the time step length of the input sequence.

[0041] Preferably, the initialization strategy of the convolutional weights in step S02 adopts exponential distribution initialization: ;

[0042] Among them, λ is the decay rate parameter, K c is the convolutional kernel size, i is the time step index where the weight is located, w i represents the convolutional weight at the i th time step, and exp represents the exponential operation; Perform proportional normalization on the convolutional weights along the sequence dimension: ;​

[0043] Among them, w ’ i represents the proportion-normalized convolutional weight at the i th time step.

[0044] In a preferred embodiment, step S03 specifically includes: Adopt a cyclic reconstruction attention mechanism to strengthen residual modeling. Segment the input sequence of residual components according to a specific cycle period, and cyclically shift each sequence segment to the end of the sequence in turn to generate a multi-version concatenated sequence: ;

[0045] Among them, R cat represents the concatenated residual component sequence, Cat represents the sequence concatenation operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H represents the cycle period; Generate the three elements of attention through linear transformation Q , K and V : ;

[0046] Among them: W Q , W K and W V represent weight matrices, b Q , b K and b V represent bias terms; Introduce a learnable scaling factor s , and use the Softmax function to perform a scaled dot product calculation on Q and K to obtain the attention score Scores Attention : ;

[0047] Among them, d k represents K 's sequence dimension size; Use the attention score to perform a weighted calculation with V to generate the attention output Output Attention : ;​

[0048] The training stability is enhanced by using two residual connections and a feed - forward neural network. The first residual connection uses post - batch normalization, and the second residual connection uses post - layer normalization: ;

[0049] Among them, Residual 1st and Residual 2nd respectively represent the outputs of the two residual connections, BatchNorm represents the batch normalization operation, LayerNorm represents the layer normalization operation, FFN represents the feed - forward neural network, and Output FFN represents the output of the feed - forward neural network, represents the residual component after being enhanced by the cyclic reconstruction attention mechanism.

[0050] Preferably, step S04 for generating the ultra - short - term wind power prediction value includes: Establishing independent linear mapping layers for the macro - trend component, meso - scale trend component, and residual component generated by the two - level trend decomposition respectively; Performing unbiased summation calculation using the linear mapping outputs of each component, and taking the component mapping fusion result as the final prediction output: ;

[0051] Among them, T Macro represents the macro - trend component, T Meso represents the meso - scale trend component, represents the residual component after being enhanced by the cyclic reconstruction attention mechanism, W macro , W meso and W Res represent the mapping weight matrices applied to the three components respectively, P pred represents the prediction output.

[0052] In another embodiment, a computer storage medium stores a computer program, and when the computer program is executed, it implements the above - mentioned univariate ultra - short - term wind power prediction method based on two - level trend decomposition. The specific implementation adopts the above - mentioned wind power prediction method and will not be elaborated here.

[0053] In another embodiment, a univariate ultra - short - term wind power prediction system based on two - level trend decomposition includes: A two - level trend decomposition module performs multiple trend decompositions on historical wind power time - series data to generate a macro - trend component, a meso - scale trend component, and a residual component; A trend modeling module respectively uses an exponential distribution to initialize causal convolution kernels for the decomposed macro - trend component and meso - scale trend component to extract multi - scale trend features, and uses proportion - normalized adaptive weights to maintain time - series causality; A residual modeling module uses a cyclic reconstruction attention mechanism to strengthen residual modeling, and through sequence splicing and double residual connections, obtains the dynamic features of the residual component; A mapping prediction module performs linear processing on each component respectively and fuses the results to generate a very - short - term wind power prediction value.

[0054] The specific implementation uses the above - mentioned wind power prediction method, which will not be elaborated here.

[0055] To effectively train the prediction model, appropriate parameter design and experimental design are required. The specific methods are as follows: Select the wind power dataset of a certain wind farm in East China for very - short - term prediction. The time resolution of the wind power sequence is 15 minutes. This dataset is divided into a training set, a validation set, and a test set according to the ratio of 7:1:2.

[0056] The parameter design of the prediction model is as follows: prediction step length PL = 2, batch size BS = 1024, learning rate LR = [0.00005, 0.0001, 0.0005], dropout rate DP = 0.4, hidden layer dimension DM = [16, 32, 64], attention layer number NL = 4.

[0057] During training, the mean squared error (MSE) is used as the guiding loss function for backpropagation, and at the same time, the mean absolute error (MAE) is provided as another perspective reference for loss measurement to comprehensively evaluate the model performance.

[0058] When DM = 16 and LR = 0.0001, the comparison chart of the power prediction result and the true value is as Figure 2 shown.

[0059] When DM = 32 and LR = 0.0001, the comparison chart of the power prediction result and the true value is as Figure 3 shown.

[0060] When DM = 64 and LR = 0.0001, the comparison chart of the power prediction result and the true value is as Figure 4 shown.

[0061] When DM = 32 and LR = 0.0005, the comparison chart of the power prediction result and the true value is as Figure 5 shown.

[0062] When DM = 32 and LR = 0.00005, the comparison diagram of the power prediction result and the true value is as Figure 6 shown.

[0063] Table 1 gives the experimental results of each group of parameters in this embodiment.

[0064] Table 1: Experimental parameters and results of each group

[0065] It can be seen from Table 1 that the embodiment of the present invention realizes the prediction of a relatively low error level of univariate wind power in the 2-step ultra-short-term power prediction experiment of the above wind farm dataset, showing the superiority of the present invention in the univariate ultra-short-term wind power prediction.

[0066] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A univariate ultra-short-term wind power prediction method based on two-stage trend decomposition, characterized in that It includes the following steps: S01: Perform multiple trend decompositions on the historical wind power time series data to generate a macro trend component, a mesoscale trend component, and a residual component; S02: Respectively use the exponential distribution to initialize the causal convolution kernel for the decomposed macro trend component and mesoscale trend component to extract multi-scale trend features, and use the proportion-normalized adaptive weight to maintain the time series causality; S03: Adopt a cyclic reconstruction attention mechanism to strengthen the residual modeling. Through sequence splicing and double residual connection, obtain the dynamic features of the residual component; S04: Perform linear processing on each component respectively and fuse the results to generate the ultra-short-term wind power prediction value.

2. The single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition according to claim 1, wherein, Before step S01, it also includes: Segment the univariate historical power dataset of the wind farm and process it in batches, including: The Z-score based standardization method calculates the mean based on the training dataset μ and the standard deviation σ and performs standardization processing on the full dataset; Divide the original time series data into a training set, a validation set, and a test set. Set a forward overlapping window at the boundary between the training set and the validation set, and set the overlapping length to the input sequence length to ensure the time series continuity; Convert the processed tensor data into a time series adapted to the input format of the prediction model ； Among them, represents the power data of each time step of the input sequence, L seq represents the time step length of the input sequence, and the tensor structure of the input sequence is ; Among them, B represents the batch size, L pred represents the prediction step, N represents the number of channels and is fixed at 1 to match the univariate prediction requirement.

3. The single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition according to claim 1, characterized in that Step S01 specifically includes: Use the first causal convolution layer to extract the macro trend component. This causal convolution layer adopts a forward unilateral zero-padding strategy to maintain the sequence length: ， Among them, T Macro represents the macro trend component, DTD represents the two-level trend decomposition operation, Cat represents the sequence concatenation operation, and Padding represents the sequence zero-padding operation. K Macro represents the convolutional kernel size used to extract the macro trend component. P is the historical wind power time series data; Obtain a temporary intermediate component : ， Use the second causal convolution layer to process the temporary intermediate state sequence to extract the mesoscale trend component: ， Among them, T Meso represents the mesoscale trend component, K Meso represents the convolution kernel size used to extract the mesoscale trend component; Remove the mesoscale trend component from the temporary intermediate state sequence to obtain the residual component R : 。 4. The single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition according to claim 3, wherein In order to achieve differential modeling of different time scales, the convolution kernel sizes of each causal convolution layer satisfy the following conditions: ， Among them, L seq represents the time step length of the input sequence.

5. The single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition according to claim 3, characterized in that In step S02, the initialization strategy of the convolution weight adopts the exponential distribution for initialization: ， Among them, λ is the attenuation rate parameter, K c is the convolution kernel size, i is the time step index where the weight is located, w i represents the convolution weight at the i th time step, and exp represents the exponential operation;​ Perform proportion normalization on the convolution weight along the sequence dimension: ， Among them, w ’ i represents the ratio-normalized convolutional weights at the i -th time step.​ 6. The single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition according to claim 1, wherein Step S03 specifically includes: Adopt a cyclic reconstruction attention mechanism to strengthen the residual modeling. Segment the input sequence of the residual component according to a specific cycle period, and cyclically shift each sequence segment to the end of the sequence in turn to generate a multi-version spliced sequence: ， Among them, R cat represents the concatenated residual component sequence, Cat represents the sequence concatenation operation, R is the residual component, Shift represents the circular shift operation of the sequence segment, H represents the cycle period; Generating the three key elements of attention through linear transformation Q , K and V : ， Wherein: W Q , W K and W V represent weight matrices, b Q , b K and b V represent bias quantities; Introduce a learnable scaling factor s , and use the Softmax function to Q and K perform scaled dot-product calculations to obtain attention scores Scores Attention : ， Among them, d k represents K the sequence dimension size of Using the attention scores and V performing weighted calculations to generate the attention output Output Attention : ， Adopt two residual connections and a feed-forward neural network to enhance the training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization: ， Among them, Residual 1st and Residual 2nd respectively represent the outputs of two residual connections. BatchNorm represents the batch normalization operation, LayerNorm represents the layer normalization operation, FFN represents the feed-forward neural network, and Output FFN represents the output of the feed-forward neural network, represents the residual component after being enhanced and modeled by the cyclic reconstruction attention mechanism.

7. The single-variable ultra-short-term wind power prediction method based on two-stage trend decomposition according to claim 1, characterized in that Step S04 to generate the ultra-short-term wind power prediction value includes: Respectively establish independent linear mapping layers for the macro trend component, mesoscale trend component, and residual component generated by the two-level trend decomposition; Use the linear mapping outputs of each component for unbiased summation calculation, and use the component mapping fusion result as the final prediction output: ， Among them, T Macro represents the macro trend component, T Meso represents the mesoscale trend component, represents the residual component after being enhanced and modeled by the cyclic reconstruction attention mechanism, W macro , W meso and W Res represent the mapping weight matrices applied to the three components respectively, P pred represents the predicted output.

8. A univariate ultra-short-term wind power prediction system based on two-stage trend decomposition, characterized in that It includes: A two-level trend decomposition module that performs multiple trend decompositions on the historical wind power time series data to generate a macro trend component, a mesoscale trend component, and a residual component; A trend modeling module that respectively uses the exponential distribution to initialize the causal convolution kernel for the decomposed macro trend component and mesoscale trend component to extract multi-scale trend features, and uses the proportion-normalized adaptive weight to maintain the time series causality; A residual modeling module that adopts a cyclic reconstruction attention mechanism to strengthen the residual modeling. Through sequence splicing and double residual connection, obtain the dynamic features of the residual component; A mapping prediction module that performs linear processing on each component respectively and fuses the results to generate the ultra-short-term wind power prediction value.

9. The single-variable ultra-short-term wind power prediction system based on two-stage trend decomposition according to claim 8, characterized in that The specific method of the residual modeling module includes: The cyclic reconstruction attention mechanism is adopted to strengthen the residual modeling. The residual component input sequence is segmented according to a specific cyclic period, and each sequence segment is cyclically shifted to the end of the sequence in turn to generate a multi-version spliced sequence: ， Among them, R cat represents the concatenated residual component sequence, Cat represents the sequence concatenation operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H represents the cycle period; Generate the three key elements of attention through linear transformation Q , K and V : ， Wherein: W Q , W K and W V represent weight matrices, b Q , b K and b V represent bias quantities; Introduce a learnable scaling factor s , and use the Softmax function to Q and K perform scaled dot-product calculations to obtain attention scores Scores Attention : ， Among them, d k represents K the sequence dimension size of; Use the attention scores and V perform weighted calculations to generate the attention output Output Attention : ， Two residual connections and a feed-forward neural network are adopted to enhance the training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization: ， Among them, Residual 1st and Residual 2nd respectively represent the outputs of two residual connections. BatchNorm represents the batch normalization operation, LayerNorm represents the layer normalization operation, FFN represents the feed-forward neural network, and Output FFN represents the output of the feed-forward neural network. represents the residual component after being enhanced and modeled by the cyclic reconstruction attention mechanism.

10. A computer storage medium, on which a computer program is stored, characterized in that, When the computer executes the computer program, it implements the univariate ultra-short-term wind power prediction method based on two-stage trend decomposition according to any one of claims 1-7.

Images