Single variable ultra-short-term wind power forecasting method based on two-level trend decomposition
By combining two-level trend decomposition and cyclic reconstruction attention mechanism, the problem of low efficiency of time series feature extraction in wind power forecasting is solved, and accurate prediction of ultra-short-term wind power and improvement of grid stability are achieved.
Patent Information
- Application Number
- CN202510905119.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2045-07-02
AI Technical Summary
Existing wind power forecasting methods are inefficient in extracting time series features for ultra-short-term forecasting, and single-variable methods lack time series causality, which affects the safety and stability of power grid operation.
A univariate ultra-short-term wind power forecasting method based on two-level trend decomposition is adopted. Macro-scale, mesoscale trend components and residual components are generated through multiple trend decompositions. The causal convolution kernel is initialized with exponential distribution to extract multi-scale features, and the cyclic reconstruction attention mechanism is used to enhance the residual modeling. Finally, independent component mapping fusion is performed.
It achieves accurate prediction of ultra-short-term wind power, breaks through the multi-scale feature coupling limitations of traditional methods, enhances temporal causality and dynamic perception capabilities, and provides mechanism explainability and single-variable data adaptability.
Smart Images
Figure CN120414533B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wind power prediction, and relates to a single variable ultra-short-term wind power prediction method based on two-level trend decomposition. Background Art
[0002] The intermittent and fluctuating nature of wind power can easily cause disruptive power disturbances to the main power grid, even impacting the operational safety and stability of local power grids. Existing methods, such as ARIMA time series analysis and LSTM multivariate fusion models, still suffer from issues such as reliance on collaborative modeling of multivariate meteorological data and inefficient extraction of time series features for ultra-short-term forecasting.
[0003] Patent application number CN2023101478202 discloses a short-term power generation forecasting method based on ensemble learning. This method uses MSTL to fully extract the trend, period, and residual terms of the original sequence, and then predicts each component separately. The residual terms from the MSTL decomposition are further decomposed using the VMD algorithm to predict each component separately. While this method considers the periodicity of historical power data, it does not account for temporal causality, and single-variable methods have certain limitations. Summary of the Invention
[0004] The purpose of the present invention is to provide a univariate ultra-short-term wind power forecasting method based on two-level trend decomposition, perform two trend decompositions on historical wind power time series data, extract multi-scale trend features, construct a cyclic reconstruction attention mechanism to model the dynamic characteristics of the residual component, and fuse the multi-component prediction results to achieve accurate prediction of ultra-short-term wind power.
[0005] The technical solutions for achieving the purpose of the present invention are:
[0006] A single variable ultra-short-term wind power forecasting method based on two-level trend decomposition includes the following steps:
[0007] S01: Perform multiple trend decompositions on the historical wind power time series data to generate macro trend components, mesoscale trend components and residual components;
[0008] S02: The decomposed macro-trend component and mesoscale trend component are respectively initialized with exponential distribution causal convolution kernel to extract multi-scale trend features, and the adaptive weight of scale normalization is used to maintain temporal causality;
[0009] S03: Adopting the cyclic reconstruction attention mechanism to enhance residual modeling, and obtaining the dynamic characteristics of the residual component through sequence splicing and double residual connection;
[0010] S04: Perform linear processing on each component and fuse the results to generate an ultra-short-term wind power forecast value.
[0011] In the preferred technical solution, before step S01, the following steps are further included:
[0012] The univariate historical power data set of the wind farm is split and processed in batches, including:
[0013] 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;
[0014] 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;
[0015] Convert the processed tensor data into a time series format suitable for the prediction model input ;
[0016] 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 ;
[0017] 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.
[0018] In the preferred technical solution, step S01 specifically includes:
[0019] 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:
[0020] ;
[0021] 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;
[0022] Obtain temporary intermediate state components :
[0023] ;
[0024] The temporal intermediate state sequence is processed using a second causal convolutional layer to extract the mesoscale trend component:
[0025] ;
[0026] in, T Meso represents the mesoscale trend component, K Meso Represents the convolution kernel size used to extract the mesoscale trend component;
[0027] Remove the mesoscale trend component from the temporary intermediate state sequence to obtain the residual component R:
[0028] .
[0029] In the preferred technical solution, in order to achieve differentiated modeling of different time scales, the convolution kernel size of each causal convolution layer meets the following conditions:
[0030] ;
[0031] in, L seq Represents the time step length of the input sequence.
[0032] In the preferred technical solution, the initialization strategy of the convolution weight in step S02 adopts exponential distribution initialization:
[0033] ;
[0034] in, λ is the decay rate parameter, K c is the convolution kernel size, i is the time step index of the weight, w [ i ] indicates the i The convolution weight of the time step, exp represents the exponential operation;
[0035] Normalize the convolution weights along the sequence dimension:
[0036] ;
[0037] in, w ’ [ i ] indicates the i Normalize the convolution weights by the ratio of time steps.
[0038] In the preferred technical solution, step S03 specifically includes:
[0039] The cyclic reconstruction attention mechanism is used to enhance 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 to generate a multi-version spliced sequence:
[0040] ;
[0041] in, R cat Represents the residual component sequence after splicing, Cat represents the sequence splicing operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H Indicates the cycle period;
[0042] Generating three elements of attention through linear transformation Q 、 K and V :
[0043] ;
[0044] in: W Q , W K and W V represents the weight matrix, b Q , b K and b V Indicates the offset;
[0045] Introducing a learnable scaling factor s , use the Softmax function to perform scaled dot product calculation on Q and K to get the attention score Scores Attention :
[0046] ;
[0047] in, d k represent K The size of the sequence dimension;
[0048] Using attention scores and V Perform weighted calculations to generate attention output Attention :
[0049] ;
[0050] Two residual connections and feedforward neural networks are used to enhance training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization:
[0051] ;
[0052] Among them, Residual 1st 、Residual 2nd Represents two residual connection outputs, BatchNorm represents batch normalization operation, LayerNorm represents layer normalization operation, FFN represents feedforward neural network, Output FFN represents the output of the feedforward neural network, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling.
[0053] In the preferred technical solution, step S04 of generating an ultra-short-term wind power forecast value includes:
[0054] Independent linear mapping layers are established for the macro-trend component, meso-scale trend component and residual component generated by the two-level trend decomposition;
[0055] Use the linear mapping output of each component to perform unbiased summation and use the component mapping fusion result as the final prediction output:
[0056] ;
[0057] in, T Macro Represents the macro trend component, T Meso represents the mesoscale trend component, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling, W macro , W meso and W Res Represents the mapping weight matrix applied to the three components respectively, P pred Represents the predicted output.
[0058] The present invention also discloses a single variable ultra-short-term wind power prediction system based on two-level trend decomposition, comprising:
[0059] The two-level trend decomposition module performs multiple trend decompositions on historical wind power time series data to generate macro trend components, mesoscale trend components, and residual components;
[0060] The trend modeling module uses exponential distribution to initialize causal convolution kernels for the decomposed macro-trend components and mesoscale trend components to extract multi-scale trend features, and uses proportionally normalized adaptive weights to maintain temporal causality;
[0061] The residual modeling module uses a cyclic reconstruction attention mechanism to enhance residual modeling and obtains the dynamic characteristics of the residual component through sequence splicing and double residual connection;
[0062] The mapping prediction module performs linear processing on each component and fuses the results to generate an ultra-short-term wind power forecast value.
[0063] The present invention further discloses a computer storage medium on which a computer program is stored. When the computer program is executed, the single variable ultra-short-term wind power prediction method based on two-level trend decomposition is realized.
[0064] Compared with the prior art, the present invention has the following significant advantages:
[0065] First, based on the cascaded causal convolution and two-level trend decomposition technology, the original power time series data is decoupled into macro trend component, mesoscale trend component and residual component. The differentiated receptive field design of the two-level convolution kernel is used to achieve spatiotemporal feature separation across time scales, breaking through the limitations of traditional single-variable methods in modeling multi-scale feature coupling.
[0066] Secondly, the proposed dual weight constraint strategy of exponential distribution initialization and scale normalization constructs a temporal causal barrier from the source of convolution kernel parameter initialization, enhances the dynamic perception ability of current moment features through exponential decay weight distribution, and cooperates with the scale normalization operation along the time dimension to eliminate parameter sensitivity in long-range dependency modeling, avoiding gradient vanishing or gradient explosion during training.
[0067] Finally, a component-independent mapping fusion mechanism is employed to perform hierarchical feature decoupling and directional mapping fusion, constructing independently optimized representation spaces for features at different time scales. The component mapping fusion results are then used as the final prediction output. This provides a solution for univariate ultra-short-term wind power forecasting that combines mechanism interpretability with univariate data adaptability. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 This is a flow chart of a single variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to this embodiment;
[0069] Figure 2 This is a comparison chart of the predicted results and true values when DM (hidden layer dimension) = 16 and LR (learning rate) = 0.0001;
[0070] Figure 3 This is a comparison chart of the predicted results and true values when DM=32 and LR=0.0001;
[0071] Figure 4 This is a comparison chart of the predicted results and true values when DM=64 and LR=0.0001;
[0072] Figure 5 This is a comparison chart of the predicted results and the true values when DM=32 and LR=0.0005;
[0073] Figure 6 This is a comparison chart of the predicted results and true values when DM=32 and LR=0.00005. DETAILED DESCRIPTION
[0074] The principles of the present invention are as follows: performing multiple trend decompositions on historical wind power time series data to generate macro-trend components, mesoscale trend components, and residual components; extracting multi-scale trend features through exponentially initialized causal convolution kernels, and using scale-normalized adaptive weights to maintain temporal causality; constructing a cyclic reconstructed attention mechanism to model the dynamic characteristics of the residual component; and employing a component-independent mapping fusion mechanism for hierarchical feature decoupling and directional mapping fusion, constructing independently optimized representation spaces for features at different time scales, and using the component mapping fusion results as the final prediction output. This invention provides a solution for univariate ultra-short-term wind power forecasting that combines mechanism interpretability with univariate data adaptability.
[0075] Example 1:
[0076] like Figure 1 As shown, a single variable ultra-short-term wind power forecasting method based on two-level trend decomposition includes the following steps:
[0077] S01: Perform multiple trend decompositions on the historical wind power time series data to generate macro trend components, mesoscale trend components and residual components;
[0078] S02: The decomposed macro-trend component and mesoscale trend component are respectively initialized with exponential distribution causal convolution kernel to extract multi-scale trend features, and the adaptive weight of scale normalization is used to maintain temporal causality;
[0079] S03: Adopting the cyclic reconstruction attention mechanism to enhance residual modeling, and obtaining the dynamic characteristics of the residual component through sequence splicing and double residual connection;
[0080] S04: Perform linear processing on each component and fuse the results to generate an ultra-short-term wind power forecast value.
[0081] Specifically, before step S01, the following steps are also included:
[0082] The univariate historical power data set of the wind farm is split and processed in batches, including:
[0083] Based on the Z-score standardization method, the mean is calculated based on the training data set μ and standard deviation σ, performs standardization on the full dataset x; its calculation formula is:
[0084] ;
[0085] Here, xstd represents the standardized time series. The time series data is divided into training set, validation set, and test set in a ratio of 7:1:2. 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.
[0086] Convert the processed tensor data into a time series format suitable for the prediction model input ;
[0087] 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 ;
[0088] 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.
[0089] In a preferred embodiment, step S01 specifically includes:
[0090] 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:
[0091] ;
[0092] 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;
[0093] Obtain temporary intermediate state components :
[0094] ;
[0095] The temporal intermediate state sequence is processed using a second causal convolutional layer to extract the mesoscale trend component:
[0096] ;
[0097] in, T Meso represents the mesoscale trend component, K Meso Represents the convolution kernel size used to extract the mesoscale trend component;
[0098] Remove the mesoscale trend component from the temporary intermediate state sequence to obtain the residual component R :
[0099] .
[0100] Preferably, in order to achieve differentiated modeling of different time scales, the convolution kernel size of each causal convolution layer meets the following conditions:
[0101] ;
[0102] in, L seq Represents the time step length of the input sequence.
[0103] Preferably, the initialization strategy of the convolution weight in step S02 adopts exponential distribution initialization:
[0104] ;
[0105] in, λ is the decay rate parameter, K c is the convolution kernel size, i is the time step index of the weight, w [ i ] indicates the i The convolution weight of the time step, exp represents the exponential operation;
[0106] Normalize the convolution weights along the sequence dimension:
[0107] ;
[0108] in, w ’ [ i ] indicates the i Normalize the convolution weights by the ratio of time steps.
[0109] In a preferred embodiment, step S03 specifically includes:
[0110] The cyclic reconstruction attention mechanism is used to enhance 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 to generate a multi-version spliced sequence:
[0111] ;
[0112] in, R cat Represents the residual component sequence after splicing, Cat represents the sequence splicing operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H Indicates the cycle period;
[0113] Generating three elements of attention through linear transformation Q 、 K and V :
[0114] ;
[0115] in: W Q , W K and W V represents the weight matrix, b Q , b K and b V Indicates the offset;
[0116] Introducing a learnable scaling factor s , using the Softmax function Q and K Perform scaled dot product calculations to obtain attention scores Attention :
[0117] ;
[0118] in, d k represent K The size of the sequence dimension;
[0119] Using attention scores and V Perform weighted calculations to generate attention output Attention :
[0120] ;
[0121] Two residual connections and feedforward neural networks are used to enhance training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization:
[0122] ;
[0123] Among them, Residual1st 、Residual 2nd Represents two residual connection outputs, BatchNorm represents batch normalization operation, LayerNorm represents layer normalization operation, FFN represents feedforward neural network, Output FFN represents the output of the feedforward neural network, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling.
[0124] Preferably, step S04 of generating an ultra-short-term wind power forecast value includes:
[0125] Independent linear mapping layers are established for the macro-trend component, meso-scale trend component and residual component generated by the two-level trend decomposition;
[0126] Use the linear mapping output of each component to perform unbiased summation and use the component mapping fusion result as the final prediction output:
[0127] ;
[0128] in, T Macro Represents the macro trend component, T Meso represents the mesoscale trend component, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling, W macro , W meso and W Res Represents the mapping weight matrix applied to the three components respectively, P pred Represents the predicted output.
[0129] In another embodiment, a computer storage medium stores a computer program that, when executed, implements the aforementioned univariate ultra-short-term wind power forecasting method based on two-level trend decomposition. The specific implementation utilizes the aforementioned wind power forecasting method and will not be further described here.
[0130] In another embodiment, a single variable ultra-short-term wind power forecasting system based on two-level trend decomposition includes:
[0131] The two-level trend decomposition module performs multiple trend decompositions on historical wind power time series data to generate macro trend components, mesoscale trend components, and residual components;
[0132] The trend modeling module uses exponential distribution to initialize causal convolution kernels for the decomposed macro-trend components and mesoscale trend components to extract multi-scale trend features, and uses proportionally normalized adaptive weights to maintain temporal causality;
[0133] The residual modeling module uses a cyclic reconstruction attention mechanism to enhance residual modeling and obtains the dynamic characteristics of the residual component through sequence splicing and double residual connection;
[0134] The mapping prediction module performs linear processing on each component and fuses the results to generate an ultra-short-term wind power forecast value.
[0135] The specific implementation adopts the above wind power prediction method, which will not be described here.
[0136] In order to effectively train the prediction model, appropriate parameter design and experimental design are required. The specific methods are as follows:
[0137] A wind power data set from a wind farm in East China was selected for ultra-short-term forecasting. The time resolution of the wind power series was 15 minutes. The data set was divided into training, validation, and test sets in a ratio of 7:1:2.
[0138] The parameters of the prediction model are designed as follows: prediction step PL=2, batch size BS=1024, learning rate LR=[0.00005, 0.0001, 0.0005], drop rate DP=0.4, hidden layer dimension DM=[16,32,64], and number of attention layers NL=4.
[0139] During training, the mean squared error (MSE) is used as the guiding loss function for backpropagation, and the mean absolute error (MAE) is provided as another angle reference for loss measurement to comprehensively evaluate model performance.
[0140] When DM=16 and LR=0.0001, the power prediction results are compared with the true value. Figure 2 shown.
[0141] When DM=32 and LR=0.0001, the power prediction results are compared with the true value as shown in the figure below. Figure 3 shown.
[0142] When DM=64 and LR=0.0001, the power prediction results are compared with the true value as shown in the figure below. Figure 4 shown.
[0143] When DM=32 and LR=0.0005, the power prediction results are compared with the true value. Figure 5 shown.
[0144] When DM=32 and LR=0.00005, the power prediction results are compared with the true value. Figure 6 shown.
[0145] Table 1 shows the experimental results of various groups of parameters in this embodiment.
[0146] Table 1: Experimental parameters and results of each group
[0147]
[0148] As can be seen from Table 1, the embodiment of the present invention achieves a low error level prediction of univariate wind power in the two-step ultra-short-term power prediction experiment of the above wind farm dataset, demonstrating the superiority of the present invention in univariate ultra-short-term wind power prediction.
[0149] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.
Claims
1. A single variable ultra-short-term wind power forecasting method based on two-level trend decomposition, characterized by: The following steps are involved: S01: Perform multiple trend decompositions on the historical wind power time series data to generate macro trend components, mesoscale trend components and residual components; S02: Use exponential distribution to initialize causal convolution kernels for the decomposed macro-trend component and mesoscale trend component to extract multi-scale trend features, and perform proportional normalization on the convolution weights along the sequence dimension to maintain temporal causality; S03: Adopting the cyclic reconstruction attention mechanism to strengthen 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 and generate an attention output; two residual connections and a feedforward neural network are used to enhance training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization to obtain the dynamic characteristics of the residual component; S04: Perform linear processing on each component and fuse the results to generate an ultra-short-term wind power forecast value.
2. The single variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to claim 1 is characterized in that: Before step S01, the following steps are also included: 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.
3. The single variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to claim 1 is characterized in that: 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: ， 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 : ， The temporal intermediate state sequence is processed using a second causal convolutional layer to extract the mesoscale trend component: ， in, 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 forecasting method based on two-level trend decomposition according to claim 3 is characterized in that: In order to achieve differentiated modeling of different time scales, the convolution kernel size of each causal convolution layer meets the following conditions: ， in, L seq Represents the time step length of the input sequence.
5. The single variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to claim 3 is characterized in that: The initialization strategy for the convolution weights in step S02 uses exponential distribution initialization: ， in, λ is the decay rate parameter, K c is the convolution kernel size, i is the time step index of the weight, w [ i ] indicates the i The convolution weight of the time step, exp represents the exponential operation; The formula for scaling the convolution weights along the sequence dimension is expressed as: ， in, w ’ [ i ] indicates the i Normalize the convolution weights by the ratio of time steps.
6. The single variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to claim 1 is characterized in that: Step S03 specifically includes: The formula for generating multi-version splicing sequences is expressed as: ， in, R cat Represents the residual component sequence after splicing, Cat represents the sequence splicing operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H Indicates the cycle period; Generating three elements of attention through linear transformation Q 、 K and V : ， in: W Q , W K and W V represents the weight matrix, b Q , b K and b V Indicates the offset; Introducing a learnable scaling factor s , using the Softmax function Q and K Perform scaled dot product calculations to obtain attention scores Attention : ， in, d k represent K The size of the sequence dimension; Using attention scores and V Perform weighted calculations to generate attention output Attention : ， Two residual connections and feedforward neural networks are used to enhance training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization. The formula is expressed as: ， Among them, Residual 1st 、Residual 2nd Represents two residual connection outputs, BatchNorm represents batch normalization operation, LayerNorm represents layer normalization operation, FFN represents feedforward neural network, Output FFN represents the output of the feedforward neural network, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling.
7. The single variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to claim 1 is characterized in that: Step S04 generates an ultra-short-term wind power forecast value, including: Independent linear mapping layers are established for the macro-trend component, meso-scale trend component and residual component generated by the two-level trend decomposition; Use the linear mapping output of each component to perform unbiased summation and use the component mapping fusion result as the final prediction output: ， in, T Macro Represents the macro trend component, T Meso represents the mesoscale trend component, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling, W macro , W meso and W Res Represents the mapping weight matrix applied to the three components respectively, P pred Represents the predicted output.
8. A single variable ultra-short-term wind power forecasting system based on two-level trend decomposition, characterized by: include: The two-level trend decomposition module performs multiple trend decompositions on historical wind power time series data to generate macro trend components, mesoscale trend components, and residual components; The trend modeling module uses exponential distribution to initialize causal convolution kernels for the decomposed macro-trend components and mesoscale trend components to extract multi-scale trend features, and normalizes the convolution weights along the sequence dimension to maintain temporal causality; The residual modeling module uses a cyclic reconstruction attention mechanism to enhance residual modeling. The residual component input sequence is segmented according to a specific cyclic period, and each sequence segment is cyclically shifted toward the end of the sequence to generate a multi-version spliced sequence and generate an attention output. Two residual connections and a feedforward neural network are used to enhance training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization to obtain the dynamic characteristics of the residual component. The mapping prediction module performs linear processing on each component and fuses the results to generate an ultra-short-term wind power forecast value.
9. The single variable ultra-short-term wind power forecasting system based on two-level trend decomposition according to claim 8, characterized in that: The residual modeling module specifically includes: The formula for generating multi-version splicing sequences is expressed as: ， in, R cat Represents the residual component sequence after splicing, Cat represents the sequence splicing operation, R is the residual component, Shift represents the cyclic shift operation of the sequence segment, H Indicates the cycle period; Generating three elements of attention through linear transformation Q 、 K and V : ， in: W Q , W K and W V represents the weight matrix, b Q , b K and b V Indicates the offset; Introducing a learnable scaling factor s , using the Softmax function Q and K Perform scaled dot product calculations to obtain attention scores Attention : ， in, d k represent K The size of the sequence dimension; Using attention scores and V Perform weighted calculations to generate attention output Attention : ， Two residual connections and feedforward neural networks are used to enhance training stability. The first residual connection uses post-batch normalization, and the second residual connection uses post-layer normalization. The formula is expressed as: ， Among them, Residual 1st 、Residual 2nd Represents two residual connection outputs, BatchNorm represents batch normalization operation, LayerNorm represents layer normalization operation, FFN represents feedforward neural network, Output FFN represents the output of the feedforward neural network, Represents the residual component after the cyclic reconstruction attention mechanism strengthens the modeling.
10. A computer storage medium having a computer program stored thereon, characterized in that: When a computer executes the computer program, the single-variable ultra-short-term wind power forecasting method based on two-level trend decomposition according to any one of claims 1 to 7 is implemented.