Short-term wind power prediction method based on multi-scale dominant component analysis and hybrid data driving

Abstract

The invention discloses a short-term wind power prediction method based on multi-scale dominant component analysis and mixed data driving, and the method comprises the steps: carrying out the decomposition of wind power data through VMD (variational mode decomposition), and decomposing a wind power sequence into n mode components; extracting a dominant component from each modal component based on singular spectrum analysis SSA and taking the dominant component as new modal data; meanwhile, superposing the residual part after extraction of each modal component with a variational modal decomposition (VMD) residual error to serve as an (n + 1) th modal; constructing a mixed data driving model based on TCN and SWLSTM, and respectively predicting multi-modal data obtained by variational mode decomposition (VMD) and singular spectrum analysis (SSA); and reconstructing prediction results of all modal data to obtain a final wind power prediction value. The invention provides a short-term wind power prediction method based on multi-scale dominant component analysis and mixed data driving, which can effectively improve the short-term wind power prediction precision and improve the prediction efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a short-term wind power prediction method based on multi-scale dominant component analysis and hybrid data driving. Background Art

[0002] With the continuous growth of global energy demand and the continuous advancement of the dual-carbon policy, the development and utilization of clean and renewable energy has become a focus of attention for governments and research institutions around the world. As an important renewable energy source, wind energy accounts for an increasing proportion of the power system year by year due to its abundant resources, wide distribution and low impact on the environment. However, the intermittent and volatile nature of wind power poses certain challenges to the operation of wind farms and the stability of power grids, and also affects their absorption. Accurate wind power forecasting is of great significance to improving the comprehensive power generation benefits of wind farms and the reliability of power grids.

[0003] Traditional wind power prediction methods mainly include physical methods and statistical methods. With the development of artificial intelligence and machine learning technology, wind power prediction methods based on deep learning have gradually become a research hotspot. Deep learning models can automatically extract features from data and effectively handle complex nonlinear relationships, so they perform well in wind power prediction.

[0004] For example, in the literature [1]: Zhang Shengjie, Zhang Xuefeng, Xie Zhongzhou, et al. Research on wind power prediction method based on LSTM [J]. Energy and Energy Conservation, 2024, (08): 11-13. LSTM is used to capture the temporal characteristics and long-term dependencies of wind power data, fully learn the changing trend of wind power series, and obtain more accurate wind power prediction results. However, due to the single prediction model, it is unable to effectively deal with the highly nonlinear and multi-scale characteristics of wind power data.

[0005] Reference [2]: Zhu Juping, Wei Xia, Xie Lirong, et al. Short-term wind power forecasting based on VMD and improved BiLSTM [J]. Acta Energiae Solaris Sinica, 2024, 45(06): 422-428. By introducing the attention mechanism, corresponding weights are assigned to the hidden layer state output of BiLSTM to highlight the influence of important input features, and extreme gradient boosting is used to correct the prediction error, thereby further improving the prediction accuracy. However, it still faces problems such as overfitting, high computational complexity and dependence on large-scale data. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive, which can effectively improve the short-term wind power forecasting accuracy and improve the forecasting efficiency.

[0007] The technical solution adopted by the present invention is:

[0008] The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data driving includes the following steps:

[0009] Step 1: Decompose the wind power data using variational mode decomposition (VMD) to decompose the wind power series into n modal components;

[0010] Step 2: Based on singular spectrum analysis (SSA), the dominant component of each modal component is extracted and used as the new modal data; at the same time, the remaining part extracted from each modal component is superimposed with the variational mode decomposition (VMD) residual as the n+1th mode;

[0011] Step 3: Construct a hybrid data-driven model based on TCN and SWLSTM to predict the multimodal data obtained by variational mode decomposition (VMD) and singular spectrum analysis (SSA) respectively;

[0012] Step 4: Reconstruct the prediction results of all modal data to obtain the final wind power prediction value.

[0013] In step 1, by decomposing the original wind power signal into multiple frequency components, variational mode decomposition (VMD) can more accurately capture changes on different time scales and provide better input features for subsequent prediction models;

[0014] Furthermore, the core of variational mode decomposition (VMD) is to achieve signal decomposition by solving constrained variational problems. For a given signal f(t), the goal is to deduce K modes u k (t) and the corresponding center frequency ω k , so that:

[0015]

[0016] Where K is the number of modes, k = 1, 2, …, K.

[0017] The optimization objective function is expressed as:

[0018]

[0019] Where: u k (t) represents the mode, w k represents the center frequency, * represents the convolution operation, j is the imaginary unit, t is the time variable, and δ(t) represents the Dirac impulse function.

[0020] After using the Lagrange multiplier method to construct the augmented variational problem, combined with the alternating direction multiplier method, the mode u can be iteratively obtained: k (t) and center frequency ω kSpecifically, the basic idea of ​​the Lagrange multiplier method is to introduce constraints into the objective function and solve it by constructing an augmented Lagrangian function. Assuming that there is a constraint condition g(u k ,w k )=0, then the optimization objective function can be:

[0021]

[0022] Where λ is the Lagrange multiplier and ρ is the penalty coefficient.

[0023] The alternating direction multiplier rule is to decompose the difficult constrained optimization problem into multiple sub-problems that are easier to solve, and then optimize these sub-problems alternately. k (t) and w k (t) Optimization:

[0024]

[0025]

[0026] In the formula, and is the updated u after iteration k and w k .

[0027] At the same time, the Lagrange multiplier λ and the multiplier ρ need to be updated to strengthen the penalty for the constraints. The update rules are as follows:

[0028]

[0029] In the formula, λ (t+1) represents the updated λ.

[0030] In step 2, singular spectrum analysis (SSA) is a non-parametric method for time series analysis. It combines the ideas of singular value decomposition and time series smoothing, and decomposes the time series into several components with specific structures to extract features of complex signals. It mainly includes the following steps:

[0031] S2.1. Data Embedding:

[0032] Convert the time series x(t) into the trajectory matrix X. For a time series x(t) of length N, when the window length is L, the trajectory matrix X can be defined as:

[0033]

[0034] Where: x(1), x(2), x(3), x(L), x(L+1), x(N-L+1), x(N-L+2), and x(N) represent the elements in the trajectory matrix.

[0035] S2.2, singular value decomposition:

[0036] Perform singular value decomposition on the trajectory matrix X and decompose it into the product of three matrices:

[0037] X=UΣV T

[0038] Where: U is an (N-L+1)×(N-L+1) orthogonal matrix; Σ is a diagonal matrix containing singular values; V is an L×L orthogonal matrix.

[0039] S2.3, Reconstruction:

[0040] Singular values ​​are selected based on their cumulative contribution rates, and the diagonal feature matrix composed of the selected singular values ​​is used to perform reverse reconstruction of the time series. The contribution rate of each singular value is the square of the singular value divided by the sum of the squares of all singular values:

[0041]

[0042] In the formula, CR represents the contribution rate, σ i is the i-th singular value, σ j is the jth singular value and r is the total number of singular values.

[0043] The cumulative contribution rate CCR is the sum of the contribution rates from the first singular value to the i-th singular value, that is:

[0044]

[0045] Where CRj is the contribution rate of the jth singular value.

[0046] Assuming that the first n singular values ​​are selected, a new matrix can be obtained based on the selected singular values ​​and the corresponding eigenvectors:

[0047]

[0048] Where, X n is the matrix reconstructed by selecting the first n singular values, U n are the first n columns and rows extracted from U, V n are the first n columns and first n rows extracted from V, Σ n are the first n singular values ​​extracted.

[0049] The reconstruction of the time series can be obtained by inverse embedding:

[0050]

[0051] In the formula, represents the reconstructed time series value, w iis the reconstruction weight, X n (t-i+1,;) represents the reconstruction matrix X n The t-i+1th row in .

[0052] In step 2, the remaining part after extraction from each modal component is superimposed with the variational mode decomposition (VMD) residual as the n+1th mode. Specifically, the VMD residual is the part that fails to fully fit the original signal during the VMD process and can be expressed as:

[0053]

[0054] Where e(t) is the VMD residual, f(t) is the original signal, and u k (t) is the model component obtained by VMD decomposition.

[0055] The remaining parts of all modes are superimposed on the VMD residual to form a new mode, namely:

[0056]

[0057] In the formula, u n+1 (t) represents the n+1th mode, r k (t) represents the remainder of all modes.

[0058] SSA can effectively remove high-frequency noise in time series, extract the main trends and periodic components, and help reduce the non-stationarity of wind power data and thus improve the accuracy of wind power prediction.

[0059] In step 3, the hybrid data-driven model is constructed by TCN and SWLSTM, wherein TCN can effectively capture local features in long time series by stacking convolutional layers and residual connections, and the weight matrix of SWLSTM is shared between different time steps and different memory units, effectively reducing the risk of overfitting and being able to more effectively learn patterns in time series;

[0060] Furthermore, TCN, as a deep learning architecture for time series prediction and modeling, combines the parallel computing capabilities of CNN and the sequence modeling capabilities of RNN, and is particularly suitable for processing time series data with long-term dependencies. Assume that the input sequence is x = (x1, x2, ..., x T ), where x1, x2, …, x T Represent the elements in the sequence respectively, T is the length of the sequence, then the general form of TCN can be expressed as:

[0061] y=f(x;θ)

[0062] Where: y is the output sequence, θ is the parameter of the model, and f(x; θ) represents a function that calculates the output y based on the input data x and the model parameters θ.

[0063] Specifically, the convolution operation and residual connection in TCN can be expressed as:

[0064] h (l+1) =ReLU(Conv d(l) (h (l) )+h (l) )

[0065] Where: h (l+1) represents the output of the l+1th layer; ReLU is the activation function; Conv d(l) It is a convolution operation with causal properties; h (l) is the output of the lth layer; d(l) is the size of the convolution kernel of the lth layer.

[0066] Furthermore, SWLSTM reduces the number of network parameters by sharing weights. Its update rule is similar to that of the long short-term memory network (LSTM), but its weight matrix is ​​restricted to a shared form. The state update rule is as follows:

[0067] i t =σ(W i x t +U i h t-1 +b i );

[0068] Where: i t represents the input gate, W i represents the input gate weight matrix, x t is the input vector; U i represents the input gate weight matrix of the previous time step, h t-1 is the hidden state of the previous time step; b i represents the input gate bias term, σ is the sigmoid function;

[0069] f t =σ(W f x t +U f h t-1 +b f );

[0070] Where: f t represents the forget gate, W f represents the forget gate weight matrix, U f represents the forget gate weight matrix of the previous time step, b f Represents the forget gate bias term.

[0071] o t =σ(W o x t +U o h t-1 +b o );

[0072] Where: t represents the output gate, W o represents the output gate weight matrix, U o represents the output gate weight matrix of the previous time step, h t-1 is the hidden state of the previous time step; b o Represents the output gate bias term.

[0073] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c );

[0074] Where: c t Indicates the cell state, c t-1 is the unit state of the previous unit step; W c The weight matrix representing the unit state, U c represents the unit state weight matrix of the previous time step, b c Represents the unit state bias term.

[0075] h t =o t ⊙tanh(c t );

[0076] Where: h t represents the hidden state and ⊙ represents element-wise multiplication.

[0077] In step 3, the multimodal data obtained by variational mode decomposition (VMD) and singular spectrum analysis (SSA) are predicted separately, and the modal data are input one by one into the hybrid data-driven model constructed by TCN and SWLSTM for prediction, and the prediction results of each modal data are obtained, which can be expressed as:

[0078]

[0079] In the formula, represents the predicted value of the mth mode; Mode represents the hybrid data-driven model; u m (t) represents the mth mode.

[0080] In step 4, the prediction results of all modal data are reconstructed to obtain the final wind power prediction value. The reconstruction formula is as follows:

[0081]

[0082] In the formula, is the final wind power forecast value, is the predicted value of the mth mode, and K is the total number of modes.

[0083] The present invention provides a short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive, and the technical effects are as follows:

[0084] 1) The present invention adopts a multi-scale dominant component analysis (MDCA) strategy based on VMD and SSA to reduce the impact of the non-stationarity of wind power data. VMD can perform multi-scale decomposition on the original wind power data to capture different frequency characteristics; SSA extracts dominant components to further reveal the potential structure and key patterns in the data.

[0085] 2) The present invention constructs a hybrid data-driven model based on TCN and SWLSTM to predict the extracted modal sequence. TCN can effectively capture local features in long time series by stacking convolutional layers and residual connections; SWLSTM reduces the number of network parameters by sharing weights, thereby accelerating the training process and improving the overall performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 is a flow chart of the present invention;

[0087] Figure 2 Schematic diagram of the TCN structure involved in the present invention.

[0088] Figure 3 This is a schematic diagram of the SWLSTM structure involved in the present invention.

[0089] Figure 4 This is a diagram of original wind power data according to a preferred embodiment of the present invention.

[0090] Figure 5 This is the wind power prediction result of the preferred embodiment of the present invention.

[0091] Figure 6 The following are the control experiment results of the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0092] A short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data-driven is as follows: Figure 1 As shown, it mainly includes the following steps:

[0093] S1: Variational mode decomposition (VMD) is used to decompose the original wind power data and decompose the complex wind power sequence into n modal components;

[0094] S2: Based on singular spectrum analysis (SSA), the dominant component of each modal component is extracted and used as the new modal data. At the same time, the remaining part extracted from each component is superimposed with the VMD decomposition residual as the n+1th mode;

[0095] S3: Construct a hybrid data-driven model based on TCN and SWLSTM to predict the multimodal data obtained by VMD and SSA respectively;

[0096] S4: Reconstruct the prediction results of all modal data to obtain the final wind power prediction value.

[0097] Furthermore, in S1, VMD is an adaptive signal decomposition method with good local characteristics and is widely used in signal processing and time series analysis. Different from the iterative solution of empirical mode decomposition, the core of VMD is to achieve signal decomposition by solving constrained variational problems. For a given signal f(t), the goal is to deduce K modes u k (t) and the corresponding center frequency ω k , so that:

[0098]

[0099] Where K is the number of modes, k = 1, 2, …, K.

[0100] The optimization objective function is expressed as:

[0101]

[0102] Where: u k (t) represents the mode, w k represents the center frequency, * represents the convolution operation, j is the imaginary unit, t is the time variable, and δ(t) represents the Dirac impulse function.

[0103] After using the Lagrange multiplier method to construct the augmented variational problem, combined with the alternating direction multiplier method, u can be iterated to obtain k (t) and ω k Specifically, the basic idea of ​​the Lagrange multiplier method is to introduce constraints into the objective function and solve it by constructing an augmented Lagrangian function. Assuming that there is a constraint condition g(u k ,w k )=0, then the optimization objective function can be written as:

[0104]

[0105] Where λ is the Lagrange multiplier and ρ is the penalty coefficient.

[0106] The alternating direction multiplier rule is to decompose the difficult constrained optimization problem into multiple sub-problems that are easier to solve, and then optimize these sub-problems alternately. k (t) and w k (t) Optimization:

[0107]

[0108] In the formula, and is the updated u after iteration k and w k .

[0109] At the same time, the Lagrange multiplier λ and the multiplier ρ need to be updated to strengthen the penalty for the constraints. The update rules are as follows:

[0110]

[0111] In the formula, λ (t+1) represents the updated λ.

[0112] By decomposing the original wind power data into multiple frequency components, VMD can more accurately capture the changes on different time scales and provide better input features for subsequent prediction models.

[0113] Furthermore, in S2, SSA, as a non-parametric method for time series analysis, combines the ideas of singular value decomposition and time series smoothing, and realizes feature extraction of complex signals by decomposing the time series into several components with specific structures. The specific steps of SSA are as follows:

[0114] s21: Data embedding. Convert the time series x(t) into a trajectory matrix. For a time series of length N, when the window length is L, the trajectory matrix X can be defined as:

[0115]

[0116] Where: x(1), x(2), x(3), x(L), x(L+1), x(N-L+1), x(N-L+2), and x(N) represent the elements in the trajectory matrix.

[0117] s22: Singular value decomposition. Perform singular value decomposition on the trajectory matrix X and decompose it into the product of three matrices:

[0118] X=UΣV T

[0119] Where: U is an (N-L+1)×(N-L+1) orthogonal matrix, Σ is a diagonal matrix containing singular values, and V is an L×L orthogonal matrix.

[0120] s23: Reconstruction. Singular values ​​are selected based on their cumulative contribution rates, and the diagonal feature matrix composed of the selected singular values ​​is used to perform reverse reconstruction of the time series. The contribution rate of each singular value is the square of the singular value divided by the sum of the squares of all singular values:

[0121]

[0122] In the formula, CR represents the contribution rate, σ i is the i-th singular value, σ j is the jth singular value and r is the total number of singular values.

[0123] The cumulative contribution rate CCR is the sum of the contribution rates from the first singular value to the i-th singular value, that is:

[0124]

[0125] Where CRj is the contribution rate of the jth singular value.

[0126] Assuming that the first n singular values ​​are selected, a new matrix can be obtained based on the selected singular values ​​and the corresponding eigenvectors:

[0127]

[0128] Where, X n is the matrix reconstructed by selecting the first n singular values, U n are the first n columns and rows extracted from U, V n are the first n columns and first n rows extracted from V, Σ n are the first n singular values ​​extracted.

[0129] The reconstruction of the time series can be obtained by inverse embedding:

[0130]

[0131] In the formula, represents the reconstructed time series value, w i is the reconstruction weight, X n (t-i+1,;) represents the reconstruction matrix X n The t-i+1th row in .

[0132] SSA can effectively remove high-frequency noise in time series, extract the main trends and periodic components, and help reduce the non-stationarity of wind power data and thus improve the accuracy of wind power prediction.

[0133] Furthermore, in S3, TCN is a deep learning architecture for time series prediction and modeling, which combines the parallel computing capability of CNN and the sequence modeling capability of RNN, and is particularly suitable for processing time series data with long-term dependencies. Traditional RNNs are prone to performance degradation when processing long-term dependencies due to the problem of gradient vanishing or gradient exploding. To solve this problem, TCN introduces convolution operations, so that the model can effectively capture long-term dependencies in time series, while taking advantage of parallel computing to accelerate the training process.

[0134] The core idea of ​​TCN is to expand the receptive field of the network by stacking convolutional layers, thereby capturing patterns at different time scales in the time series. Its structure is as follows: Figure 2 shown.

[0135] Assume that the input sequence is x=(x1,x2,…,x T ), where T is the length of the sequence, the general form of TCN can be expressed as:

[0136] y=f(x;θ)

[0137] Where: y is the output sequence, θ is the parameter of the model, and f(x; θ) represents a function that calculates the output y based on the input data x and the model parameters θ.

[0138] Specifically, the convolution operation and residual connection in TCN can be expressed as:

[0139] h (l+1) =ReLU(Conv d(l) (h (l) )+h (l) )

[0140] Where: h (l+1) Represents the output of the l+1th layer, ReLU is the activation function; Conv d(l) It is a convolution operation with causal properties; h (l) is the output of the lth layer; d(l) is the size of the convolution kernel of the lth layer.

[0141] TCN can effectively capture local features in long time series by stacking convolutional layers and residual connections. Through effective parameter sharing and parallel computing, TCN has good scalability and performance when processing large-scale time series data.

[0142] Furthermore, in S4, SWLSTM is an optimized LSTM that is effectively used to process and predict time series data. The core idea of ​​LSTM is to store long-term dependencies through memory cells and control the flow of information through gating cells, including forget gates, input gates, and output gates. SWLSTM reduces the number of parameters while retaining the characteristics of LSTM to improve the efficiency and generalization ability of the model. Its structure is as follows: Figure 3 shown.

[0143] SWLSTM reduces the number of network parameters by sharing weights, thereby accelerating the training process and improving the overall performance of the model. In wind power forecasting, this design can more effectively capture the long-term dependencies and nonlinear dynamic characteristics in wind power sequences. The update rule of SWLSTM is similar to that of traditional LSTM, but its weight matrix is ​​restricted to a shared form. The state update rule is as follows:

[0144] i t =σ(W i x t +U i h t-1 +b i )

[0145] f t =σ(W f x t +U f h t-1 +b f )

[0146] o t =σ(W o x t +U o h t-1 +b o )

[0147] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c )

[0148] h t =o t ⊙tanh(c t )

[0149] Where: i t represents the input gate, W i represents the input gate weight matrix, x t is the input vector; Ui represents the input gate weight matrix of the previous time step, h t-1 is the hidden state of the previous time step; b i represents the input gate bias term, σ is the sigmoid function; f t represents the forget gate, W f represents the forget gate weight matrix, U f represents the forget gate weight matrix of the previous time step, b f represents the forget gate bias term; o t represents the output gate, W o represents the output gate weight matrix, U o represents the output gate weight matrix of the previous time step, h t-1 is the hidden state of the previous time step; b o represents the output gate bias term; c t Indicates the unit state, c t-1 is the unit state of the previous unit step; W c The weight matrix representing the unit state, U c represents the unit state weight matrix of the previous time step, b c represents the unit state bias; h t represents the hidden state and ⊙ represents element-wise multiplication.

[0150] Compared with LSTM, the weight matrix of SWLSTM is shared between different time steps and different memory units, which means that the number of network parameters is greatly reduced, thereby reducing the risk of overfitting and being able to learn patterns in time series more effectively. At the same time, shared weights reduce the amount of calculation per time step, speeding up training and reasoning, especially when processing long sequences.

[0151] Furthermore, in S4, the reconstruction method is to accumulate the prediction results of each modal data.

[0152] The following is a detailed description with reference to specific embodiments:

[0153] A preferred embodiment of the present invention is an example verification of a short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data-driven model on wind power data of a wind farm in Gansu Province.

[0154] The wind power data sampling period is from January 1 to January 15 of a certain year, with a sampling interval of 15 minutes, for a total of 1440 samples. Figure 4 Among them, the last 288 samples are selected as the test set to verify the prediction effect of the proposed method.

[0155] In the numerical experiment, the particle swarm optimization algorithm was used to determine the K value of VMD to be 5, the convolution kernel size of TCN to be 16, the number of nodes of SWLSTM to be 64, and 100 iterations of training were performed. The experimental results are shown in Figure 2. Figure 5 As shown. Figure 5 It can be seen that the curve trend between the predicted value and the actual value is very close, indicating that the method proposed in the present invention has good performance in wind power forecasting. Specifically, the predicted value can better follow the change trend of the actual value, especially in the period when the wind power changes greatly, the prediction accuracy of the model still remains at a high level.

[0156] In order to further verify the effectiveness of each module of the method proposed in this invention, different control experiments are set up, including module settings with and without MDCA, and the selection of different hybrid data-driven models such as TCN-LSTM, CNN-SWLSTM, and TCN-SWLSTM.

[0157] Preferably, in order to quantitatively and comprehensively describe the prediction performance of the prediction model, the present invention uses four performance indicators commonly used in the prediction field to evaluate the performance of the prediction model, namely: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and determination coefficient R 2 The calculation formulas for each evaluation index are as follows:

[0158]

[0159] Where: N is the sequence length, y i , and They are the actual value, predicted value and average value respectively.

[0160] The results of the control experiment are as follows Figure 6 The performance indicators of each prediction method are shown in Table 1.

[0161] Table 1 Prediction performance of different methods

[0162]

[0163] Depend on Figure 6 The comparison of the prediction result curves shown in the figure shows that the prediction result of the proposed method has the smallest deviation from the actual value. At the same time, it has the highest fitting degree among all experimental methods and has achieved the best prediction effect. Specifically, Figure 6 The prediction effect of the upper part is significantly better than that of the lower part, which shows that the MDCA method has higher accuracy in capturing the trend of wind power changes. Figure 6In both the upper and lower parts, the combined method of TCN-SWLSTM achieved the best prediction results, which reflects the effectiveness of the constructed hybrid data-driven model in wind power forecasting.

[0164] As can be seen from Table 1, the introduction of MDCA significantly improves the prediction performance of the model. Compared with the method without adding MDCA, the method after adding MDCA shows obvious improvement in various indicators, especially in MAE, RMSE and MAPE. At the same time, the combination of TCN-SWLSTM shows the best prediction performance both with and without adding MDCA. This shows that in the task of wind power prediction, the introduction of the hybrid data-driven model proposed in the present invention can effectively improve the prediction accuracy and the data fitting ability of the model.

Claims

1. A short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data driven, characterized by The following steps are involved: Step 1: Decompose the wind power data using variational mode decomposition (VMD) to decompose the wind power series into n modal components; Step 2: Based on singular spectrum analysis (SSA), the dominant component of each modal component is extracted and used as the new modal data; at the same time, the remaining part extracted from each modal component is superimposed with the variational mode decomposition (VMD) residual as the n+1th mode; Step 3: Construct a hybrid data-driven model based on TCN and SWLSTM to predict the multimodal data obtained by variational mode decomposition (VMD) and singular spectrum analysis (SSA) respectively; Step 4: Reconstruct the prediction results of all modal data to obtain the final wind power prediction value.

2. The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive according to claim 1 is characterized by: In step 1, variational mode decomposition (VMD) realizes signal decomposition by solving constrained variational problems; for a given signal f(t), the goal is to deduce K modes u k (t) and the corresponding center frequency ω k , so that: Where K is the number of modes, k = 1, 2, ..., K; The optimization objective function is expressed as: Where: u k (t) represents the mode, w k represents the center frequency, * represents the convolution operation; j is the imaginary unit; t is the time variable, δ(t) represents the Dirac impulse function; After using the Lagrange multiplier method to construct the augmented variational problem, combined with the alternating direction multiplier method, the mode u can be iteratively obtained: k (t) and center frequency ω k ; The details are as follows: The Lagrange multiplier method introduces constraints into the objective function and solves it by constructing an augmented Lagrangian function. Assuming that there is a constraint condition g(u k ,w k )=0, then the optimization objective function is: Where λ is the Lagrange multiplier and ρ is the penalty coefficient; The alternating direction multiplier rule is to decompose the difficult constrained optimization problem into multiple sub-problems that are easier to solve, and then optimize these sub-problems alternately. k (t) and w k (t) Optimization: In the formula, and is the updated u after iteration k and w k ; At the same time, the Lagrange multiplier λ and the multiplier ρ need to be updated to strengthen the penalty for the constraints. The update rules are as follows: In the formula, λ (t+1) represents the updated λ.

3. The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive according to claim 1 is characterized by: In step 2, singular spectrum analysis (SSA) decomposes the time series into several components with specific structures to achieve feature extraction of complex signals, including the following steps: S2.

1. Data Embedding: Convert the time series x(t) into the trajectory matrix X. For a time series x(t) of length N, when the window length is L, the trajectory matrix X can be defined as: Where: x(1), x(2), x(3), x(L), x(L+1), x(N-L+1), x(N-L+2), x(N) represent the elements in the trajectory matrix; S2.2, singular value decomposition: Perform singular value decomposition on the trajectory matrix X and decompose it into the product of three matrices: X=UΣV T Where: U is an orthogonal matrix of (N-L+1)×(N-L+1); Σ is a diagonal matrix containing singular values; V is an orthogonal matrix of L×L; S2.3, Reconstruction: Singular values ​​are selected based on their cumulative contribution rates, and the diagonal feature matrix composed of the selected singular values ​​is used to perform reverse reconstruction of the time series; the contribution rate of each singular value is the square of the singular value divided by the sum of the squares of all singular values: In the formula, CR represents the contribution rate, σ i is the i-th singular value, σ j is the jth singular value, r is the total number of singular values; The cumulative contribution rate CCR is the sum of the contribution rates from the first singular value to the i-th singular value, that is: Where CRj is the contribution rate of the jth singular value; Suppose we select the first n singular values, and get a new matrix based on the selected singular values ​​and the corresponding eigenvectors: Where, X n is the matrix reconstructed by selecting the first n singular values, U n are the first n columns and rows extracted from U, V n are the first n columns and first n rows extracted from V, Σ n are the first n singular values ​​extracted; The reconstruction of the time series is obtained by inverse embedding: In the formula, represents the reconstructed time series value, w i is the reconstruction weight, X n (t-i+1,;) represents the reconstruction matrix X n The t-i+1th row in .

4. The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive according to claim 3 is characterized by: The remaining part extracted from each modal component is superimposed with the variational mode decomposition (VMD) residual as the n+1th mode, as follows: The VMD residual is the part that fails to fully fit the original signal during the VMD process and is expressed as: Where e(t) is the VMD residual, f(t) is the original signal, and u k (t) is the model component obtained by VMD decomposition; The remaining parts of all modes are superimposed on the VMD residual to form a new mode, namely: In the formula, u n+1 (t) represents the n+1th mode, r k (t) represents the remainder of all modes.

5. The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive according to claim 1 is characterized by: In step 3, the hybrid data-driven model is constructed by TCN and SWLSTM; TCN is suitable for processing time series data with long-term dependencies; let the input sequence be x = (x1, x2, ..., x T ), where x1, x2, …, x T Represent the elements in the sequence respectively, T is the length of the sequence, then the general form of TCN is expressed as: y=f(x;θ) Where: y is the output sequence, θ is the parameter of the model, f(x;θ) represents a function that calculates the output y based on the input data x and the model parameters θ; Specifically, the convolution operation and residual connection in TCN are expressed as: h (l+1) =ReLU(Conv d(l) (h (l) )+h (l) ) Where: h (l+1) represents the output of the l+1th layer; ReLU is the activation function; Conv d(l) It is a convolution operation with causal properties; h (l) is the output of the lth layer; d(l) is the size of the convolution kernel of the lth layer; SWLSTM reduces the number of network parameters by sharing weights. Its weight matrix is ​​restricted to a shared form, and the state update rules are as follows: i t =σ(W i x t +U i h t-1 +b i ); Where: i t represents the input gate, W i represents the input gate weight matrix, x t is the input vector; U i represents the input gate weight matrix of the previous time step, h t-1 is the hidden state of the previous time step; b i represents the input gate bias term, σ is the sigmoid function; f t =σ(W f x t +U f h t-1 +b f ); Where: f t represents the forget gate, W f represents the forget gate weight matrix, U f represents the forget gate weight matrix of the previous time step, b f Represents the forget gate bias term; the t =σ(W o x t +U o h t-1 +b o ); In the formula: o t represents the output gate, W o represents the output gate weight matrix, U o represents the output gate weight matrix of the previous time step, h t-1 is the hidden state of the previous time step; b o represents the output gate bias term; c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c ); Where: c t Indicates the cell state, c t-1 is the unit state of the previous unit step; W c The weight matrix representing the unit state, U c represents the unit state weight matrix of the previous time step, b c Represents the unit state bias term; h t =o t ⊙tanh(c t ); Where: h t represents the hidden state and ⊙ represents element-wise multiplication.

6. The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive according to claim 5 is characterized by: In step 3, the multimodal data obtained by variational mode decomposition (VMD) and singular spectrum analysis (SSA) are predicted respectively, and the modal data are input one by one into the hybrid data-driven model constructed by TCN and SWLSTM for prediction, and the prediction results of each modal data are obtained, which are expressed as: In the formula, represents the predicted value of the mth mode; Mode represents the hybrid data-driven model; u m (t) represents the mth mode.

7. The short-term wind power forecasting method based on multi-scale dominant component analysis and hybrid data drive according to claim 1 is characterized by: In step 4, the prediction results of all modal data are reconstructed to obtain the final wind power prediction value. The reconstruction formula is as follows: In the formula, is the final wind power forecast value, is the predicted value of the mth mode, and K is the total number of modes.

Images