Wind power multi-step prediction system based on SVMD algorithm and DIATE loss function

Abstract

The invention relates to the field of wind power multi-step prediction, and discloses a wind power multi-step prediction system based on an SVMD algorithm and a DIATE loss function, comprising: (1) missing value interpolation: interpolating missing values in an original wind power data set by using an artificial neural network; (2) data decomposition: adaptively decomposing original wind power data based on an SVMD algorithm; (3) constructing a prediction model: training an LSTM prediction model by using a DISATE loss function supporting shape and time change detection; and (4) model optimization: optimizing trainable parameters of the LSTM-DIATE prediction model through an Adam optimization algorithm. According to the method, the SVMD algorithm is utilized to realize self-adaptive decomposition of the original wind power data, so that the negative effect of subjectively setting the number of modes on the prediction result is avoided, and meanwhile, the DIATE loss function covering the shape and the time change detection standard is taken as the training target of the prediction model; the method improves the capability of a prediction model to timely and accurately capture sharp changes in wind power data, and has a high practical application value.

Description

Technical Field

[0001] This invention relates to the field of multi-step wind power prediction, specifically a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function. Background Technology

[0002] The excessive use of fossil fuels in industrial production has resulted in the emission of large amounts of greenhouse gases into the atmosphere, triggering a series of environmental problems such as global warming and natural disasters. Therefore, under the dual pressures of a conventional energy shortage and global ecological degradation, renewable energy is receiving increasing attention and importance from countries around the world. Among these, wind power, as a practical, environmentally friendly, and safe new energy source, has enormous development potential in the electricity market.

[0003] However, the intermittent, fluctuating, and uncertain characteristics of wind power pose significant challenges to power system dispatch. A key solution to this problem is multi-step wind power prediction. By predicting wind power in multiple steps, dispatching departments can anticipate trends in wind power changes, thereby rationally arranging operational modes and response measures to improve the economy, safety, and reliability of the power grid. In previous studies on multi-step wind power prediction, the data decomposition algorithms used mostly required subjective setting of the number of modes, which to some extent reduced the accuracy of multi-step wind power prediction. Furthermore, most current multi-step wind power prediction systems use mean squared error for model training, making them lack the ability to capture sharp changes in non-stationary wind power data in a timely and accurate manner. Based on these considerations, this invention comprehensively considers the accuracy, shape distortion, and time delay of multi-step wind power prediction, proposing a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function. This system solves the problem that most current multi-step wind power prediction systems use mean squared error for model training, which makes them lack the ability to capture various sharp changes in non-stationary wind power data in a timely and accurate manner.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function, comprising the following steps:

[0006] S100: Obtain the wind power dataset of the wind farm;

[0007] S200: Uses artificial neural networks to interpolate missing values ​​in the original wind power dataset, thereby avoiding a decrease in the accuracy of model predictions due to missing original data;

[0008] S300: The original wind power data is adaptively decomposed into a series of sub-modes of different frequencies through the SVMD algorithm. The SVMD algorithm uses parameter optimization method instead of empirical parameter method to eliminate the impact of subjective selection of the number of modes on prediction performance during the decomposition process.

[0009] S400: Set the DILATE loss function, which explicitly supports the detection of shape and time changes, as the training objective of the LSTM prediction model, thereby improving the ability of the LSTM prediction model to capture sharp changes in non-stationary wind power data in a timely and accurate manner.

[0010] S500: Optimizes the trainable parameters of the LSTM-DILATE prediction model based on the Adam optimization algorithm;

[0011] S600: Multiple trained LSTM-DILATE models are used to make multi-step predictions for each sub-mode, and the prediction values ​​of each sub-mode are accumulated to obtain the final multi-step prediction result of wind power.

[0012] Preferably, in step S200, the internal calculation process of the neurons in the artificial neural network specifically adopts the following formula:

[0013]

[0014] Among them, w i It is the weight of the i-th input signal, x i is the value of the i-th input signal, and b is the bias term of the neuron. It is the activation function, and Opt is the output of the neuron.

[0015] Preferably, in step S300, the SVMD algorithm performs signal decomposition using the following calculation formula:

[0016] f(t) = u L (t)+f r (t) (2)

[0017]

[0018] Where f(t) is the input signal, u L (t) is the Lth mode, f r (t) is the residual signal. It is the sum of previously obtained modes, f u (t) represents the unprocessed portion of the signal;

[0019] The constraint criteria for the SVMD algorithm specifically adopt the following formula:

[0020]

[0021] Where δ(t) is the impulse function, j is the imaginary number, and w L The center frequency. It is the square of the L2 norm, and * represents the convolution operation. For each mode function u L The analytic signal (t) is subjected to Hilbert transform, and the single-sided frequency is finally obtained. Constraint criterion J1 ensures that each mode is compactly around its center frequency in the frequency domain. and These are the center frequencies w of the Lth mode, respectively. L And the center frequency w of the previously extracted i-th mode i The frequency response of the designed filter, α is a regularization parameter, β L (t) and β i (t) is the impulse response of the filter, and constraint criterion J2 ensures that the residual signal f is within the effective frequency range of the current mode. r Minimize the energy of (t), and constrain criterion J3 to avoid the overlap between the Lth mode obtained by the algorithm decomposition and the previous L-1 modes;

[0022] To ensure that the input signal can be completely reconstructed, the SVMD algorithm specifically adopts the following formula:

[0023]

[0024] Preferably, in step S400, the cell state update of the LSTM model specifically adopts the following formula:

[0025]

[0026] in, It is the sigmoid activation function, w forgetting The weights h are the forget gate weights in the LSTM model. t-1 It is the hidden state at time t-1, x t It is the input at time t, b forgetting It is the bias of the forgetting gate in the LSTM model. t This is the output of the forget gate in the LSTM model. Input These are the weights of the input gate of the LSTM model, b Input It is the bias of the input gate of the LSTM model. t This is the calculation result of the input gate of the LSTM model. τ(·) is the tanh activation function, and w C b is a weighting factor when calculating candidate values ​​for memory cells. C It is the bias used when calculating candidate values ​​for memory cells. C represents the candidate values ​​for memory cells at time t. t-1C is the value of the memory cells at time t-1. t It is the memory cell value at time t;

[0027] The output of the LSTM model is specifically expressed using the following formula:

[0028]

[0029] h t =Output t ·τ(C t (16)

[0030] Among them, w Output These are the weights of the output gates of the LSTM model, b Output It is the bias of the output gate of the LSTM model. t This is the calculation result of the output gate of the LSTM model. h t This is the output of the LSTM model;

[0031] The total error of the DILATE loss function is specifically expressed by the following formula:

[0032]

[0033]

[0034] Where, y = (y 1 y 2 , ..., y m ) is the true value of the time series in the next m steps. It is the predicted value of the time series in the next m steps, and φ(·) is the Euclidean distance. It is y and The cost matrix between them. It is a matrix with {0,1} elements. It is y and Dynamic time regularization. It is a smoothing operator for the min function. It is dynamic time warping The differentiable form, namely soft-DTW, This is the shape error predicted by the model. (B) best This is the optimal path. It is B best A smooth approximation. It is an m×m matrix used to penalize the path B in the optimal path. best p≠q of y p and The time error of the model prediction is θ, and the balance factor is θ. It is the total error of the model prediction calculated by the DILATE loss function.

[0035] Preferably, in step S500, the Adam optimization algorithm updates the trainable parameters of the prediction model using the following formula:

[0036]

[0037] q t =δ1q t-1 +(1-δ1)ψ t (26)

[0038]

[0039] Where, ψ t It is the gradient of the parameters, p t It is the exponentially weighted average of the squared gradients, q t It is the exponentially weighted average of the gradient, and δ1 and δ2 are the two exponentially weighted averages q. t and p t The attenuation coefficient, Here, η is the parameter at time t, η is the learning rate, and μ is a very small constant. These are the parameters at time t+1, i.e., the updated parameters.

[0040] This invention provides a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function. It has the following advantages:

[0041] 1. The wind power multi-step prediction system developed in this invention utilizes the DILATE loss function, which explicitly supports the detection of shape and time changes, to train the prediction model LSTM, thereby improving the ability of the prediction model LSTM to capture step changes in non-stationary wind power data in a timely and accurate manner. Furthermore, since some missing values ​​often exist in the original wind power data, this invention uses an artificial neural network to impute missing values ​​to improve the quality of the original data. Simultaneously, it achieves adaptive decomposition of wind power data based on the SVMD algorithm, which avoids the negative impact on prediction performance caused by empirical judgments about the number of modes. Attached Figure Description

[0042] Figure 1 A flowchart of a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function is provided for this invention.

[0043] Figure 2 This is a schematic diagram of the framework of the wind power multi-step prediction system provided in an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the original wind power data provided in an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of the decomposition results of the SVMD algorithm provided in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram illustrating the operation process of the LSTM model provided in an embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram illustrating the error convergence of the LSTM-DILATE model on the training set provided in an embodiment of the present invention.

[0048] Figure 7 This diagram illustrates the prediction results and prediction errors of the wind power multi-step prediction system proposed in this invention on a test set, as provided in an embodiment of the invention. Detailed Implementation

[0049] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see the appendix Figure 1 - Appendix Figure 7 This invention provides a multi-step wind power prediction system based on the SVMD algorithm and the DILATE loss function, comprising:

[0051] S100: Obtain the wind power dataset of the wind farm.

[0052] The research data comes from historical records of the Sotavento wind farm in Spain, using wind power data from April 1st to April 30th, 2024 for the experiment. The temporal resolution of the data observation is 10 minutes, with a total of 4320 observation points, including 10 missing values. The research data is divided in a 7:3 ratio, with the first 70% of the data used as the training set to train the LSTM-DILATE prediction model for each modality, and the last 30% of the data used as the test set to verify the performance of the proposed multi-step wind power prediction system. Figure 3 A schematic diagram of the original wind power data provided for an embodiment of the present invention.

[0053] S200: Uses artificial neural networks to interpolate missing values ​​in the original wind power dataset, thereby avoiding a decrease in the accuracy of model predictions due to missing original data.

[0054] The artificial neural network in the missing value imputation module has 20 hidden layers and 10 neurons in each hidden layer. The training objective of the model is the Mean Squared Error loss function. Missing values ​​in the original wind power dataset are imputed according to formula (1).

[0055] S300: The original wind power data is adaptively decomposed into a series of sub-modes of different frequencies using the SVMD algorithm. The SVMD algorithm uses parameter optimization methods instead of empirical parameter methods to eliminate the impact of subjective selection of the number of modes on prediction performance during the decomposition process.

[0056] The sequence lengths of the training and test data are 3024 and 1296, respectively. The constraint criteria and the balance parameter are 200. The time step used to update the dual variable in the dual ascent method is 0. The energy of the last mode is used as the stopping criterion of the SVMD algorithm. The tolerance of the convergence criterion is 1e-6. According to formulas (2) to (10), the original wind power data can be adaptively decomposed using the SVMD algorithm. Figure 4 This is a schematic diagram of the decomposition results of the SVMD algorithm provided in an embodiment of the present invention.

[0057] S400: Set the DILATE loss function, which explicitly supports the detection of shape and time changes, as the training objective of the LSTM prediction model, thereby improving the LSTM prediction model's ability to capture sharp changes in non-stationary wind power data in a timely and accurate manner.

[0058] The time step of the input sequence of the prediction model LSTM is 6, the time step of the output sequence is 3, the batch size is 100, the feature dimension of the input sequence is 1, the number of stacked LSTM layers is 1, the dimension of the hidden state of the LSTM unit is 64, and the number of training iterations is 500. The smoothing operator of the DILATE loss function is 0.1, and the balance factor between shape loss and time loss is 0.5. According to formulas (11) to (24), the prediction models of each submodality can be trained. Figure 5 This is a schematic diagram illustrating the operation process of the LSTM model provided in an embodiment of the present invention.

[0059] S500: Optimizes the trainable parameters of the LSTM-DILATE prediction model based on the Adam optimization algorithm.

[0060] The decay coefficient of the exponentially weighted average of the gradient is 0.9, the decay coefficient of the exponentially weighted average of the squared gradient is 0.999, the constant is 10⁻⁸, and the learning rate is 0.001. The trainable parameters of the LSTM-DILATE prediction model can be updated according to formulas (25), (26), and (27). Figure 6 This is a schematic diagram illustrating the error convergence of the LSTM-DILATE model on the training set, as provided in an embodiment of the present invention.

[0061] S600: Multiple trained LSTM-DILATE models are used to make multi-step predictions for each sub-mode, and the prediction values ​​of each sub-mode are accumulated to obtain the final multi-step prediction result of wind power.

[0062] On the test set, multiple trained LSTM-DILATE models are used to perform three-step predictions for each sub-mode. The predicted values ​​of each sub-mode are summed to obtain the final multi-step wind power prediction result and prediction error, so as to verify the performance of the wind power multi-step prediction system proposed in this invention. Figure 7 This diagram illustrates the prediction results and prediction errors of the wind power multi-step prediction system proposed in this invention on a test set, as provided in an embodiment of the invention.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind power multi-step prediction system based on SVMD algorithm and DILATE loss function, characterized in that: include: S100: Acquire a wind power data set of a wind farm; S200: Using artificial neural networks to interpolate missing values ​​in the original wind power data set, thereby avoiding the reduction of model prediction accuracy due to missing original data; S300: The original wind power data is adaptively decomposed into a series of sub-modes with different frequencies through the SVMD algorithm. The SVMD algorithm uses parameter optimization method instead of empirical parameter method to eliminate the influence of subjective selection of the number of modes in the decomposition process on the prediction performance; S400: Set the DILATE loss function that explicitly supports shape and time change detection as the training objective of the LSTM prediction model to improve the ability of the LSTM prediction model to capture sharp changes in non-stationary wind power data in a timely and accurate manner; S500: Optimizing the trainable parameters of the LSTM-DILATE prediction model based on the Adam optimization algorithm; S600: Use the trained multiple LSTM-DILATE models to perform future multi-step predictions for each sub-mode respectively, and accumulate the predicted values ​​of each sub-mode to obtain the final wind power multi-step prediction result.

2. The wind power multi-step prediction system based on the SVMD algorithm and the DILATE loss function according to claim 1 is characterized in that: The internal calculation process of the neurons of the artificial neural network in S200 specifically adopts the following formula: Among them, w i is the weight of the ith input signal, x i is the value of the i-th input signal, b is the bias term of the neuron, is the activation function, and Opt is the output of the neuron.

3. The wind power multi-step prediction system based on the SVMD algorithm and the DILATE loss function according to claim 1 is characterized in that: The SVMD algorithm described in S300 specifically uses the following formula to perform signal decomposition: f(t)=u L (t)+f r (t) Where f(t) is the input signal, u L (t) is the Lth mode, is the residual signal, is the sum of the previously obtained modes, f u (t) is the unprocessed part of the signal; The constraint criterion 1 of the SVMD algorithm described in S300 specifically adopts the following formula: in, is the square of the L2 norm, w L is the center frequency, * is the convolution operation, For each modal function u L The analytical signal of (t) is Hilbert transformed, and finally the single-sided frequency is obtained, is the impulse function, j is an imaginary number, and the constraint criterion J1 ensures that each mode is compact around its center frequency in the frequency domain; The constraint criterion 2 of the SVMD algorithm described in S300 specifically adopts the following formula: Among them, β L (t) is the frequency response The impulse response of the filter, α is a regularization parameter, and the constraint criterion J2 ensures that the residual signal f is within the effective frequency range of the current mode. r (t) energy minimization; The constraint criterion 3 of the SVMD algorithm described in S300 specifically adopts the following formula: Among them, β i (t) is the center frequency w of the previously extracted i-th mode i The impulse response of the designed filter, filter β i The frequency response of (t) is The constraint criterion J3 prevents the Lth mode obtained by the algorithm decomposition from overlapping with the previous L-1 modes; The SVMD algorithm described in S300 specifically adopts the following formula to ensure that the input signal can be completely reconstructed:

4. The wind power multi-step prediction system based on the SVMD algorithm and the DILATE loss function according to claim 1 is characterized in that: The output of the forget gate of the LSTM model described in S400 specifically adopts the following formula: in, is the sigmoid activation function, w forgetting is the weight of the forget gate of the LSTM model, h t-1 is the hidden state at time t-1, x t is the input at time t, b forgetting is the bias of the forget gate of the LSTM model; The memory cell candidate value of the LSTM model described in S400 specifically adopts the following formula: in, is the tanh activation function, w C is the weight factor when calculating the candidate value of the memory cell, b C is the bias when calculating candidate values ​​for memory cells; The input gate of the LSTM model described in S400 specifically adopts the following formula: Among them, Input t is the calculation result of the LSTM model input gate, w Input is the weight of the input gate of the LSTM model, b Input is the bias of the input gate of the LSTM model; The cell state update of the LSTM model described in S400 specifically adopts the following formula: Among them, C t is the memory cell value at time t, forgetting t is the output result of the forget gate of the LSTM model, C t-1 is the memory cell value at time t-1, is the candidate value of the memory cell at time t; The output of the LSTM model described in S400 is specifically calculated using the following formula: Among them, h t is the output of the LSTM model. is the calculation result of the output gate of the LSTM model, w Output is the weight of the output gate of the LSTM model, b Output is the bias of the output gate of the LSTM model; The shape error of the DILATE loss function described in S400 is specifically calculated using the following formula: in, Dynamic Time Warping The differentiable form of is the smoothing operator of the min function, y=(y 1 ,y 2 ,…,y m ) is the true value of the time series for the next m steps, is the predicted value of the prediction model for the time series of m steps in the future, is y and The cost matrix between , φ(·) is the Euclidean distance, is a matrix of {0,1} elements; The time error of the DILATE loss function described in S400 is specifically calculated using the following formula: in, Is the best path A smooth approximation of is an m×m matrix, which is used to penalize the optimal path B best y on p≠q p and The total error of the DILATE loss function described in S400 is specifically calculated using the following formula: in, is the shape error predicted by the model, is the time error predicted by the model and θ is the balancing factor.

5. The wind power multi-step prediction system based on the SVMD algorithm and the DILATE loss function according to claim 1, characterized in that: The Adam optimization algorithm described in S500 updates the trainable parameters of the prediction model using the following calculation formula: in, is the parameter at time t+1, is the parameter at time t, η is the learning rate, and μ is a small constant. is the gradient ψ t The exponentially weighted average of the squares of t =δ1q t-1 +(1-δ1)ψ t is the gradient ψ t The exponentially weighted average of , δ1 and δ2 are two exponentially weighted averages q t and p t The attenuation coefficient.

Images