Offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction

Abstract

The application discloses a kind of offshore wind power prediction methods based on wind speed-power combination decomposition reconstruction, comprising: based on combination decomposition method, offshore wind power sequence is carried out stationary treatment;According to the complexity and similarity characteristics of component, the data after decomposition is reconstructed dimension reduction;According to the coupling correlation between offshore wind speed and power, power is decomposed and combined with LSTM to establish prediction submodel;According to the network structure characteristics of deep learning, the hyperparameters are optimized using Bayesian algorithm, and BO-LSTM model is constructed, and the final prediction value is obtained by superimposing the prediction results of each BO-LSTM submodel.The method of the application can improve the problems of strong data volatility, incomplete decomposition and large data size caused by decomposition in offshore wind power prediction, enhance the ability of neural network to explore the correlation between wind speed data and power data, and improve the accuracy of offshore wind power prediction.

Description

Technical Field

[0001] This invention belongs to the field of offshore wind power prediction, specifically involving an offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction and Bayesian optimization. Background Technology

[0002] Against the backdrop of vigorously developing renewable and clean energy and striving to achieve the goals of "carbon peaking and carbon neutrality," my country's offshore wind power industry has developed rapidly due to its advantages such as abundant resources, high power generation efficiency, and proximity to load centers. However, due to the randomness and volatility of offshore wind power output, large-scale offshore wind power grid connection poses a severe challenge to the safe operation of the power grid in my country's coastal provinces. Accurately predicting offshore wind power output can provide an important basis for flexibly allocating system reserve capacity, ensuring the safety and stability of the power system, and improving the absorption capacity of offshore wind power. With the rapid development of machine learning, statistical models combined with machine learning prediction methods are widely used in new energy power prediction. For example, methods such as neural networks, support vector machines, and deep learning can explore the feature relationships between input and output, and have faster calculation speeds and higher accuracy. Although machine learning algorithms can effectively improve the accuracy of new energy power generation prediction, due to the randomness and volatility of offshore wind power output, traditional machine learning prediction models struggle to capture the feature information of offshore wind power data, resulting in poor power prediction accuracy and failing to achieve ideal results.

[0003] Therefore, a new technical solution is needed to solve this problem. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of poor power prediction accuracy in the existing technology, this invention provides an offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction and Bayesian optimization. This method can improve the problems of strong data volatility, incomplete decomposition and large data scale caused by decomposition in offshore wind power prediction, enhance the ability of neural networks to discover the correlation between wind speed data and power data, and improve the accuracy of offshore wind power prediction.

[0005] Technical solution: This invention provides a method for predicting offshore wind power based on wind speed-power combination decomposition and reconstruction, comprising the following steps:

[0006] S1: Wind speed data is stabilized based on the combinatorial decomposition method;

[0007] S2: Based on the complexity and similarity characteristics of the components, the data decomposed in step S1 is reconstructed and dimensionality reduced using fuzzy entropy FE.

[0008] S3: Based on the coupling correlation between sea wind speed and power, historical power is decomposed and a Long Short-Term Memory (LSTM) prediction sub-model is established. Wind speed reconstruction data and power data at a certain time scale are selected as inputs to the prediction sub-model.

[0009] The time scale of the input data is determined by analyzing the Pearson correlation coefficient between the unit output at time t and the historical wind speed and power data of the previous n times, that is, using historical data of the previous n times for prediction.

[0010] S4: Based on the characteristics of deep learning network structure, the Bayesian algorithm (BO) is used to optimize the hyperparameters of the LSTM to construct the BO-LSTM model. The prediction results of each BO-LSTM sub-model are then superimposed to obtain the final prediction value.

[0011] Furthermore, the method for combining and decomposing the wind speed data in step S1 is as follows:

[0012] The wind speed sequence is decomposed using ICEEMDAN, which includes defining the wind speed sequence as v and the k-th modal component of the EMD decomposition as E. k (·), add noise to the wind speed sequence, and calculate the residual R of the first decomposition. es1 And the first-order intrinsic mode component D1.

[0013] D1 = vR es1 (1)

[0014]

[0015] In the formula, ε is the signal-to-noise ratio of white noise, and w (i) Let M be the number of Gaussian white noise groups added, representing the i-th group. Repeat the steps to add more white noise and calculate Ri for the second decomposition. es2 And D2. Continue calculating the residual of the kth decomposition and the kth modal component until no further decomposition can be performed.

[0016] To address the strong fluctuations in the high-frequency component IMF1 after the first decomposition, the VMD method is used to further decompose IMF1. Specifically, this involves defining the intrinsic mode functions (IMFs) with finite bandwidth. k The constrained variational problem is constructed as shown in formula (4), and the modal components are obtained by solving the variational problem.

[0017] IMF k (t)=A k (t)cos(Φ(t)) (3)

[0018]

[0019] In the formula: k is the number of components after decomposition, A kω(t) and φ(t) are the amplitude and phase, respectively. k The center frequencies of each modal component are... These are the gradient and convolution operators, respectively. δ(t) is the Dirac function, and X(t) is the sequence to be decomposed.

[0020] Solving equation (4), a penalty factor α is introduced to transform the constrained variational problem into an unconstrained variational problem to reduce Gaussian noise interference. The expression is as follows:

[0021]

[0022] In the formula: α is the penalty factor, and λ is the Lagrange operator.

[0023] The SSA algorithm is used to optimize VMD decomposition parameters to address the problem that VMD cannot adaptively select the number of components k and the penalty coefficient α, avoiding the influence of subjective factors from manually setting parameters and improving decomposition quality. Specifically, this includes:

[0024] A1: Initialize the SSA population size, set the initial [K,α] and the maximum number of iterations.

[0025] A2: Update the location of the finder in the sparrow population. The finder is the member with the best fitness in the population, responsible for foraging. The formula for updating its location is:

[0026]

[0027] In the formula, Iter represents the position of the i-th discoverer in dimension d. max R represents the maximum number of iterations; R2 is the warning value, ranging from [0,1]; S represents the safety value, ranging from [0.5,1]; Q is a random number following a normal distribution; and L is a 1×d matrix with all elements equal to 1.

[0028] A3: The newcomer constantly observes the discoverer's behavior and adjusts its own position accordingly. The position update formula is:

[0029]

[0030] In the formula, x worst Represents the worst current position. Represents the current global optimal position; A is a matrix whose elements are all 1 or -1 and A + =A T (A T ) -1 .

[0031] A4: In each iteration, a certain number of sparrows in the population will act as watchdogs. Upon detecting danger, they will abandon foraging and move to a safe area to continue foraging. The position update formula is:

[0032]

[0033] In the formula, x best f represents the current global optimal position; β is a random number following a normal distribution; i f represents the current fitness. g denoted as the optimal fitness in the population, I is a random value taken within the range [0,1], and μ is a constant.

[0034] A5: Repeat the iteration until the optimal value of [K,α] is obtained.

[0035] Furthermore, the method for reconstructing and reducing the dimensionality of the decomposed data based on FE in step S2 is as follows: analyze the FE value of each component, group them according to the complexity and similarity characteristics of each component, and reconstruct components of the same component into one component. The specific steps are as follows:

[0036] B1: Let the subsequence obtained after the combination decomposition of the sea wind speed data be x1, x2, ..., x n X is obtained by reconstructing the phase space of the subsequence. i ;

[0037]

[0038] In the formula, N is x i The length of ; m represents the embedding dimension, which is usually taken as 2; This represents the mean of m adjacent sequences;

[0039] B2: Define the fuzzy membership function Calculate each sequence X i fuzzy entropy;

[0040]

[0041]

[0042]

[0043] In the formula, This represents the maximum value of the difference between the corresponding endpoints of two sequences X(i) and X(j); r is the noise margin, which is 0.15 times the standard deviation of the processed data.

[0044] B3: The FE of a subsequence is a dimensionless number taking values ​​in the range [0,1]. The larger the value, the higher the complexity of the sequence and the greater the probability of generating a new mode. Based on the fuzzy entropy value of each subsequence, they are divided into 3 groups, and each group of subsequences is reconstructed into a component.

[0045] The method for SSA-VMD decomposition and modeling of power in step S3 is as follows: Similar to step S1, where SSA-VMD is used to decompose the high-frequency component IMF1 after the first decomposition, the historical power sequence is decomposed using SSA-VMD, and a prediction sub-model is established for each power component. Simultaneously, the Pearson correlation coefficient between offshore wind power output and the historical wind speed and historical power of the previous n times is analyzed to determine the time scale of the model input.

[0046] Furthermore, the method for establishing an LSTM prediction sub-model for the power component in step S3 is as follows:

[0047] An LSTM network is constructed, with each LSTM layer consisting of multiple LSTM units. Each LSTM unit has three gates: a forget gate, an input gate, and an output gate. The forget gate determines the current time c. t Save the cell state c from the previous time step t-1 The number of input gates determines the cell state X. t At the current time c t The number of network inputs, and the output gate control output to the current output value h of the LSTM. t cell state c t The quantity is expressed by the following formula:

[0048] f t =σ(W f ·(h t-1 ,x t )+b f (13)

[0049] i t =σ(W i ·(h t-1 ,x t )+b i (14)

[0050] o t =σ(W o ·(h t-1 ,x t )+b o (15)

[0051] c′ t =tanh(W c ·(h t-1 ,x t )+b c (16)

[0052] c t =f t ×c t-1 +i t ×c′t (17)

[0053] h t =o t ×tanh(c t (18)

[0054] Among them, f t i t o t Information for the forget gate, input gate, and output gate, respectively; W f W i W o Here is the weight matrix; x t and h t-1 These are the input at time t and the hidden layer state at time t-1, respectively; b f b i , b o is the bias; σ is the sigmoid function, c' t Input the information of the memory module at time t; W c is the weight matrix; tanh is the hyperbolic tangent function.

[0055] Furthermore, the method for optimizing the hyperparameters of the LSTM network using the Bayesian algorithm in step S4 is as follows:

[0056] By evaluating two sets of hyperparameters, a sequentially updated probabilistic model is established to obtain the prior probabilities of the optimization problem, and then the optimal solution satisfying the objective function f(z) is sought. The objective function formula is as follows:

[0057]

[0058] in The optimal value of Z is the range for hyperparameter optimization.

[0059] Based on the above scheme, the overall process of the present invention can be summarized as follows: First, the offshore wind power data is combined and decomposed using signal decomposition methods to improve the impact of the randomness of the original signal and the incomplete decomposition on the prediction accuracy. Then, fuzzy entropy analysis is used to reconstruct the modal components, reducing the dimensionality of the input data to the prediction model. Finally, a prediction sub-model is established for each power component. The reconstructed wind speed data and historical power data are input into a Bayesian-optimized long short-term memory neural network, and the prediction results of each sub-model are superimposed to obtain the final power prediction value.

[0060] The following is a comprehensive description of the technical aspects involved in this invention:

[0061] Combinatorial decomposition of offshore wind speed data: To address the non-stationarity of offshore wind power data and the impact of the decomposed high-frequency mode components on power prediction accuracy, ICEEMDAN and VMD are used to perform combined decomposition of offshore wind speed data, which stabilizes the data. The SSA algorithm is then used to optimize VMD to solve the problem that VMD cannot adaptively select decomposition parameters.

[0062] Data Reconstruction Based on FE: After two decompositions, the raw offshore wind power data has too many components, leading to excessively high dimensionality and making direct processing extremely complex. This invention analyzes the complexity of each sub-component using the FE method, reconstructing subsequences with correlated FE characteristics by grouping them, thereby reducing data dimensionality and improving computational speed.

[0063] Historical Power Decomposition Modeling using SSA-VMD: In the prediction modeling stage, offshore wind power data exhibits strong temporal correlation, and the non-stationarity of the power data significantly impacts prediction accuracy. Therefore, SSA-VMD is used to decompose the historical offshore wind power sequence, establishing a prediction sub-model for each power component to reduce the interference of historical power sequence volatility on prediction accuracy. This invention selects wind speed reconstruction signals and historical power signals at a certain time scale as inputs to the prediction model, improving the LSTM's ability to learn the correlation and temporal characteristics between wind speed data and historical power data.

[0064] BO-LSTM prediction model: The hyperparameter settings of neural networks have a significant impact on model performance. Since the selection of hyperparameters is very difficult, the BO algorithm, which has stronger generalization ability and faster convergence speed, is chosen to optimize the hyperparameters of the LSTM network.

[0065] Beneficial effects: Compared with the prior art, this invention proposes an offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction, which effectively reduces the impact of historical data volatility on power prediction, improves the problems of incomplete decomposition, mode mixing and large data scale after multiple decompositions in the signal decomposition process, and enhances the performance of LSTM network, enabling the neural network to better discover the correlation feature information between offshore wind speed sequence and historical power sequence, thus effectively improving the accuracy of offshore wind power prediction. Attached Figure Description

[0066] Figure 1 A framework diagram of an offshore wind power prediction model;

[0067] Figure 2 Here is a diagram of the LSTM network structure;

[0068] Figure 3 This is a display diagram of the ICEEMDAN decomposition of the wind speed sequence;

[0069] Figure 4 A diagram showing the iterative evolution of the SSA algorithm;

[0070] Figure 5 The SSA-VMD results for the high-frequency component IMF1 of wind speed are shown in the figure.

[0071] Figure 6 A graph showing the SSA-VMD results for historical power sequences;

[0072] Figure 7 The results of the component FE analysis are shown in the figure.

[0073] Figure 8 The result of component reconstruction is shown in the figure;

[0074] Figure 9 A graph showing the RMSE error of different models for predicting the four seasons.

[0075] Figure 10 The graph shows the seasonal prediction results of the decomposition model. Detailed Implementation

[0076] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0077] This invention provides a method for predicting offshore wind power based on wind speed-power combination decomposition and reconstruction, referring to... Figure 1 It includes the following steps:

[0078] S1: Wind speed data is stabilized based on the combinatorial decomposition method;

[0079] Decomposing the original signal to reduce the impact of volatility and randomness on prediction accuracy is an effective method to improve the prediction accuracy of new energy power generation. In this step, ICEEMDAN is first used to decompose the wind speed sequence for offshore wind power. Specifically, the wind speed sequence is defined as v, and the k-th order mode component of the EMD decomposition is E. k (·), add noise to the wind speed sequence, and calculate the residual R of the first decomposition. es1 And the first-order intrinsic mode component D1.

[0080] D1 = vR es1 (1)

[0081]

[0082] In the formula, ε is the signal-to-noise ratio of white noise, and w (i)Let M be the number of Gaussian white noise groups added, representing the i-th group. Repeat the steps to add more white noise and calculate Ri for the second decomposition. es2 And D2. Continue calculating the residual of the kth decomposition and the kth modal component until no further decomposition is possible.

[0083] The VMD method is used to further decompose the high-frequency modal components IMF1 after the first decomposition. Specifically, this includes defining the intrinsic mode functions (IMFs) with finite bandwidth. k The constrained variational problem is constructed as shown in formula (4), and the modal components are obtained by solving the variational problem.

[0084] IMF k (t)=A k (t)cos(Φ(t)) (3)

[0085]

[0086] In the formula: k is the number of components after decomposition, A k ω(t) and φ(t) are the amplitude and phase, respectively. k The center frequencies of each modal component are... * represents the gradient and convolution operators, respectively; δ(t) is the Dirac function; and X(t) is the sequence to be decomposed.

[0087] Solving equation (4), a penalty factor α is introduced to transform the constrained variational problem into an unconstrained variational problem to reduce Gaussian noise interference. The expression is as follows:

[0088]

[0089] In the formula: α is the penalty factor, and λ is the Lagrange operator.

[0090] The SSA algorithm is used to optimize VMD decomposition parameters to address the problem that VMD cannot adaptively select the number of components k and the penalty coefficient α, avoiding the influence of subjective factors from manually setting parameters and improving decomposition quality. Specifically, this includes:

[0091] (1) Initialize the SSA population size, set the initial [K,α] and the maximum number of iterations.

[0092] (2) Update the position of the finder in the sparrow population. The finder is the member with the best fitness in the population and is responsible for foraging. The formula for updating its position is:

[0093]

[0094] In the formula, Iter represents the position of the i-th discoverer in dimension d. maxR represents the maximum number of iterations; R2 is the warning value, ranging from [0,1]; S represents the safety value, ranging from [0.5,1]; Q is a random number following a normal distribution; and L is a 1×d matrix with all elements equal to 1.

[0095] (3) The newcomer constantly observes the discoverer's behavior and adjusts its own position accordingly. The position update formula is:

[0096]

[0097] In the formula, x worst Represents the worst current position. Represents the current global optimal position; A is a matrix whose elements are all 1 or -1 and A + =A T (A T ) -1 .

[0098] (4) In each iteration, a certain number of sparrows in the population will act as watchdogs. Upon detecting danger, they will abandon foraging and move to a safe area to continue foraging. The position update formula is:

[0099]

[0100] In the formula, x best f represents the current global optimal position; β is a random number following a normal distribution; i f represents the current fitness. g denoted as the optimal fitness in the population, I is a random value taken within the range [0,1], and μ is a constant.

[0101] (5) Repeat the iteration until the optimal value of [K,α] is obtained.

[0102] S2: Based on the complexity and similarity characteristics of the components, the decomposed data is reconstructed and its dimensionality reduced using FE: The original offshore wind power data, after two decompositions, has too many components, leading to excessively high dimensionality of the input data. The fuzzy entropy of each component is analyzed, and the components are grouped according to their complexity and similarity characteristics. Components of the same type are reconstructed into a single component, specifically including:

[0103] (1) Let the subsequence obtained after the combination decomposition of the sea wind speed data be x1, x2, ..., x n X is obtained by reconstructing the phase space of the subsequence. i .

[0104]

[0105] In the formula, N is x i The length of ; m represents the embedding dimension, which is usually taken as 2; This represents the mean of m adjacent sequences.

[0106] (2) Define the fuzzy membership function Calculate each sequence X i The fuzzy entropy.

[0107]

[0108]

[0109]

[0110] In the formula, represents the maximum value of the difference between the corresponding endpoints of two sequences X(i) and X(j); r is the noise tolerance, which is 0.15 times the standard deviation of the processed data.

[0111] (3) The FE of a subsequence is a dimensionless number with a value in [0,1]. The larger the value, the higher the complexity of the sequence and the greater the probability of generating a new mode. Based on the fuzzy entropy value of each subsequence, they are divided into 3 groups, and each group of subsequences is reconstructed into a component.

[0112] S3: Perform SSA-VMD decomposition on the power and establish an LSTM prediction sub-model. Select wind speed reconstruction data and power data at a certain time scale as input to the sub-model. Similar to step S1, use SSA-VMD to decompose the high-frequency component IMF1 again after the first decomposition. Perform SSA-VMD decomposition on the historical power sequence and establish a prediction sub-model for each power component. At the same time, analyze the Pearson correlation coefficient between offshore wind power output and the historical wind speed and historical power of the previous n times to determine the time scale of the model input.

[0113] The formula for establishing an LSTM prediction sub-model for each power component in this step is as follows:

[0114] f t =σ(W f ·(h t-1 ,x t )+b f (13)

[0115] i t =σ(W i ·(h t-1 ,x t )+b i (14)

[0116] o t =σ(W o ·(h t-1 ,x t )+b o (15)

[0117] c t '=tanh(W c ·(h t-1 ,x t )+b c (16)

[0118] c t =f t ×c t-1 +i t ×c t (17)

[0119] h t =o t ×tanh(c t (18)

[0120] Where f t i t o t Information for the forget gate, input gate, and output gate, respectively; W f W i W o Here is the weight matrix; x t and h t-1 These are the input at time t and the hidden layer state at time t-1, respectively. f b i b o Let c' be the bias, σ be the sigmoid function, and c' be the bias. t W is the information input into the memory module at time t. c Let be the weight matrix, and tanh be the hyperbolic tangent function.

[0121] S4: Use the Bayesian algorithm to optimize the LSTM hyperparameters to construct a BO-LSTM model, and superimpose the prediction results of each BO-LSTM sub-model to obtain the final prediction value. By evaluating two sets of hyperparameters, establish a sequentially updated probabilistic model to obtain the prior probability of the optimization problem, and then find the optimal solution that satisfies the objective function f(z). The objective function formula is as follows:

[0122]

[0123] in The optimal value of Z is the range for hyperparameter optimization.

[0124] Based on the above-mentioned offshore wind power prediction method derived from data decomposition and reconstruction, this embodiment performs simulation analysis, as follows:

[0125] This analysis uses measured data from a 100MW offshore wind farm in the United States in 2018 as an example. The data sampling interval was 10 minutes, with 144 recording points per day. The seasonal climate variations of offshore wind farms have a significant impact on offshore wind power prediction. To verify the model's practicality in different seasons, this example selects data from four typical months: January, April, July, and October, representing winter, spring, summer, and autumn, respectively. For each month, the first 30 or 29 days are used as the training set, and the last day as the validation set. One-step, two-step, and three-step predictions are performed, predicting offshore wind power at 10 minutes, 20 minutes, and 30 minutes, respectively. To eliminate the influence of randomness in deep learning, the average of 30 prediction results is taken.

[0126] 1. Combinatorial decomposition of marine wind speed data

[0127] Taking January's winter data as an example, such as Figure 2 As shown, the original wind speed sequence, after ICEEMDAN decomposition, yields a total of 12 IMF components and one residual component (only IMF1-IMF5 are shown). It can be seen that IMF1 after ICEEMDAN decomposition still exhibits strong volatility.

[0128] The process of SSA algorithm optimizing VMD decomposition parameters is as follows: Figure 3 The optimal penalty coefficient was found to be 205, and the number of components was 4. The result of the IMF1 decomposition after SSA-VMD is as follows: Figure 4 As shown. After combination and decomposition, a total of 16 components are obtained.

[0129] 2. Wind speed component reconstruction based on FE

[0130] Calculate the FE values ​​of the 16 wind speed sequences obtained after the second decomposition (the four iMF sub-components obtained from the IMF1 decomposition are numbered as sequences 13-16, and the results are as follows). Figure 5 As shown in the diagram. Based on the calculation results, sequences with FE values ​​greater than 0.6 are reconstructed into component R1, which has the highest entropy, the most complex fluctuations, and the highest random information content. Sequences with FE values ​​below 0.2 are reconstructed into R2, which has the lowest entropy, indicating smaller fluctuations and representing the basic trend of wind speed changes. Sequences with FE values ​​between 0.2 and 0.6 are reconstructed into R3, whose complexity and information content are between R1 and R2. After FE reconstruction, the data size of wind speed data is reduced from 16 time series to 3.

[0131] 3. Decompose the power using SSA-VMD and establish an LSTM prediction sub-model. Select wind speed reconstruction data and power data at a certain time scale as inputs to the sub-model.

[0132] In the historical power decomposition modeling, the SSA algorithm optimization result is an optimal penalty coefficient of 303 and a number of components of 4. The decomposition result is as follows: Figure 6 As shown.

[0133] Offshore wind power exhibits a time lag between wind speed and power output. The output of an offshore wind turbine at time t is closely correlated with the historical wind speed and power output of the previous n time points. The time-series Pearson correlation of offshore wind power data is shown in Table 1. In this example, historical data from the previous 12 time points with a correlation greater than 0.85 (ρ≥0.85) with the power output at time t are selected as input data for the prediction model.

[0134] Table 1. Time-series Pearson coefficients for offshore wind power data

[0135]

[0136] The LSTM network structure used in this example is as follows: Figure 7 As shown, each LSTM unit contains 3 control gates, and the forget gate determines the current time c. t Save the cell state c from the previous time step t-1 The number of input gates determines the cell state X. t At the current time c t The number of network inputs, and the output gate control output to the current output value h of the LSTM. t cell state c t The quantity.

[0137] 4. Optimize LSTM hyperparameters using Bayesian algorithm to construct BO-LSTM model: The optimization range and results of the Bayesian algorithm parameters are shown in Table 2. Figure 8 The process of Bayesian optimization is presented, in which the model is continuously updated based on the observed points, and the next point with the smallest error is found based on the updated model. After multiple iterations, the LSTM hyperparameters with the smallest prediction error are obtained.

[0138] Table 2 Parameter Optimization Results

[0139]

[0140] To provide a clear and intuitive evaluation of the model's predictive performance, this embodiment selects root mean square error (RMSE), mean absolute error (MAE), and mean relative percentage error (MAPE) for evaluation, as shown in the formula below.

[0141]

[0142]

[0143]

[0144] This invention is compared with the classic model, and the following control group is also set up:

[0145] Model 1: The wind speed sequence is decomposed using a combination of ICEEMDAN and SSA-VMD. The decomposed components are then reconstructed and input into the LSTM model for prediction.

[0146] Model 2: The SSA-VMD method is used to decompose and model the wind power sequence, and the predictions are performed separately and then superimposed to obtain the final result.

[0147] Model 3: The wind speed sequence is decomposed using ICEEMDAN and SSA-VMD combined decomposition and reconstructed using the FE method. At the same time, the historical power is modeled using SSA-VMD decomposition. The results are predicted and superimposed to obtain the final results.

[0148] Model 4: Based on Model 3, Bayesian optimization of the hyperparameters of LSTM is used, which is the method proposed in this invention.

[0149] Tables 3 and 4 compare the results of the method of this invention with those of other methods in seasonal power prediction and multi-step prediction. Figure 9 , 10 As shown:

[0150] Table 3 Comparison of Prediction Error Evaluation for Different Models

[0151]

[0152] Table 4. Multi-step forecasting error index (January)

[0153]

[0154] Therefore, it can be seen that the method of the present invention has higher accuracy in offshore wind power prediction and multi-step prediction in four seasons, and can effectively reduce the impact of historical data fluctuations on prediction accuracy. Moreover, the stronger the data volatility, the more obvious the improvement effect of the method of the present invention.

Claims

1. A method for predicting offshore wind power based on wind speed-power combination decomposition and reconstruction, characterized in that, Includes the following steps: S1: Wind speed data is stabilized based on the combinatorial decomposition method; S2: Based on the complexity and similarity characteristics of the components, the data decomposed in step S1 is reconstructed and dimensionality reduced using fuzzy entropy FE. S3: Based on the coupling correlation between sea wind speed and power, historical power is decomposed and a Long Short-Term Memory (LSTM) prediction sub-model is established. Wind speed reconstruction data and power data at a certain time scale are selected as inputs to the prediction sub-model. S4: Based on the characteristics of deep learning network structure, use the Bayesian algorithm (BO) to optimize the hyperparameters of LSTM to construct a BO-LSTM model, and then superimpose the prediction results of each BO-LSTM sub-model to obtain the final prediction value. The method for combining and decomposing wind speed data in step S1 is as follows: ICEEMDAN (Complete Ensemble Empirical Mode Decomposition) with improved adaptive noise and VMD (Variable Mode Decomposition) are used to combine and decompose the sea wind speed data, and the SSA (Sparrow Search Algorithm) is used to solve for the optimal decomposition parameters of VMD. The method for variable mode decomposition (VMD) in step S1 is as follows: The decomposed high-frequency components are subjected to VMD secondary decomposition. A variational problem is constructed according to the following formula, and the intrinsic mode functions are obtained by solving the variational problem. (3)； (4)； In the formula: The number of components after decomposition. and For amplitude and phase, The center frequencies of each modal component are... , These are the gradient and convolution operators, respectively. It is the Dirac function. The sequence to be decomposed; Solving formula (4) involves introducing a penalty factor. The constrained variational problem is transformed into an unconstrained variational problem to reduce Gaussian noise interference, as shown in the following expression: (5)； In the formula: As a penalty factor, For Lagrange operators; The method for decomposing and modeling the historical power sequence in step S3 is as follows: the historical power sequence is decomposed into SSA-VMD, an LSTM prediction sub-model is established for each decomposed power component, and the Pearson correlation coefficient between offshore wind power output and the historical wind speed and historical power of the previous n times is analyzed to determine the time scale of the model input.

2. The offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction according to claim 1, characterized in that, The method of using the complete set empirical mode decomposition ICEEMDAN with improved adaptive noise in step S1 is as follows: Define wind speed sequence The wind speed sequence is decomposed using ICEEMDAN, and its eigenmode functions are obtained. and residual components The calculation formula is: (1)； (2)； In the formula, The signal-to-noise ratio (SNR) of white noise; For the added first Group of Gaussian white noise; The number of groups to which white noise is added.

3. The offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction according to claim 2, characterized in that, In step S1, during the secondary decomposition of wind speed data, the SSA algorithm is used to optimize the VMD decomposition parameters to address the issue that VMD cannot adaptively select the number of components. and penalty coefficient The problem specifically includes the following steps: A1: Initialize the SSA population size, set the initial... and maximum number of iterations; A2: Update the location of the finder in the sparrow population. The finder is the member with the best fitness in the population, responsible for foraging. The formula for updating its location is: (6)； In the formula, represent In the dimension Only the location of the discoverer, Represents the maximum number of iterations. This is the warning value, with a range of [0,1]. Represents a safety value; These are random numbers that follow a normal distribution. For elements all equal to 1 matrix; A3: The newcomer constantly observes the discoverer's behavior and adjusts its own position accordingly. The position update formula is: (7)； In the formula, Represents the worst current position. Represents the current globally optimal position; A matrix whose elements are all 1 or -1 and ; A4: In each iteration, sparrows in the population will be responsible for guarding the area. If they detect danger, they will abandon foraging and move to a safe area to continue foraging. The position update formula is: (8)； In the formula, Represents the current globally optimal position; These are random numbers that follow a normal distribution. Current fitness; The optimal fitness in the population. For random values ​​within the range [0,1], It is a constant; A5: Repeat the iteration until the optimal result is obtained. Values.

4. The offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction according to claim 1, characterized in that: The method for reconstructing and reducing the dimensionality of the decomposed data based on FE in step S2 is as follows: B1: Let the subsequence obtained after the combination decomposition of the sea wind speed data be... The phase space is reconstructed from the subsequence to obtain ; (9)； In the formula, for Length; Represents the embedding dimension; represent The mean of 10 adjacent sequences; B2: Define the fuzzy membership function Calculate each sequence Fuzzy entropy; (10)； (11)； (12)； In the formula, Representing two sequences The maximum value of the difference between corresponding endpoints; For noise margin; B3: The FE of the subsequence is a dimensionless number with a value in [0,1]. Based on the fuzzy entropy value of each subsequence, it is divided into 3 groups, and each group of subsequences is reconstructed into a component.

5. The offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction according to claim 4, characterized in that: The method for establishing an LSTM prediction sub-model for the power component in step S3 is as follows: An LSTM network is constructed, with each LSTM layer consisting of multiple LSTM units. Each LSTM unit has three gates: a forget gate, an input gate, and an output gate. The forget gate determines the current time step. Save the cell state from the previous time step. The number of input gates determines the cell state. Current moment The number of network inputs, and the output gate control output to the current output value of the LSTM. cell state The quantity is expressed as follows: (13)； (14)； (15)； (16)； (17)； (18)； in, , , These are the information for the forget gate, input gate, and output gate, respectively. , , This is the weight matrix; and They are Input time and The hidden layer state at any given time; , , This is the bias value; For the sigmoid function, for Constantly input information into the memory module; The weight matrix; It is the hyperbolic tangent function.

6. The offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction according to claim 1, characterized in that: The method for building the BO-LSTM model using the Bayesian algorithm in step S4 is as follows: by evaluating two sets of hyperparameters, a sequentially updated probability model is established to obtain the prior probability of the optimization problem, and then a model that satisfies the objective function is found. optimal solution .

7. The offshore wind power prediction method based on wind speed-power combination decomposition and reconstruction according to claim 6, characterized in that: The optimal solution The objective function formula is as follows: (19)； in, Range of hyperparameter optimization The optimal value.

Images