Offshore wind power prediction method and device based on multi-model collaborative fusion

Abstract

The invention discloses an offshore wind power prediction method and device based on multi-model collaborative fusion. The method comprises the steps that historical data and original complex environment data of offshore wind power are acquired; the method comprises the following steps: decomposing multivariate data by utilizing multivariate joint mode decomposition, decomposing a signal into a plurality of mode component IMFS with different frequencies, and introducing mutual information entropy to divide a mode component data set into high-frequency components and low-frequency components; through generalized S transformation, obtaining a time-frequency diagram feature of each modal component; a VTransform prediction model and a TiDE prediction model are established to predict a high-frequency component and a low-frequency component respectively, and an improved Newton-Raphson algorithm INRBO is used to optimize model hyper-parameters; and carrying out weighted fusion on prediction results of the two models by utilizing an INRBO algorithm, and carrying out error correction on weighted results by adopting an extended Kalman filter (EKF) model to obtain a final prediction result. According to the method, the precision of offshore wind power prediction can be improved, and the power grid dispatching and the operation of the offshore wind power plant are optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and particularly relates to a method and device for predicting the power of offshore wind farms based on multi-model collaborative fusion. Background Art

[0002] In the wind power prediction industry, accurately predicting the power of offshore wind farms is crucial for the stable operation of the power grid and the economic benefits of wind farms. With the increasing dependence on renewable energy and the strengthening of environmental protection regulations, accurately predicting the power of offshore wind farms has become a key task in the power industry and environmental management. The prediction of offshore wind power is affected by various factors, including offshore meteorological parameters, topography, and the performance of wind turbines. The complex interaction of these factors makes the prediction of offshore wind power a highly challenging problem.

[0003] Currently, the methods for predicting the power of offshore wind farms can be roughly divided into three categories: physics-based methods, statistics-based methods, and artificial intelligence-based methods. Physics-based methods need to consider the influence of various factors, and the calculation difficulty and cost are very high. Although statistics-based methods overcome the shortcomings of physics-based methods, there are still many deficiencies in dealing with non-linear data. In recent years, artificial intelligence-based methods have been widely used in the prediction of offshore wind power.

[0004] In view of the above problems and challenges, in the field of offshore wind power prediction, there is an urgent need to develop new prediction models and algorithms that can better process complex, non-linear, and multi-variable data. In particular, it can capture the non-linear relationship between meteorological variables such as wind speed and wind direction and wind power, extract multi-scale features from the data, and effectively decompose and analyze the data to improve the accuracy and robustness of the prediction. These new models should have stronger generalization ability and be able to handle prediction tasks in the case of offshore extreme meteorological events or sudden failures, so as to improve the stability and reliability of the prediction of the power of offshore wind farms. Summary of the Invention

[0005] Object of the Invention: Aiming at the problems existing in the prior art, the present invention provides a method and device for predicting the power of offshore wind farms based on multi-model collaborative fusion, which can effectively improve the accuracy of the wind farm power prediction model.

[0006] Technical Solution: A method for predicting the power of offshore wind farms based on multi-model collaborative fusion according to the present invention includes the following steps:

[0007] (1) Collect historical data of the offshore wind farm and the original complex environmental data collected by the environmental sensors used in the wind farm, and preprocess the data set;

[0008] (2) Multivariate data are decomposed in the frequency domain using multivariate joint modal decomposition (MJMD) to capture the multi-scale characteristics of the signal, decompose each variable into K IMF components, and introduce mutual information entropy (MIE) to classify the variables into high and low frequencies. Generalized S transform is performed on the classified data set to obtain the time-frequency diagram characteristics of each modal component.

[0009] (3) The improved Newton-Raphson algorithm INRBO is used to simultaneously optimize the parameters of the time series dense encoder TIDE and the parameters of the VTransformer model to search for the optimal solution of the two model parameters; the Cauchy-Gaussian mutation strategy and point symmetry strategy are used to improve the Newton-Raphson algorithm NRBO, and the grey correlation theory is introduced to improve the adaptive weight to obtain the INRBO algorithm;

[0010] (4) The complex decoder in the Transformer prediction model for processing time series is replaced by a fully connected layer to obtain an improved VTransformer model; at the same time, a wind power prediction model based on TIDE is established to predict high-frequency and low-frequency components respectively;

[0011] (5) The INRBO algorithm is used to perform weight fusion on the prediction results of the TIDE and VTransformer models to obtain the optimal weight combination, and the extended Kalman filter model EKF is used to correct the errors of the weighted results to obtain the final prediction results.

[0012] Furthermore, the implementation process of step (1) is as follows:

[0013] Construct a historical data set of offshore wind farms, collect wind turbine power, weather data and environmental sensor data at multiple times within a certain time period; weather data includes surface air pressure, humidity, temperature, precipitation rate, wind speed, wind direction, and air density at various altitudes; environmental sensor data includes local airflow characteristics, noise, and vibration; combine the collected offshore weather data and offshore wind power at each moment into a sample point, remove outliers, use interpolation to complete missing data within the set time period, and generate a complete data set; normalize the data.

[0014] Furthermore, the implementation process of step (2) is as follows:

[0015] (2.1) The multivariate joint modal decomposition (MJMD) method is introduced to decompose the signal components into jump variables h(t) and AM-FM modal components u k (t), where AM-FM modal component is the main component of IMF component, and the decomposition model is as follows:

[0016]

[0017] Among them, f(t) is a multivariate signal composed of C data channels, f1(t) = [f1(t), … f c (t)]; u k (t) is a set of multivariate AM - FM decomposition modes, and the number of data channels is C; h(t) and I(t) respectively refer to the multivariate jump components in the data and noise;

[0018] (2.2) Align the AM - FM mode frequencies, and use the mvmd cost function to calculate the harmonic shift gradient norm to determine the bandwidth of u k (t), so that the bandwidth of the decomposed IMF components is narrow:

[0019]

[0020] Among them, C = 1, 2, …, C represents the number of channels;

[0021] (2.3) Define an auxiliary variable to decompose the variable, and the calculation formula is as follows:

[0022]

[0023] h c +∑ k u k,c =f c (5)

[0024]

[0025] Among them, h c is the jump component, and ∑ k u k,c is the superposition of AM - FM modes;

[0026] (2.4) Update the oscillation mode u k (t), ensure that the separated IMF components can accurately reflect the characteristics of the signal, so as to obtain K different frequency IMF components of wind power, and the specific formula is as follows:

[0027]

[0028] (2.5) After MJMD decomposition, decompose the signal into different IMF1, IMF2, …, IMF n components, calculate the energies E1, E1, …, E n of each component respectively, and calculate the intrinsic energy entropy of each IMF component based on these energy values, and the specific formula is as follows:

[0029]

[0030] In the formula, E t is the energy entropy of the t-th component, Rt is the energy of the t-th component, R is the total energy of the signal, and N is the number of components obtained by MJMD decomposition;

[0031] (2.6) Calculate the mutual information entropy of adjacent components to measure the correlation degree between the energy entropies of adjacent components after MJMD decomposition:

[0032] I(imf t ,imf t+1 ) = Z(imf t ) + Z(imf t+1 ) - Z(imf t imf t+1 ) (10)

[0033] (2.7) The mutual information entropy of adjacent IMF components first decreases and then increases as the frequency decreases from high to low, and there is a turning point; according to the mutual information principle, at this turning point, the components are divided into high-frequency and low-frequency parts, which is expressed as:

[0034] x(t) = h i，k (x) + l j，k (x) + r n (x) (11)

[0035] ifI(imf t , imf t+1 )↓andI(imf t+1 , imf t+2 )↑ (12)

[0036]

[0037] where h i，k (x) is the high-frequency component; l j，k (x) is the low-frequency component; i = 1, 2,..., k; j = k, k + 1,..., n;

[0038] (2.8) For the x(t) signal, perform a generalized S transform on the decomposed high-frequency IMF H and low-frequency IMF L components. The window width of the transform is inversely proportional to the frequency, and the corresponding window function and the expression of the generalized S transform are as follows:

[0039]

[0040] where x(t) is the original signal;

[0041] (2.9) The generalized S-transform can adjust the values of λ and v to optimize the time-frequency resolution. According to the time-frequency focusing criterion, the time-frequency energy concentration degree is measured by the magnitude of energy, so as to obtain better time-frequency map features TIMF. H and TIMF L , and the specific calculation formula is as follows:

[0042]

[0043] In the formula, q is a value greater than 1, and the smaller M v , the better the time-frequency focusing degree of the generalized S-transform, and the corresponding is the optimal window function parameter.

[0044] Further, the parameters of the TiDE described in step (3) include the total number of TiDE input features, the size of the hidden layer, and the number of stacked residual blocks used in the TiDE encoder and decoder; the parameters of the VTransformer include the VTransformer hidden layer dimension, the number of heads in the multi-head attention mechanism, the size of the hidden layer in the feed-forward network, the learning rate, and the parameters of Attention.

[0045] Further, the implementation process of the step (3) is as follows:

[0046] (3.1) Set the population size, dimension, number of iterations of the NRBO algorithm, and the upper and lower limits of the search space, and randomly select an initial point X0 from the parameter domain as the starting point of the algorithm;

[0047] (3.2) The Cauchy-Gaussian mutation strategy can generate a balance between long-tailed and short-tailed distributions during the search process, increase the population diversity, and jump out of the local optimum; the point symmetry strategy is introduced to enhance the search ability of the population. The improved formula is as follows:

[0048]

[0049] Among them, is the position information of the population's optimal individual after mutation; W is the standard deviation of the Cauchy-Gaussian mutation theorem, and cauchy(0, w) and Gauss(0, w) respectively represent random variables subject to the Cauchy distribution and the Gaussian distribution; η1 and η2 are parameters that change dynamically with the iteration process;

[0050] (3.3) The NRBO algorithm uses the mean square error MSE for adaptive fitness evaluation, and introduces an adaptive coefficient δ to balance the search capabilities of the exploration stage and the exploitation stage; to ensure dynamic balance, the parameter δ adaptively changes from 1 to -1 during the iteration process. The specific formula is as follows:

[0051]

[0052] where \(t\) represents the current iteration number, \(max\_t\) represents the maximum iteration number, \(y\) i represents the \(i\)-th true value, and \(\hat{y}\)

[0053] (3.4) The grey relational degree theory is introduced to improve the adaptive weight:

[0054]

[0055] where \(T\) i is the adaptive weight, and \(\gamma\) i is the grey relational degree theory;

[0056] (3.5) The Newton-Raphson search rule NRSR is introduced to accelerate the convergence of individuals in the population to the optimal solution, improving the search efficiency; using the second derivative information to obtain a better position solution, and the second derivative expression is:

[0057]

[0058] (3.6) Use the first few components of the Taylor series of \(f(k)\) to locate the root of the function in the assumed root region and calculate the position of the next iteration point. The specific calculation formula is as follows:

[0059]

[0060] where \(k_n\) n represents the solution of the \(n\)-th term in the Taylor series, \(\Delta k\) represents the difference between the known solution and the actual solution;

[0061] (3.7) NRSR is changed to use the positional \(k\) n to replace the non-fitness value \(f(k\) n ) and search based on the position of the individual. NRSR modifies the formula to:

[0062]

[0063] \(k\) n+1 = \(k\) n - NRSR(27)

[0064] where \(randn\) is a normally distributed random number, \(L_w\) w is the worst position, and \(L_b\) b is the best position;

[0065] (3.8) NRBO introduces a trap avoidance operator TAO. When the random number \(rand\in[0,1]\) is less than 0.6, TAO generates a new position b by combining the optimal position \(L_b\) and the current position Adjust the position for the next iteration The specific calculation formula is as follows:

[0066]

[0067] Among them, θ1 and θ2 are random numbers between -1 and 1 and random numbers between -0.5 and 0.5 respectively;

[0068] μ1 = α × 3 × rand + (1 - α) (31)

[0069] μ2 = α × rand + (1 - α) (32)

[0070]

[0071] Among them, rand represents a random number between 0 and 1, and Δ represents a number between 0 and 1; Since the parameters μ1 and μ2 are random numbers;

[0072] (3.9) Search for the global optimal solution according to the position update formula to obtain the optimal parameter combination of the TIDE and VTransformer models;

[0073] (3.10) If the maximum number of iterations is not reached, return to step (3.3) until the maximum number of iterations is reached, and output the optimal parameter configuration; The TIDE and VTransformer models use the optimal parameter configuration to predict the components to obtain the prediction results of the high-frequency and low-frequency components.

[0074] Furthermore, the implementation process of the said step (4) is as follows:

[0075] (4.1) The input of the L-th layer encoder of the Transformer is the output of the previous layer. Each encoder is used to extract the features of the input data layer by layer, such as the formula:

[0076] U0 = Embed(M) + PosEnc(M) (34)

[0077] Among them, Embed(M) is the embedding mapping, and PosEnc(M) is the position encoding;

[0078] (4.2) Improve the Transformer, use a linear layer to replace the complex decoder in the original Transformer. The input of the VTransformer first performs position encoding on the data after the embedding layer, and then extracts the information of the required positions in the sequence; The formula is as follows:

[0079]

[0080] Where 2a represents the dimension of even numbers, 2a + 1 represents the dimension of odd numbers, pos represents the position of data in the sequence, and d model is the model dimension;

[0081] (4.3) In the improved VTransformer model, the attention weights are first obtained by query vector matrix Q and key vector matrix K, and then the calculated weights are applied to the value vector matrix V to obtain the overall weighted output vector. The calculation formula is:

[0082]

[0083] (4.4) The feed-forward neural network consists of two fully connected layers and a Relu activation function. The calculation formula is as follows:

[0084] FNN(x) = G2max(0, G1x + o1) + o2 (38)

[0085] Where X is the input vector data with dimension d model , o1 and o2 are bias terms, and G1 and G2 are parameter matrices of linear mapping;

[0086] (4.5) The prediction result of TiDE is regarded as the joint mapping relationship of historical time series y, attribute a, and dynamic covariate x, expressed as:

[0087]

[0088] Where K is the backtracking time step, H is the prediction step, and Q is the number of time series;

[0089] (4.6) The encoder part in TiDE extracts time series features through the feature projection module, and uses the residual block to map the input feature matrix of each time step from the original dimension r to a lower-dimensional space The specific formula is as follows:

[0090]

[0091] (4.7) In the dense coding part, the dynamic covariate feature projection is flattened, concatenated with the static attribute and historical time series, and sent as input to the encoder containing multiple residual blocks. The specific formula is as follows:

[0092]

[0093] (4.8) The dense decoder takes the encoded B (i) as input, maps it to a vector g of size H × P (i) , and the vector g (i) is reshaped into a matrix D (i) ∈Rp×H , each column represents the decoded vector of a predicted time step, and the specific calculation formula is as follows:

[0094] g (i) = Decoder(e (i) ) ∈ R H×p (42)

[0095] D (i) = Reshape(g (i) ) ∈ R p×H (43)

[0096] (4.9) The time decoder maps the output of the dense decoder to the time series prediction value; the decoded vector of the t-th predicted time step is concatenated with the corresponding covariate projection , and then the feature transformation is performed through the residual block to generate the final prediction value The formula is as follows:

[0097]

[0098] where

[0099] (4.10) Establish a marine wind power prediction model based on VTransformer and TiDE, and input the high-frequency component and low-frequency component obtained in step (2) into the improved VTransformer prediction model and TiDE prediction model respectively for feature learning and prediction.

[0100] Further, the process of using the INRBO algorithm to perform weight fusion on the prediction results of the TIDE and VTransformer models in step (5) is as follows:

[0101] Establish an input matrix with the obtained high- and low-frequency component prediction results; construct a weight optimization objective function to minimize the prediction error, and the formula is as follows:

[0102]

[0103] In the formula, L(w) is the weight result, y j is the true value, N is the number of samples, and w i is the weight of each component;

[0104] Calculate the gradient of the objective function and the Hessian matrix H, and the formula is as follows:

[0105]

[0106]

[0107] The weights of each component are adaptively optimized using the INRBO algorithm to obtain the final wind power prediction result.

[0108] Furthermore, the implementation process of using the Extended Kalman Filter model EKF to correct the error of the weighted result in step (5) is as follows:

[0109] Predict the state estimate value and covariance matrix at the current moment through the state transition equation, and the calculation formula is as follows:

[0110] x′ = Ux + AS (48)

[0111] M' = UMU T + Q (49)

[0112] Among them, x' represents the current estimated value, U represents the state transition matrix, x represents the optimal estimated value at the previous moment, A represents the external input matrix, S represents the external state matrix, M' represents the current prior estimated covariance matrix,; M represents the optimal estimated covariance matrix at the previous moment; Q represents the covariance matrix of the process;

[0113] During the update process, use the linearized measurement equation to combine the prior estimated covariance matrix with the observation matrix to calculate the Kalman gain at the current moment, and the formula is as follows:

[0114] K = M'HT(HM'HT + R) -1 (50)

[0115] Among them, K represents the Kalman gain at the current moment; H represents the observation matrix;

[0116] Use the Kalman gain to incorporate the actual measurement value into the estimation process, and correct and calibrate the predicted state estimate value and covariance matrix by means of the deviation between the actual measurement value and the current state estimate value, so as to obtain the final prediction result; its update formula is as follows:

[0117] x = x′ + K(z - Hx′) (51)

[0118] P = (E - KH)M′ (52)

[0119] Among them, x represents the optimal estimated value at the current moment; z represents the actual measurement value; M represents the covariance matrix at the current moment; E represents the identity matrix;

[0120] The EKF continuously corrects the state estimation value until a more accurate result is obtained. The EKF regards the offshore wind power state as a dynamic system, uses sensors to collect data to iteratively update the wind power, and gradually reduces the estimation error by continuously repeating the prediction and update iteration process. The EKF can obtain a more accurate estimation value of the offshore wind power state and get the final prediction result.

[0121] An electronic device according to the present invention includes a memory and a processor, wherein:

[0122] The memory is used to store a computer program that can run on the processor;

[0123] The processor is used to execute the steps of the offshore wind power prediction method based on multi-model collaborative fusion as described above when running the computer program.

[0124] A storage medium according to the present invention stores a computer program, and when the computer program is executed by at least one processor, the steps of the offshore wind power prediction method based on multi-model collaborative fusion as described above are implemented.

[0125] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention performs frequency-domain decomposition on multivariate data by introducing MJMD multivariate joint mode decomposition to capture the multi-scale characteristics of signals. Each variable is decomposed into K IMF components, which helps to identify and separate useful information and noise in the signals; The mutual information entropy (MIE) is introduced to classify variables into high and low frequencies, enhancing the model's ability to identify different frequency characteristics of signals; And a generalized S transform is performed on the classified dataset to obtain the time-frequency diagram features of each modal component, which helps to capture the dynamic characteristics of wind power changes; The present invention improves the Newton-Raphson algorithm NRBO by introducing a Cauchy-Gaussian mutation strategy and a point symmetry strategy to improve the initialization process of the original population, which helps the algorithm to jump out of the local optimal region in the global search stage and enhances the global exploration ability of the algorithm; Secondly, the adaptive weight is improved by introducing the grey relational degree theory, improving the adaptive ability and search efficiency of the algorithm; The present invention replaces the complex-structured decoder in the Transformer prediction model for processing time series with a single fully connected layer to obtain an improved VTransformer model, reducing the complexity of the model and improving the calculation efficiency. The multi-head self-attention mechanism retained in the encoder can better capture the internal correlations of data; And TiDE demonstrates strong capabilities in wind power prediction for handling complex nonlinear relationships, capturing long-term dependencies, extracting useful features, and adapting to different time steps, improving the accuracy and efficiency of prediction; The present invention uses the improved Newton-Raphson algorithm INRBO to optimize the hyperparameters of the prediction model, enhancing the generalization ability of the model. The Newton-Raphson search rule is introduced to accelerate the convergence of population individuals to the optimal solution, improving the search speed of the algorithm; The introduction of a trap avoidance operator (TAO) and an adaptive weight strategy enhances the global search ability of the algorithm, further improving the accuracy and reliability of prediction; Additionally, by weighted integration of the prediction results of high-frequency and low-frequency components through the INRBO algorithm, a more accurate wind power prediction result can be output; The present invention performs error correction on the weighted result through an extended Kalman filter model (EKF), further reducing the prediction error to obtain the final prediction result. Description of the Drawings

[0126] Figure 1 It is a flowchart of a method for predicting offshore wind power based on multi-model collaborative fusion;

[0127] Figure 2 It is a schematic structural diagram of the VTransformer model;

[0128] Figure 3 It is a schematic structural diagram of the TiDE model. Detailed Embodiments

[0129] The present invention will be further described in detail below with reference to the drawings.

[0130] like Figure 1 As shown, the present invention provides an offshore wind power prediction method based on multi-model collaborative fusion, and the specific steps are as follows:

[0131] Step 1: Collect historical data of offshore wind farms and original complex environmental data collected by environmental sensors used in wind farms, and preprocess the data set.

[0132] (1.1) Construct a historical data set for offshore wind farms and collect wind turbine power, weather data, and environmental sensor data at multiple times within a certain period of time. Weather data covers surface pressure, humidity, temperature, precipitation rate, wind speed, wind direction, and air density at various altitudes, while environmental sensor data contains complex environmental information such as local airflow characteristics, noise, and vibration.

[0133] (1.2) The collected offshore weather data and wind power at each moment are combined into a sample point, outliers are removed, and the missing data within the set time period are supplemented by interpolation, and finally a complete data set is generated.

[0134] (1.3) In order to improve the prediction fit of the model, the data is normalized. The specific formula is as follows:

[0135]

[0136] Among them, N represents the original data, N max and N min They represent the maximum and minimum values ​​of the column data respectively, and N′ represents the normalized data.

[0137] Step 2: Use multivariate joint mode decomposition (MJMD) to decompose the multivariate data in the frequency domain, decompose each variable into K IMF components, introduce mutual information entropy (MIE) to classify the variables into high and low frequencies, and perform generalized S transform in the classified data set.

[0138] (2.1) The multivariate joint modal decomposition (MJMD) method is introduced to decompose the signal components into jump variables h(t) and AM-FM modal components u k (t), where AM-FM modal component is the main component of IMF component, and the decomposition model is as follows:

[0139]

[0140] Where f(t) is a multivariate signal consisting of C data channels f1(t) = [f1(t), (…f c (t)]. k(t) is a set of multivariate AM - FM decomposition modes, and the number of data channels is C. Finally, h(t) and I(t) refer to the multivariate jump components in the data and noise respectively.

[0141] (2.2) To directly compare the modes of different channels, it is necessary to align the AM - FM mode frequencies. The harmonic shift gradient norm is calculated using the mvmd cost function to determine the bandwidth of u k (t), so that the decomposed IMF components have narrow bandwidths, improve the decomposition accuracy, and reduce mode aliasing. The specific calculation formula is as follows:

[0142]

[0143] where C = 1, 2, …, C represents the number of channels.

[0144] (2.3) By defining an auxiliary variable to decompose the variable, the calculation formula is as follows:

[0145]

[0146] h c +∑ k u k,c =f c (5)

[0147]

[0148] where: h c is the jump component, and ∑ k u k,c is the superposition of AM - FM modes.

[0149] (2.4) Update the oscillatory mode u k (t) to ensure that each separated oscillatory mode (IMF component) can accurately reflect the characteristics of the signal, so as to obtain K different frequency IMF components of the wind power. The specific formula is as follows:

[0150]

[0151] (2.5) After MJMD decomposition, the signal can be decomposed into different IMF1, IMF2, …, IMF n components, and the energies E1, E1, …, E n of each component are calculated respectively. Based on these energy values, the intrinsic energy entropy of each IMF component is obtained. The specific formula is as follows:

[0152]

[0153] In the formula, E t$E_{t}$ is the energy entropy of the $t$-th component, $R_{t}$ is the energy of the $t$-th component, $R$ is the total energy of the signal, and $K$ is the number of components obtained by MJMD decomposition.

[0154] (2.6) To measure the correlation degree between the energy entropies of adjacent components after MJMD decomposition, the mutual information entropy of adjacent components can be calculated. The specific calculation formula is as follows:

[0155] $I(IMF_{i},IMF_{j})$ t , $IMF_{j}$ t+1 ) = $Z(IMF_{i})$ t ) + $Z(IMF_{j})$ t+1 ) - $Z(IMF_{i}$ t $IMF_{j}$ t+1 ) (10)

[0156] (2.7) The mutual information entropy of adjacent IMF components shows a trend of first decreasing and then increasing as the frequency changes from high to low, and there is a turning point. According to the mutual information principle, the components can be divided into high-frequency and low-frequency parts at this turning point, which can be expressed as:

[0157] $x(t)=h_{i}(x)+l_{j}(x)+r(x)$ (11) i，k $(x)+l_{j}$ j，k $(x)+r$ n $(x)$ (11)

[0158] if $I(IMF_{i}$ t , $IMF_{j}$ t+1 )↓ and $I(IMF_{i}$ t+1 , $IMF_{j}$ t+2 )↑ (12)

[0159]

[0160] where $h_{i}$ i，k $(x)$ is the high-frequency component; $l_{j}$ j，k $(x)$ is the low-frequency component; $i = 1,2,\cdots,k$; $j = k,k + 1,\cdots,n$.

[0161] (2.8) For the $x(t)$ signal, the high-frequency IMF H and low-frequency IMF L components obtained by decomposition are subjected to the generalized S transform. The window width of the transform is inversely proportional to the frequency. The corresponding window function and the expression of the generalized S transform are as follows:

[0162]

[0163] where $x(t)$ is the original signal.

[0164] (2.9) The generalized S-transform can adjust the values of λ and v to optimize the time-frequency resolution. According to the time-frequency focusing criterion, the time-frequency energy concentration is measured by the magnitude of energy, so as to obtain better time-frequency map features TIMF. H and TIMF L , and the specific calculation formula is:

[0165]

[0166] In the formula, q is a value greater than 1, and the smaller M v , the better the time-frequency focusing of the generalized S-transform, and the corresponding is the optimal window function parameter.

[0167] Step 3: Use the improved Newton-Raphson algorithm INRBO to synchronously optimize the parameters of the time series dense encoder TIDE and the parameters of the VTransformer model, and search for the optimal solutions of the two model parameters; adopt the Cauchy-Gaussian mutation strategy and the point symmetry strategy to improve the Newton-Raphson algorithm NRBO, and introduce the grey relational degree theory to improve the adaptive weight to obtain the INRBO algorithm.

[0168] (3.1) First, set the population size, dimension, number of iterations of the NRBO algorithm, and the upper and lower limits of the search space, and randomly select an initial point X0 from the parameter domain as the starting point of the algorithm.

[0169] (3.2) Introduce the Cauchy-Gaussian mutation strategy and the point symmetry strategy to improve the algorithm. The Cauchy-Gaussian mutation strategy can generate a balance between long-tailed and short-tailed distributions during the search process, increase population diversity, and jump out of local optima; introducing the point symmetry strategy enhances the search ability of the population. The improved formula is as follows:

[0170]

[0171] Among them, is the position information of the mutated optimal individual in the population; W is the standard deviation of the Cauchy-Gaussian mutation theorem, and cauchy(0, w) and Gauss(0, w) respectively represent random variables subject to the Cauchy distribution and the Gaussian distribution; η1 and η2 are parameters that change dynamically with the iteration process.

[0172] (3.3) The NRBO algorithm uses the mean square error (MSE) for adaptive fitness evaluation, and introduces an adaptive coefficient δ to balance the search capabilities of the exploration stage and the exploitation stage. To ensure dynamic balance, the parameter δ adaptively changes from 1 to -1 during the iteration process. The specific formula is as follows:

[0173]

[0174] where \(t\) represents the current iteration number, \(max\_t\) represents the maximum iteration number, \(y\) i represents the \(i\)-th true value, and \(\hat{y}\)

[0175] (3.4) The grey relational degree theory is introduced to improve the adaptive weight, and the formula is as follows:

[0176]

[0177] where \(T\) i is the adaptive weight, and \(\gamma\) i is the grey relational degree theory.

[0178] (3.5) The Newton-Raphson search rule NRSR is the core search mechanism of the NRBO algorithm. NRSR improves the search efficiency by accelerating the convergence of individuals in the population towards the optimal solution. The key lies in using the second derivative information to more precisely guide the control vector to explore the feasible solution space, thereby obtaining a better position solution. The second derivative expression is:

[0179]

[0180] (3.6) The Newton-Raphson method starts from a hypothetical initial solution, makes a local approximation at the current point, and iteratively updates along a determined direction to gradually approach the root of the function. The Taylor series of the first few components of \(f(k)\) is used in the region of the assumed root to locate the root of the function and calculate the position of the next iteration point. The specific calculation formula is as follows:

[0181]

[0182] where \(k\) n respectively represents the solution of the \(n\)-th term in the Taylor series, and \(\Delta k\) represents the difference between the known solution and the actual solution.

[0183] (3.7) To save computational time, NRSR changes to use the positional \(k\) n instead of the fitness value \(f(k\) n ). By directly searching based on the position of the individual, the computational complexity is reduced. NRSR modifies the formula as:

[0184]

[0185] k n+1 =k n - NRSR(27)

[0186] where \(randn\) is a normally distributed random number, \(L\) w is the worst position, and \(L\) b is the best position.

[0187] (3.8) NRBO introduces the trap avoidance operator (TAO). When the random number rand∈[0,1] is less than 0.6, TAO passes through the optimal position L b With current location Combinations of So as to adjust the position of the next iteration This mechanism avoids falling into local optimality, enables individuals to jump out of the current trap, and improves the global search ability of the algorithm. The specific calculation formula is as follows:

[0188]

[0189] Among them, θ1 and θ2 are random numbers between -1 and 1 and between -0.5 and 0.5 respectively.

[0190] μ1=α×3×rand+(1-α) (31)

[0191] μ2=α×rand+(1-α) (32)

[0192]

[0193] Among them, rand represents a random number between 0 and 1, and Δ represents a number between 0 and 1. Since the selection of parameters μ1 and μ2 is random, the population diversity becomes richer, which further helps to avoid falling into the local optimal solution.

[0194] (3.9) According to the position update formula, the global optimal solution is searched to obtain the optimal parameter combination of TIDE and VTransformer models.

[0195] (3.10) If the maximum number of iterations is not reached, return to step (3.3) until the maximum number of iterations is reached and the optimal parameter configuration is output; the TIDE and VTransformer models use the optimal parameters to configure the prediction components and obtain the prediction results of the high-frequency components and the low-frequency components.

[0196] Hyperparameters of the VTransformer and TiDE models, including VTransformer hidden layer dimensions, the number of heads in the multi-head attention mechanism, the size of the hidden layers in the feedforward network, the learning rate, the parameters of Attention (query, key, and value), the total number of TiDE input features and the size of the hidden layers, the number of stacked residual blocks used in the TiDE encoder and decoder, etc.

[0197] Step 4: Replace the complex decoder in the Transformer prediction model for processing time series with a single fully connected layer to obtain an improved VTransformer model. At the same time, establish a wind power prediction model based on TIDE to predict high-frequency and low-frequency components respectively.

[0198] (4.1) As Figure 2 shown, the Transformer is stacked by multiple encoder modules. The one-dimensional time series input is first mapped to a high-dimensional space through a data embedding layer, and the time position information is introduced through position encoding to retain the time series characteristics. The input of the L-th layer encoder is the output of the previous layer, and each encoder is used to extract the features of the input data layer by layer, as shown in the formula:

[0199] U0 = Embed(M) + PosEnc(M) (34)

[0200] where Embed(M) is the embedding mapping and PosEnc(M) is the position encoding.

[0201] (4.2) Aiming at the non-linearity of the wind power environment data time series, the present invention proposes an improved VTransformer model, which replaces the complex decoder in the original Transformer with a single linear layer. The input of the VTransformer first performs position encoding on the data after the embedding layer, and then extracts the information of the required positions in the sequence. The formula is as follows:

[0202]

[0203] In the formula, 2a represents the dimension of even numbers, 2a + 1 represents the dimension of odd numbers, pos is the position of the data in the sequence, and d model is the model dimension.

[0204] (4.3) In the improved VTransformer model, first calculate the attention weights through the query vector matrix Q and the key vector matrix K, and then apply the calculated weights to the value vector matrix V to obtain the overall weighted output vector. The calculation formula is:

[0205]

[0206] (4.4) The feedforward neural network consists of two fully connected layers and a Relu activation function. The calculation formula is as follows:

[0207] FNN(x) = G2max(0, G1x + o1) + o2 (38)

[0208] In the formula, X is the dimension of d modelThe input vector data, o1 and o2 are bias terms, and G1 and G2 are parameter matrices of linear mappings.

[0209] (4.5) The prediction result of TiDE can be regarded as the joint mapping relationship of the historical time series y, the attribute a, and the dynamic covariate x, expressed as:

[0210]

[0211] Among them, K is the backtracking time step, H is the prediction step, and Q is the number of time series.

[0212] (4.6) As Figure 3 shown, TiDE is an encoder-decoder architecture based on a multi-layer perceptron. The encoder part extracts time series features through a feature projection module and uses a residual block to map the input feature matrix at each time step from the original dimension r to a lower-dimensional space The specific formula is as follows:

[0213]

[0214] (4.7) In the dense encoding part, the dynamic covariate features are projected and flattened, concatenated with the static attributes and the historical time series, and sent as input to an encoder containing multiple residual blocks. The specific formula is as follows:

[0215]

[0216] (4.8) The dense decoder takes the encoded B (i) as input and maps it to a vector g (i) of size H×P. The vector g (i) is reshaped into a matrix D (i) ∈R p×H , and each column represents the decoded vector for a prediction time step. The specific calculation formula is as follows:

[0217] g (i) = Decoder(e (i) ) ∈R H×p (42)

[0218] D (i) = Reshape(g (i) ) ∈R p×H (43)

[0219] (4.9) The time decoder maps the output of the dense decoder to the time series prediction value. It concatenates the decoded vector for the t-th prediction time step with the corresponding covariate projection , and then performs feature transformation through a residual block to generate the final prediction value The formula is as follows:

[0220]

[0221] Among them,

[0222] (4.10) Establish a wind power prediction model based on VTransformer and TiDE, and input the high-frequency component and low-frequency component obtained in step (2) into the improved VTransformer prediction module and TiDE prediction module respectively for feature learning and prediction.

[0223] Step 5: Use the INRBO algorithm to fuse the prediction results of the TIDE and VTransformer models to obtain the optimal weight combination, and use the extended Kalman filter model EKF to correct the error of the weighted result to obtain the final prediction result.

[0224] (5.1) Establish an input matrix based on the high-frequency and low-frequency component prediction results obtained in step (3.9);

[0225] (5.2) Construct a weight optimization objective function to minimize the prediction error. The formula is as follows:

[0226]

[0227] In the formula, L(w) is the weight result, y j is the true value, N is the number of samples, and w i is the weight of each component.

[0228] (5.3) Calculate the gradient of the objective function and the Hessian matrix H. The formula is as follows:

[0229]

[0230] (5.4) Use the INRBO algorithm to adaptively optimize the weights of each component to obtain the final wind power prediction result.

[0231] (5.5) According to the dynamic model of the system, predict the state estimate value and covariance matrix at the current moment through the state transition equation. The calculation formula is as follows:

[0232] x′ = Ux + AS (48)

[0233] M′ = UMUT + Q (49)

[0234] Wherein, x′ represents the current estimated value, U represents the state transition matrix, x represents the optimal estimated value at the previous moment, A represents the external input matrix, S represents the external state matrix, M′ represents the current prior estimated covariance matrix; M represents the optimal estimated covariance matrix at the previous moment; Q represents the covariance matrix of the process.

[0235] (5.6) During the update process, using the linearized measurement equation, combining the prior estimated covariance matrix with the observation matrix, calculate the Kalman gain at the current moment, and the formula is as follows:

[0236] K = M'H T (HM'H T +R) -1 (50)

[0237] Wherein, K represents the Kalman gain at the current moment; H represents the observation matrix.

[0238] (5.7) Use the Kalman gain to incorporate the actual measurement value into the estimation process, and correct and calibrate the predicted state estimated value and the covariance matrix by means of the deviation between the actual measurement value and the current state estimated value, so as to obtain the final prediction result. The update formula is as follows:

[0239] x = x′ + K(z - Hx′) (51)

[0240] P = (E - KH)M′ (52)

[0241] Wherein, x represents the optimal estimated value at the current moment; z represents the actual measurement value; M represents the covariance matrix at the current moment; E represents the identity matrix.

[0242] (5.8) Judge whether the requirements are met. If not, return to step (5.5). The EKF continuously corrects the state estimated value until a more accurate result is obtained. The EKF regards the offshore wind power state as a dynamic system, uses sensors to collect data to iteratively update the wind power, and gradually reduces the estimation error by continuously repeating the prediction and update iteration process. The EKF can obtain a more accurate offshore wind power state estimated value, so as to obtain the final prediction result.

[0243] The present invention also provides an electronic device, including a memory and a processor, wherein: the memory is used for storing a computer program that can run on the processor; the processor is used for executing the steps of the above-mentioned offshore wind power prediction method based on multi-model collaborative fusion when running the computer program.

[0244] The present invention also provides a storage medium, on which a computer program is stored. When the computer program is executed by at least one processor, the steps of the above-mentioned method for predicting the power of an offshore wind farm based on multi-model collaborative fusion are implemented.

[0245] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for offshore wind power prediction based on multi-model collaborative fusion, characterized in that: The following steps are involved: (1) Collect historical data of offshore wind farms and original complex environmental data collected by environmental sensors used in wind farms, and preprocess the data sets; (2) Multivariate data are decomposed in the frequency domain using multivariate joint modal decomposition (MJMD) to capture the multi-scale characteristics of the signal, decompose each variable into K IMF components, and introduce mutual information entropy (MIE) to classify the variables into high and low frequencies. Generalized S transform is performed on the classified data set to obtain the time-frequency diagram characteristics of each modal component. (3) The improved Newton-Raphson algorithm INRBO is used to simultaneously optimize the parameters of the time series dense encoder TIDE and the parameters of the VTransformer model to search for the optimal solution of the two model parameters; the Cauchy-Gaussian mutation strategy and point symmetry strategy are used to improve the Newton-Raphson algorithm NRBO, and the grey correlation theory is introduced to improve the adaptive weight to obtain the INRBO algorithm; (4) The complex decoder in the Transformer prediction model for processing time series is replaced by a fully connected layer to obtain an improved VTransformer model; at the same time, a wind power prediction model based on TIDE is established to predict high-frequency and low-frequency components respectively; (5) The INRBO algorithm is used to perform weight fusion on the prediction results of the TIDE and VTransformer models to obtain the optimal weight combination, and the extended Kalman filter model EKF is used to correct the errors of the weighted results to obtain the final prediction results.

2. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The implementation process of step (1) is as follows: Construct a historical data set of offshore wind farms, collect wind turbine power, weather data and environmental sensor data at multiple times within a certain time period; weather data includes surface air pressure, humidity, temperature, precipitation rate, wind speed, wind direction, and air density at various altitudes; environmental sensor data includes local airflow characteristics, noise, and vibration; combine the collected offshore weather data and offshore wind power at each moment into a sample point, remove outliers, use interpolation to complete missing data within the set time period, and generate a complete data set; normalize the data.

3. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The implementation process of step (2) is as follows: (2.1) The multivariate joint modal decomposition (MJMD) method is introduced to decompose the signal components into jump variables h(t) and AM-FM modal components u k (t), where AM-FM modal component is the main component of IMF component, and the decomposition model is as follows: Where f(t) is a multivariate signal consisting of C data channels f1(t) = [f1(t),…f c (t)];u k (t) is the set of multivariate AM-FM decomposition modes, and the number of data channels is C; h(t) and I(t) refer to the multivariate jump components in the data and noise, respectively; (2.2) For the frequency alignment of AM-FM mode, the harmonic shift gradient norm is calculated using the mvmd cost function to determine u k (t) makes the bandwidth of the decomposed IMF component narrow: Where, C = 1, 2, ..., C represents the number of channels; (2.3) Define an auxiliary variable To decompose the variables, the calculation formula is as follows: h c +∑ k u k,c =f c (5) Among them, h c is the jump component, ∑ k u k,c It is the superposition of AM-FM mode; (2.4) Update the oscillation mode u k (t), ensuring that the separated IMF components can accurately reflect the characteristics of the signal, thereby obtaining K different frequency IMF components of wind power. The specific formula is as follows: (2.5) After MJMD decomposition, the signal is decomposed into different IMF1, IMF2, …, IMF n Components, calculate the energy of each component E1,E1,…,E n , based on these energy values, the inherent energy entropy of each IMF component is calculated, and the specific formula is as follows: In the formula, E t is the energy entropy of the t-th component, Rt is the energy of the t-th component, R is the total energy of the signal, and N is the number of components obtained by MJMD decomposition; (2.6) Calculate the mutual information entropy of adjacent components to measure the correlation between the energy entropies of adjacent components after MJMD decomposition: I(imf t ,imf t+1 )=Z(imf t )+Z(imf t+1 )-Z(imf t imf t+1 ) (10) (2.7) The mutual information entropy of adjacent IMF components decreases first and then increases as the frequency changes from high frequency to low frequency, and there is a turning point. According to the mutual information principle, the component is divided into a high-frequency part and a low-frequency part at this turning point, which can be expressed as: x(t)=h i,k (x)+l j,k (x)+r n (x) (11) ifI(imf t ,imf t+1 )↓and I(imf t+1 ,imf t+2 )↑ (12) Among them, h i,k (x) is the high frequency component; l j,k (x) is the low-frequency component; i = 1, 2, ..., k; j = k, k+1, ..., n; (2.8) For the x(t) signal, the decomposed high-frequency IMF H , low frequency IMF L The components are generalized S transform, the transform window width is inversely proportional to the frequency, and the corresponding window function and the expression of generalized S transform are expressed as follows: Among them, x(t) is the original signal; (2.9) The generalized S transform can adjust the values ​​of λ and v to optimize the time-frequency resolution. According to the time-frequency focusing criterion, the time-frequency energy concentration is measured by the energy size, so as to obtain a better time-frequency feature TIMF. H ,TIMF L , the specific calculation formula is: Where q is a value greater than 1, M v The smaller it is, the better the time-frequency focusing of the generalized S transform is, and the corresponding window function parameter is the optimal one.

4. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The parameters of the TiDE in step (3) include the total number of TiDE input features and the size of the hidden layer, and the number of stacked residual blocks used in the TiDE encoder and decoder; the parameters of the VTransformer include the VTransformer hidden layer dimension, the number of heads in the multi-head attention mechanism, the size of the hidden layer in the feedforward network, the learning rate, and the parameters of Attention.

5. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The implementation process of step (3) is as follows: (3.1) Set the population size, dimension, number of iterations, and upper and lower limits of the search space of the NRBO algorithm, and randomly select an initial point X0 from the parameter definition domain as the starting point of the algorithm; (3.2) The Cauchy-Gauss mutation strategy can produce a balance between long-tail and short-tail distributions during the search process, increase population diversity, and escape from local optimality; the point symmetry strategy is introduced to enhance the search ability of the population. The improved formula is as follows: in, is the position information of the optimal individual in the population after mutation; W is the standard deviation of the Cauchy-Gaussian mutation theorem, cauchy(0,w) and Gauss(0,w) represent random variables that obey the Cauchy distribution and Gaussian distribution respectively; η1 and η2 are parameters that change dynamically with the iteration process; (3.3) The NRBO algorithm uses mean square error MSE for adaptability evaluation and introduces an adaptive coefficient δ to balance the search capabilities in the exploration phase and the development phase. In order to ensure dynamic balance, the parameter δ adaptively changes from 1 to -1 during the iteration process. The specific formula is as follows: In the formula, t represents the current number of iterations, max_t represents the maximum number of iterations, and y i represents the i-th true value, represents the i-th predicted value; (3.4) Introduce the grey relational theory to improve the adaptive weight: Among them, T i is the adaptive weight, γ i It is the grey relational theory; (3.5) The Newton-Raphson search rule NRSR is introduced to accelerate the convergence of individuals in the population to the optimal solution, thereby improving the search efficiency. The second-order derivative information is used to obtain a better position solution. The second-order derivative expression is: (3.6) Use the first few components of the Taylor series of f(k) to locate the root of the function in the region of the assumed root and calculate the position of the next iteration point. The specific calculation formula is as follows: Among them, k n They represent the nth solution in the Taylor series, the difference between the known solution and the actual solution; (3.7) NRSR uses the positional k in the iteration process instead. n Replace the non-fitness value f(k n ), search based on the individual's location, NRSR modified the formula to: to n+1 =k n -NRSR (27) Among them, randn is a normally distributed random number, L w is the worst position, L b For the best position; (3.8) NRBO introduces the trap avoidance operator TAO. When the random number rand∈[0,1] is less than 0.6, TAO passes through the optimal position L b With current location Combinations of Adjust the position of the next iteration The specific calculation formula is as follows: Among them, θ1 and θ2 are random numbers between -1 and 1 and between -0.5 and 0.5 respectively; μ1=α×3×rand+(1-α) (31) μ2=α×rand+(1-α) (32) Among them, rand represents a random number between 0 and 1, and Δ represents a number between 0 and 1; since the parameters μ1 and μ2 are random numbers; (3.9) Search for the global optimal solution according to the position update formula and obtain the optimal parameter combination of TIDE and VTransformer models; (3.10) If the maximum number of iterations is not reached, return to step (3.3) until the maximum number of iterations is reached and the optimal parameter configuration is output; the TIDE and VTransformer models use the optimal parameters to configure the prediction components and obtain the prediction results of the high-frequency components and the low-frequency components.

6. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The implementation process of step (4) is as follows: (4.1) The input of the L-th layer encoder of Transformer is the output of the previous layer. Each encoder is used to extract the features of the input data layer by layer, as shown in the formula: U0=Embed(M)+PosEnc(M) (34) Among them, Embed(M) is the embedding map, PosEnc(M) is the position encoding; (4.2) Improve the Transformer and replace the complex decoder in the original Transformer with a linear layer. The input of VTransformer first encodes the position of the data after the embedding layer, and then extracts the information of the required position in the sequence; the formula is as follows: In the formula, 2a represents an even dimension, 2a+1 represents an odd dimension, pos represents the position of the data in the sequence, and d model is the model dimension; (4.3) In the improved VTransformer model, the attention weights are first calculated by the query vector matrix Q and the key vector matrix K, and then the calculated weights are applied to the value vector matrix V to obtain the overall weighted output vector. The calculation formula is: (4.4) The feedforward neural network consists of two fully connected layers and a Relu activation function, and the calculation formula is as follows: FNN(x)=G2max(0,G1x+o1)+o2 (38) Where X is the dimension d model The input vector data, o1 and o2 are bias terms, G1 and G2 are parameter matrices of the linear mapping; (4.5) The prediction result of TiDE is regarded as the joint mapping relationship between the historical time series y, attribute a and dynamic covariate x, which can be expressed as: Among them, K is the lookback time step, H is the prediction step, and Q is the number of time series; (4.6) The encoder part of TiDE extracts time series features through the feature projection module and uses the residual block to transform the input feature matrix of each time step Mapping from the original dimension r to a lower dimensional space The specific formula is as follows: (4.7) In the dense coding part, the dynamic covariate feature projection is flattened, concatenated with the static attributes and historical time series, and fed into the encoder containing multiple residual blocks as input. The specific formula is as follows: (4.8) The dense decoder is encoded with B (i) As input, it is mapped to a vector g of size H×P (i) In the vector g (i) is reshaped into the matrix D (i) ∈R p×H , each column represents a decoding vector of a predicted time step, and the specific calculation formula is as follows: g (i) =Decoder(e (i) )∈R H×p (42) D (i) =Reshape(g (i) )∈R p×H (43) (4.9) The temporal decoder maps the output of the dense decoder to a time series prediction value; the decoded vector of the t-th prediction time step is With the corresponding covariate projection Splicing is performed, and then feature transformation is performed through the residual block to generate the final prediction value The formula is as follows: in, (4.10) An offshore wind power prediction model based on VTransformer and TiDE is established, and the high-frequency component and low-frequency component obtained in step (2) are respectively input into the improved VTransformer prediction model and TiDE prediction model for feature learning and prediction.

7. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The implementation process of using the INRBO algorithm to perform weight fusion on the prediction results of the TIDE and VTransformer models in step (5) is as follows: The obtained high and low frequency component prediction results are used to establish an input matrix; a weight optimization objective function is constructed to minimize the prediction error. The formula is as follows: In the formula, L(w) is the weight result, y j is the true value, N is the number of samples, w i is the weight of each component; Calculate the gradient of the objective function And the Hessian matrix H, the formula is as follows: The INRBO algorithm is used to adaptively optimize the weights of each component to obtain the final wind power prediction result.

8. The offshore wind power prediction method based on multi-model collaborative fusion according to claim 1 is characterized in that: The implementation process of using the extended Kalman filter model EKF to perform error correction on the weighted result in step (5) is as follows: The state estimation value and covariance matrix at the current moment are predicted by the state transfer equation. The calculation formula is as follows: x'=Ux+AS (48) M'=UMU T +Q (49) Among them, x' represents the current estimate, U represents the state transfer matrix, x represents the optimal estimate at the previous moment, A represents the external input matrix, S represents the external state matrix, M' represents the current prior estimate covariance matrix, M represents the optimal estimate covariance matrix at the previous moment, Q represents the covariance matrix of the process; In the updating process, the linearized measurement equation is used to combine the prior estimated covariance matrix with the observation matrix to calculate the Kalman gain at the current moment. The formula is as follows: K=M'H T (HM'H T +R) -1 (50) Where K represents the Kalman gain at the current moment; H represents the observation matrix; The Kalman gain is used to integrate the actual measurement value into the estimation process. With the help of the deviation between the actual measurement value and the current state estimation value, the predicted state estimation value and the covariance matrix are corrected and calibrated to obtain the final prediction result. The update formula is as follows: x=x′+K(z-Hx′) (51) P=(E-KH)M′ (52) Among them, x represents the optimal estimate at the current moment; z represents the actual measurement value; M represents the covariance matrix at the current moment; E represents the identity matrix; EKF continuously corrects the state estimate until a more accurate result is obtained. EKF regards the offshore wind power state as a dynamic system and uses sensor data to iteratively update the wind power. By repeatedly predicting and updating the iterative process, the estimation error is gradually reduced. EKF can obtain a more accurate estimate of the offshore wind power state and obtain the final prediction result.

9. An electronic device, characterized in that: comprising a memory and a processor, wherein: A memory for storing computer programs that can be run on the processor; A processor is used to execute the steps of the offshore wind power prediction method based on multi-model collaborative fusion as described in any one of claims 1 to 8 when running the computer program.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of the offshore wind power prediction method based on multi-model collaborative fusion as described in any one of claims 1 to 8.

Images