Method and apparatus for offshore wind turbine power prediction based on hybrid model

Abstract

The application discloses a kind of based on offshore wind turbine power prediction of mixed model method and device, the method includes: first, offshore wind turbine power data is fitted and analyzed using the SARIMA model trained to obtain power fitting result and residual data;Then using the random forest model trained to carry out offshore wind turbine data optimal data selection, obtain offshore wind turbine power data influence optimal data and its probability;Then the fitting value and residual value obtained by SARIMA model are input into the BIGRU neural network model trained together with offshore wind turbine power unit optimal influence factor and offshore wind turbine power data to predict offshore wind turbine power, and continuously input offshore wind turbine SCADA data to obtain real-time offshore wind turbine power prediction value.The application can realize offshore wind turbine power prediction method and device based on mixed model.

Description

TECHNICAL FIELD

[0001] The application relates to the field of offshore wind turbine power prediction, in particular to a method and device for offshore wind turbine power prediction based on a hybrid model. BACKGROUND

[0002] Most of the existing offshore unit power prediction models are analyzed by combining a neural network method with other methods, the neural network method is combined with other models to consider one or several factors that affect the offshore unit power, or the neural network method is combined with other models to optimize the original method, and in general, the prediction model is established by considering the interaction between the offshore environmental factors and the wind turbine components. Although many factors are considered to affect the unit power, these indicators are reflections of the same thing, which inevitably leads to overlapping information and increases the complexity of the problem, so it is necessary to sort the importance of these factors. The offshore wind turbine power data has obvious seasonal characteristics, in the months with relatively average wind speed and the months with large wind speed variation, the offshore wind turbine power data often presents different change rules, that is, there are linear characteristics and nonlinear characteristics. SUMMARY

[0003] The application aims to provide a method and device for offshore wind turbine power prediction based on a hybrid model, which aims to solve the problem of offshore wind turbine power prediction.

[0004] The application provides a method for offshore wind turbine power prediction based on a hybrid model, which comprises the following steps:

[0005] S1, acquiring historical unit data acquisition and control system data, the historical unit data acquisition and control system data comprising offshore wind turbine power data, average wind speed, average generator speed, average humidity, average temperature, steering angle and unit yawing frequency, and preprocessing the historical unit data acquisition and control system data;

[0006] S2, importing dependent packages through Python, reading and visually displaying the offshore wind turbine power data, drawing autocorrelation graphs and partial autocorrelation graphs through a self-correlation function, and performing stationarity test;

[0007] S3, performing stationarity analysis on the seasonal characteristics of the offshore wind turbine power data, and then performing difference and seasonal difference operations to determine the first SARIMA model parameters, and determining the second SARIMA model parameters through the autocorrelation graph and the partial autocorrelation graph;

[0008] S4, a SARIMA model is established based on the first SARIMA model parameter and the second SARIMA model parameter, fitting analysis is performed on the pretreated offshore wind turbine power data based on the SARIMA model to obtain a fitting result and a corresponding residual value, and a white noise test is performed on the residual by using a Ljung-Box method;

[0009] S5, the pretreated unit data acquisition and control system data are normalized, and the normalized data are input into a random forest model for training to obtain a trained random forest model;

[0010] S6, the trained random forest model is used for factor selection on the offshore wind turbine SCADA data to obtain a selection result, and the selection result includes the number and proportion of the offshore wind turbine power influencing factors;

[0011] S7: the offshore wind turbine power fitting value and the residual value obtained by the SARIMA model and the offshore wind turbine power influencing factor selection result and the offshore wind turbine active power data obtained by the random forest model are spliced to obtain time series data, and the time series data are normalized to obtain normalized data;

[0012] S8: a BIGRU recurrent neural network model is constructed, the normalized data are input into a BIGRU recurrent neural network combined with an attention mechanism for training to obtain a trained BIGRU recurrent neural network combined with an attention mechanism;

[0013] S9: the trained BIGRU recurrent neural network is used for predicting the offshore wind turbine power data to be predicted to obtain output features, and the output features are de-normalized to obtain offshore wind turbine predicted power values.

[0014] The embodiment of the present application also provides a device for offshore wind turbine power prediction based on a hybrid model, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the computer program realizes the steps of the above method when executed by the processor.

[0015] The embodiment of the present application can realize offshore wind turbine power prediction.

[0016] The above description is only a summary of the technical scheme of the present application, in order to more clearly understand the technical means of the present application, the content of the specification is implemented, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the drawings required to be used in the specific embodiments or prior art description will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0018] Figure 1 is a method flowchart of offshore wind turbine power prediction based on a hybrid model according to an embodiment of the present application;

[0019] Figure 2 is a device schematic diagram of offshore wind turbine power prediction based on a hybrid model according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] The technical solutions of the present application will be described below in conjunction with the embodiments, obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0021] Method embodiment

[0022] According to an embodiment of the present application, a method for offshore wind turbine power prediction based on a hybrid model is provided, Figure 1 is a method flowchart of offshore wind turbine power prediction based on a hybrid model according to an embodiment of the present application, as Figure 1 shown, specifically comprising:

[0023] S1, obtaining historical unit data acquisition and control system data, the historical unit data acquisition and control system data including: offshore wind turbine power data, average wind speed, average generator speed, average humidity, average temperature, steering angle and unit yawing frequency, preprocessing the historical unit data acquisition and control system data;

[0024] The preprocessing of the historical unit data acquisition and control system data specifically includes: data division, time series processing and standardization processing of the historical unit data acquisition and control system.

[0025] S2, import the dependent package through Python, read and visually display the offshore wind turbine power data, draw the autocorrelation graph and partial autocorrelation graph through the autocorrelation function, and perform stationarity test;

[0026] S3, the seasonal characteristics of the offshore wind turbine power data are analyzed for stationarity, and then difference and seasonal difference operations are performed to determine the first SARIMA model parameters, and the autocorrelation graph and partial autocorrelation graph are used to determine the second SARIMA model parameters;

[0027] S3 specifically includes: the S3 specifically includes: the offshore wind turbine power data has seasonal characteristics, and after stationarity analysis, difference and seasonal difference operations are performed to determine the seasonal difference order and difference order, and the autocorrelation graph and partial autocorrelation graph are used to determine the seasonal autoregressive parameters, seasonal moving average order, autoregressive parameters and moving average order.

[0028] S4, based on the first SARIMA model parameters and the second SARIMA model parameters, a SARIMA model is established, and based on the SARIMA model, the preprocessed offshore wind turbine power data is fitted and analyzed to obtain a fitting result and corresponding residual value, and the Ljung-Box method is used to test the white noise of the residual.

[0029] S4 specifically includes: based on the seasonal difference order and the seasonal autoregressive parameters, a SARIMA model is established, and the SARIMA model is as follows:

[0030] Wherein, S is the length of the seasonal cycle; represents a seasonal P-order autoregressive operator, and q (B S )μ t represents a seasonal Q-order moving average operator; P represents a seasonal autoregressive parameter; D is a seasonal difference order; Q represents a seasonal moving average order, p represents an autoregressive parameter; d is a seasonal difference order; q represents a seasonal moving average order;

[0031] Based on the SARIMA model, the preprocessed offshore wind turbine power data is fitted and analyzed to obtain a fitting result and corresponding residual value, and the Ljung-Box method is used to test the white noise of the residual.

[0032] S5, the preprocessed unit data acquisition and control system data are normalized, and the normalized data are input into a random forest model for training to obtain a trained random forest model.

[0033] S6, using the trained random forest model, the offshore wind turbine SCADA data is selected for factors to obtain a selection result, and the selection result includes the number of influencing factors of the offshore wind turbine power and the proportion occupied.

[0034] S6 specifically includes:

[0035] The sample is extracted in a random and replaceable manner to generate a sub-data set;

[0036] Each sub-decision tree is independently trained on the generated sub-data set, wherein, when training the sub-decision tree, an optimal partition feature is selected by using the feature information, specifically, the optimal partition feature is selected by obtaining an average impurity reduction value by Gini impurity, and the average impurity reduction value calculation formula is:

[0037]

[0038] T is the result sum at a certain time, P i is the probability of an event;

[0039] The out-of-bag error rate is used for verification and analysis, and finally a trained random forest model is obtained.

[0040] S7: The offshore wind turbine power fitting value and residual value obtained by the SARIMA model and the offshore wind turbine power influencing factor selection result and offshore wind turbine active power data obtained by the random forest model are spliced to obtain time series data, and the time series data is normalized to obtain normalized data;

[0041] The time series data is normalized to obtain normalized data, and the normalization formula is as follows:

[0042]

[0043] Wherein x0 is the original data, x is the normalized data, xmin is the minimum value of the sample, and xmax is the maximum value of the sample.

[0044] S8: A BIGRU recurrent neural network model is constructed, the normalized data is input into the BIGRU recurrent neural network combined with an attention mechanism, and a trained BIGRU recurrent neural network model combined with an attention mechanism is obtained.

[0045] The BIGRU recurrent neural network model is constructed, and specifically includes:

[0046] S61: An input layer and an output layer are constructed, wherein the number of nodes of the input layer is the same as the feature attributes of the time series data to be input, and the output layer includes fitting results and prediction results, the fitting results are used to calculate fitting errors, and the prediction results are offshore wind turbine time series prediction results;

[0047] S62: A hidden layer is constructed, which is a recurrent neural network layer built by using GRU cells;

[0048] The hidden layer is a recurrent neural network layer built with GRU cells, which specifically includes: a single-layer recurrent neural network structure built with GRU cells, the GRU cell contains an update gate, a reset gate and an output gate gating structure, the update gate is used to control the update of the hidden state, the reset gate decides whether to ignore the previous hidden state, the input xt at t time and the output ht-1 at t-1 time are concatenated to form the input of the GRU cell, and the output of the GRU cell is ht t-1 are input into the cell for calculation; the forward propagation calculation formula is as follows:

[0049]

[0050] h t-1 represents the hidden layer information at the previous moment, x t represents the current input vector, sigma is a sigmoid function, W xu and W xr are weight matrices of the reset gate and the update gate respectively; is a dot product operation, W is a weight matrix, and the small subscript represents different stages, b u , b r , and b h are bias vectors, the data obtained after the neurons at the middle moment and x t are concatenated is substituted into tanh to obtain

[0051] S63: The fitting result, the prediction result and the historical offshore wind turbine power data are taken as training data to perform parameter optimization, a 3-layer neural network structure is determined for the hidden layer, one layer is a recurrent BIGRU neural network layer, one layer is an Attention layer, and the third layer is a Dense layer, and the number of training times is determined; a loss function is defined, the gradient descent method adopts the adam descent method, and a trained BIGRU recurrent neural network model is obtained.

[0052] The normalized data is input into the BIGRU recurrent neural network combined with the attention mechanism for training, which specifically includes:

[0053] The BIGRU cell output is calculated by forward propagation;

[0054] The error term of each BIGRU cell is calculated in reverse, and the error is propagated in the time and network level directions; then, the gradient of each weight is calculated according to the corresponding error term;

[0055] The weights are updated based on the gradient, wherein the mean absolute error is selected as the error calculation method, and the loss function in the training process is:

[0056]

[0057] y i is the predicted value, x iThat is the input value, h(x i ) is the neural unit after input.

[0058] S9: The trained BIGRU recurrent neural network is used for predicting the to-be-predicted offshore wind turbine power data to obtain output features, and the output features are subjected to inverse normalization processing to obtain offshore wind turbine predicted power numerical values.

[0059] The specific implementation method is as follows:

[0060] The application provides an offshore wind turbine power prediction method and device based on a SARIMA-RF-BIGRU hybrid model, to solve or at least partially solve the technical problem of inaccurate prediction results in the prior art.

[0061] To solve the above technical problems, the first aspect of the application provides an offshore wind turbine power prediction method based on a factor selection and residual optimization hybrid neural network model, comprising:

[0062] S1: Obtain offshore wind turbine historical data acquisition and control system (SCADA) data, the SCADA data including 84-dimensional data such as offshore wind turbine active power, reactive power, average wind speed, average generator speed, average humidity, average temperature, steering angle, and unit yawing frequency, and pre-process the SCADA data;

[0063] S2: Import dependent packages through Python, read offshore wind turbine power data for visual display, draw autocorrelation and partial autocorrelation graphs through an autocorrelation function (ACF), and perform stationarity test.

[0064] S3: The offshore wind turbine power data has seasonal characteristics, and after stationarity analysis, difference and seasonal difference operations are performed to determine the difference number D, and PACF and ACF graphs are obtained to determine parameters (P, Q), and s represents the seasonal length.

[0065] S4: The determined SARIMA (p, d, q) (P, D, Q) s is used for fitting analysis of offshore wind turbine power data, and the mathematical expression of the SARIMA model is:

[0066]

[0067] Wherein, S is the seasonal period length; represents a seasonal P-order autoregressive operator, and q (B S )mu tQ represents the seasonal moving average order; P represents the seasonal autoregressive parameter; D is the seasonal difference order; Q represents the seasonal moving average order.

[0068] The fitting result and the corresponding residual value are obtained, and the Ljung-Box method is used to perform white noise test on the residual.

[0069] S5: The obtained offshore wind turbine SCADA data of a total of 84 dimensions are normalized.

[0070] S6: The normalized data are input into a random forest model for training, and a trained random forest model is obtained.

[0071] S7: The trained random forest model is used for factor selection of the offshore wind turbine SCADA data, and a selection result is obtained, the selection result including the number and proportion of the offshore wind turbine power influencing factors.

[0072] S8: The offshore wind turbine power fitting value and the residual value obtained by the SARIMA model, the offshore wind turbine power influencing factor selection result obtained by the random forest model, and the offshore wind turbine active power data are spliced, and the spliced data are normalized.

[0073] S9: The normalized data are input into a BIGRU recurrent neural network combined with an attention mechanism for training, and a trained BIGRU recurrent neural network model combined with an attention mechanism is obtained.

[0074] S10: The trained BIGRU recurrent neural network is used to predict the offshore wind turbine power data, and offshore wind turbine predicted power values are obtained.

[0075] In an embodiment, the historical SCADA information is preprocessed in S1, including: data division, time series processing and standardization processing of the historical SCADA information.

[0076] In an embodiment, the Python imports the statsmodels database in S2, which provides functions for establishing a SARIMA model. It includes:

[0077] S2.1: Correlation measurement is performed on the offshore unit power data time series, and the AutoCorrelation Function (ACF) expression is:

[0078]

[0079] S2.2: Unit root ADF test is performed on the offshore wind turbine power data to test the stationarity of the original data. The ADF test result significantly rejects the null hypothesis, indicating that the time series is stationary, otherwise, it is a non-stationary sequence and needs to be differentiated.

[0080] In an embodiment, S7 specifically comprises:

[0081] S7.1: Samples are extracted in a random with replacement manner to generate a sub-data set;

[0082] S7.2: Each sub-decision tree is independently trained on the generated sub-data set, wherein, when training the sub-decision tree, the optimal partition feature is selected by using the feature information, specifically, the optimal partition feature is selected by the average impurity reduction value obtained by Gini impurity, wherein the average impurity reduction value calculation formula is:

[0083]

[0084] S7.3: The out-of-bag error rate is used for verification and analysis, and finally the trained random forest model is obtained.

[0085] In an embodiment, the preprocessed historical offshore wind turbine SCADA information is time series data, which includes feature attributes, and S9 specifically comprises:

[0086] S9.1: Parameter optimization is performed by network search method, and it is determined that the model adopts a 3-layer neural network structure, wherein one layer is a cyclic BIGRU neural network layer, one layer is an Attention layer, and the third layer is a Dense layer. The number of neurons in each layer is 50, 128, and 2 respectively. The number of training times is determined, the mini-batch number is 128, and the gradient descent method adopts the adam descent method.

[0087] S9.2: Model training, the power time series data obtained above including power residual value, average wind speed, average generator speed value, offshore wind turbine power value, is used as all data samples, and all data is divided into training set and test set, wherein the training set data is 4 / 5, and the test set test data is 1 / 5.

[0088] S9.3: The sample is normalized, and the normalization formula is:

[0089]

[0090] Wherein x0 is the original data, and x is the normalized data.

[0091] S9.1 specifically comprises:

[0092] S9.1.1: Forward propagation calculation of BIGRU cell output;

[0093] S9.1.2: Reverse calculation of error term of each BIGRU cell, reverse propagation according to time and network level in two directions;According to the corresponding error term, the gradient of each weight is calculated;

[0094] S9.1.3: Gradient-based optimization algorithm updates the weight, wherein the mean absolute error is selected as the error calculation method, and the loss function in the training process is:

[0095]

[0096] S9.4: The test set is input to the trained cyclic BIGRU neural network to predict, and the predicted data is de-normalized to obtain real offshore wind turbine data, and the de-normalization formula is:

[0097]

[0098] The offshore wind turbine power prediction method based on the SARIMA-RF-BIGRU hybrid model provided by the application firstly uses the trained SARIMA model to analyze the offshore wind turbine power data to obtain the power fitting result and residual data;Then the trained random forest model is used to select the optimal data of offshore wind turbine data, and the optimal data and its probability of offshore wind turbine power data are obtained;The fitting value and residual value obtained by the SARIMA model are input into the trained BIGRU neural network model together with the optimal influencing factors of offshore wind turbine power and offshore wind turbine power data to predict the offshore wind turbine power, and the real-time offshore wind turbine power prediction value is obtained by continuously inputting the offshore wind turbine SCADA data, compared with the general trend analysis result in the prior art, which can only get the reliability lower and lower with the longer use time, the model prediction result is more accurate with the increase of data amount, which improves the prediction accuracy, so that the offshore wind turbine power risk warning can be improved, and timely prevention and loss prevention can be realized.

[0099] Device embodiment one

[0100] The embodiment of the application provides a device for offshore wind turbine power prediction based on a hybrid model, as shown in Figure 2 The device comprises a memory 20, a processor 22, and a computer program stored in the memory 20 and executable on the processor 22, and the computer program realizes the steps in the method embodiments when executed by the processor.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the present application.

Claims

1. A method for predicting the power of offshore wind turbines based on a hybrid model, characterized in that, include: S1. Acquire historical unit data acquisition and control system data, which includes: offshore wind turbine power data, average wind speed, average generator speed, average humidity, average temperature, steering angle, and number of yaws of the unit. Preprocess the historical unit data acquisition and control system data. S2. Import the dependent packages using Python, read and visualize the power data of offshore wind turbines, draw autocorrelation and partial autocorrelation plots using the autocorrelation function, and perform stationarity checks. S3. After performing stationarity analysis on the seasonal characteristics of offshore wind turbine power data, differential and seasonal differential operations are performed to determine the parameters of the first SARIMA model, and the parameters of the second SARIMA model are determined through autocorrelation plots and partial autocorrelation plots. S4. Establish a SARIMA model based on the parameters of the first SARIMA model and the parameters of the second SARIMA model. Based on the SARIMA model, perform fitting analysis on the preprocessed offshore wind turbine power data to obtain the fitting results and corresponding residual values. Use the Ljung-Box method to perform white noise verification on the residuals. S5. Normalize the preprocessed unit data acquisition and control system data, and input the normalized data into the random forest model for training to obtain the trained random forest model. S6. Use the trained random forest model to select factors for the SCADA data of offshore wind turbines and obtain the selection results. The selection results include the number and proportion of factors affecting the power of offshore wind turbines. S7: The power fitting values ​​and residual values ​​of offshore wind turbines obtained from the SARIMA model, the selection results of the power influencing factors of offshore wind turbines obtained from the random forest model, and the active power data of offshore wind turbines are spliced ​​together to obtain time series data. The time series data is then normalized to obtain normalized data. S8: Construct a BIGRU recurrent neural network model. Input normalized data into the BIGRU recurrent neural network with attention mechanism and train it to obtain a trained BIGRU recurrent neural network model with attention mechanism. S9: The trained BIGRU recurrent neural network is used to predict the power data of the offshore wind turbine to be predicted and the output features are obtained. The output features are then inversely normalized to obtain the predicted power value of the offshore wind turbine.

2. The method according to claim 1, characterized in that, The preprocessing of historical unit data acquisition and control system data specifically includes: data segmentation, time series processing, and standardization processing of historical unit data acquisition and control system data.

3. The method according to claim 2, characterized in that, S3 specifically includes: for the seasonal characteristics of offshore wind turbine power data, after performing stationarity analysis, performing difference and seasonal difference operations to determine the seasonal difference order and the difference order, and determining the seasonal autoregressive parameter, the seasonal moving average order, the autoregressive parameter, and the moving average order through autocorrelation plots and partial autocorrelation plots.

4. The method according to claim 3, characterized in that, S4 specifically includes: establishing a SARIMA model based on the seasonal difference order and seasonal autoregressive parameters, the SARIMA model being as follows: Where S is the length of the seasonal cycle; Represents the seasonal P-order autoregressive operator, Θ q (B S )μ t This represents the seasonal Q-order moving average operator; P represents the seasonal autoregressive parameter; D is the seasonal differencing order; Q represents the seasonal moving average order; p represents the autoregressive parameter; d is the seasonal differencing order; q represents the seasonal moving average order. The preprocessed offshore wind turbine power data was fitted and analyzed using the SARIMA model to obtain the fitting results and corresponding residual values. The Ljung-Box method was then used to perform white noise testing on the residuals.

5. The method according to claim 4, characterized in that, S6 specifically includes: A subset of data is generated by randomly selecting samples with replacement. Each sub-decision tree is trained independently on the generated subset dataset. During training, the optimal splitting feature is selected using feature information, specifically by using the average impurity reduction value obtained from the Gini impurity to select the optimal splitting feature. The formula for calculating the average impurity reduction value is as follows: T represents the sum of results at a certain time, and P represents the sum of results at a certain time. i The probability of a certain event; The out-of-bag error rate was used for verification and analysis, and the trained random forest model was finally obtained.

6. The method according to claim 5, characterized in that, The time series data is normalized to obtain normalized data. The normalization formula is as follows: Where x0 represents the original data, x represents the normalized data, and x0 represents the normalized data. min x is the minimum value of the sample. max This represents the maximum value of the sample.

7. The method according to claim 6, characterized in that, The construction of the BIGRU recurrent neural network model specifically includes: S61: Construct an input layer and an output layer. The number of nodes in the input layer is the same as the feature attributes of the time series data to be input. The output layer contains the fitting result and the prediction result. The fitting result is used to calculate the fitting error, and the prediction result is the time series prediction result of the offshore wind turbine. S62: Construct a hidden layer, wherein the hidden layer is a recurrent neural network layer built using GRU cells; S63: Using the obtained fitting results, prediction results, and historical offshore wind turbine power data as training data, perform parameter optimization, determine that the hidden layer adopts a 3-layer neural network structure, including one recurrent BIGRU neural network layer, one attention layer, and the third dense layer, and determine the number of training iterations; define the loss function, and adopt the adam descent method for gradient descent to obtain the trained BIGRU recurrent neural network model.

8. The method according to claim 7, characterized in that, The hidden layer is a recurrent neural network layer built using GRU cells, specifically comprising: a single-layer recurrent neural network structure built using GRU cells, wherein each GRU cell includes an update gate, a reset gate, and an output gate gating structure. The update gate controls the update of the hidden state, and the reset gate determines whether to ignore the previous hidden state, taking the input x at time t as... t and the output h at time t-1 t-1 Input the cells together for calculation; the forward propagation calculation formula is as follows: h t-1 x represents the hidden information from the previous time step. t Let W represent the current input vector, σ be the sigmoid function, and W be the input vector. xu and W xr , respectively, are the weight matrices for the reset gate and the update gate; ⊙ represents the dot product operation.

9. The method according to claim 8, characterized in that, The training of the BIGRU recurrent neural network by inputting normalized data and combining it with an attention mechanism specifically includes: Forward propagation calculates the output of the BIGRU cell; The error term for each BIGRU cell is calculated in reverse, and propagated backward in both time and network layer directions; then, the gradient of each weight is calculated based on the corresponding error term. Gradient-based optimization algorithms update weights, using mean absolute error as the error calculation method. The loss function during training is:

10. A device for predicting the power of offshore wind turbines based on a hybrid model, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for predicting the power of offshore wind turbines based on a hybrid model as described in any one of claims 1 to 9.

Images