Floating fan motion response real-time forecasting method based on time sequence dynamic decomposition
By performing time-series dynamic decomposition and neural network prediction on the motion response data of floating wind turbines, the problems of large computing resource consumption and low efficiency of traditional methods are solved, efficient real-time prediction of floating wind turbine motion responses is achieved, and the reliability and safety of the system are improved.
Patent Information
- Application Number
- CN202510858607.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-25
- Publication Date
- 2025-10-10
AI Technical Summary
Traditional methods consume huge computational resources and are inefficient when predicting the motion response of floating wind turbines. They cannot meet the needs of real-time offshore forecasting, and the model has poor generalization performance, resulting in inaccurate prediction results.
A method based on time series dynamic decomposition is used to decompose the motion response data of floating wind turbines into trend components and fluctuation components. Prediction is performed using CNN and GRU neural network models, and a prediction model library that adapts to different sea conditions is constructed to achieve real-time forecasting.
The calculation efficiency and forecast accuracy are improved, and the online real-time forecast of the motion response of floating wind turbines can be realized on the basis of ensuring the forecast accuracy, thereby enhancing the reliability and safety of the system.
Smart Images
Figure CN120764348A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of offshore floating wind power, and in particular to a real-time prediction method for motion response of a floating wind turbine based on time series dynamic decomposition. Background Art
[0002] Floating wind power systems are essential marine equipment for developing advantageous deep-sea wind resources. They consist of wind turbines, towers, floating platforms, mooring systems, and dynamic cables. Under the combined influence of wind, waves, and currents in the ocean environment, floating wind turbines can generate irregular, random motions in six degrees of freedom, increasing fatigue loads on the system, reducing power generation efficiency, and threatening system safety. Therefore, real-time prediction of floating wind turbine motion response not only provides advance information for floating wind turbine motion suppression systems but also provides early warning of floating wind turbine operating status, enhancing system reliability and safety.
[0003] Traditional methods typically use integrated coupled simulation to predict the motion response of floating wind turbines. However, the coupled dynamic characteristics of floating wind turbine systems are complex, and integrated coupled simulation requires solving complex aerodynamic, hydrodynamic, and structural dynamics governing equations. The numerical modeling involves numerous parameters and a complex process. This results in significant computational resource consumption and low computational efficiency, making it inadequate for real-time offshore forecasting applications. Summary of the Invention
[0004] In response to the above-mentioned problems and technical needs, this application proposes a real-time prediction method for the motion response of a floating wind turbine based on time series dynamic decomposition. The technical solution of this application is as follows:
[0005] A real-time prediction method for the motion response of a floating wind turbine based on time series dynamic decomposition includes the following steps:
[0006] Determine a set of sea conditions at each level in the target operating sea area, and obtain training samples of the motion response of the floating wind turbine under each set of sea conditions at each level; each set of sea conditions includes multiple sea conditions, and the training samples of the motion response of the floating wind turbine under each sea condition include a historical time series consisting of motion response data at N consecutive moments and a predicted time series consisting of motion response data at M consecutive moments after the N consecutive moments; N and M are integer parameters and N>M;
[0007] The network architecture of the floating wind turbine motion response prediction model is constructed by sequentially cascading a time series dynamic decomposition module, an image encoding module, a feature extraction module, and a prediction module. The historical time series of motion response training samples of each level of sea condition set is used as input, and the predicted time series of the motion response training samples is used as output to train the floating wind turbine motion response prediction model corresponding to each level of sea condition set; the time series dynamic decomposition module decomposes the historical time series of the motion response training samples into a trend component and a fluctuation component, the image encoding module encodes the trend component and the fluctuation component respectively to obtain a trend image and a fluctuation image, the feature extraction module extracts the features of the trend image and the fluctuation image respectively to obtain a trend feature vector and a fluctuation feature vector, the prediction module extracts the long-term time dependency in the trend feature vector and the fluctuation feature vector respectively and predicts the motion response data for the next M moments; the trend component represents the long-term evolution of the historical time series, and the fluctuation component represents the short-term fluctuation of the historical time series;
[0008] Using the environmental monitoring system to measure the current sea conditions of the floating wind turbine in real time, and determining the floating wind turbine motion response prediction model corresponding to the sea condition set of the current sea conditions;
[0009] The motion response data of the floating wind turbine at the current moment and the N previous historical moments are obtained to form the historical time series to be predicted. The motion response prediction model of the floating wind turbine corresponding to the sea condition set of the current sea condition is used to predict the motion response data of the floating wind turbine at the next M moments based on the historical time series to be predicted.
[0010] A further technical solution is to decompose the historical time series of motion response training samples into trend components and fluctuation components, including:
[0011] Historical time series of motion response training samples Perform data expansion and use the time series window to perform sliding average calculation on the historical time series after data expansion to obtain the trend component
[0012] The historical sequence and trend component The fluctuation component is obtained by subtraction Trend Component and the fluctuation component Length and historical sequence are of equal length.
[0013] Its further technical solution is that the feature extraction module includes two CNN sub-networks with the same structure arranged in parallel, and the trend image and the fluctuation image are respectively input into the two CNN sub-networks to obtain the trend feature vector and the fluctuation feature vector; each CNN sub-network includes a convolution layer, a pooling layer and an average pooling layer cascaded in sequence, the convolution layer is used to extract the features of the input image and mine the correlation between the features of the input image to obtain a feature map, the pooling layer is used to downsample the feature map to reduce the dimension of the feature map, and the average pooling layer is used to compress the feature map after dimensionality reduction into a feature vector of length N.
[0014] Its further technical solution is that the prediction module includes two hidden layers with the same structure arranged in parallel and an output layer, the trend feature vector and the fluctuation feature vector are respectively input into the two hidden layers to obtain the trend hidden state and the fluctuation hidden state, the trend hidden state and the fluctuation hidden state are transmitted to the output layer to obtain the motion response data of the next M moments; each hidden layer includes multiple cascaded GRU units, the output layer adopts a fully connected layer, the hidden layer is used to capture the long-term time dependency in the feature vector to obtain the hidden state, and the output layer adds the two hidden states and maps them into the motion response data of the next M moments.
[0015] A further technical solution is that the image encoding module uses the Gram angular field method to perform polar coordinate conversion on the trend component and the fluctuation component respectively and calculates the Gram matrix to obtain a trend image and a fluctuation image with a resolution of N×N.
[0016] A further technical solution is to decompose the historical time series of motion response training samples into trend components including:
[0017] In historical time The first data Previously, the predetermined number d1 was increased In historical time The last data Then increase the predetermined number d2
[0018] Align the starting position of the time series window with the starting position of the historical time series after data expansion, calculate the average value of each data covered by the time series window to obtain the first data of the trend component, and slide the time series window backward by one data until the end position of the time series window is aligned with the end position of the historical time series after data expansion to obtain the trend component.
[0019] Its further technical solution is that the length of the time series window is w, The first data Previously increased The number of d1=w-1-(w-1) / / 2; in the historical time series The last data Then increase The number of d2 = (w-1) / / 2; / / represents rounding after division.
[0020] A further technical solution is that obtaining motion response training samples includes:
[0021] A time-domain integrated coupled numerical simulation model for floating wind turbines was established. The sea conditions of various levels of sea condition sets were input into the time-domain integrated coupled numerical simulation model for numerical simulation calculations to obtain the motion response data of floating wind turbines under different levels of sea condition sets.
[0022] Normalizing the motion response data under each level of sea condition set, and randomly cutting the normalized motion response data into predetermined lengths to obtain motion response data segments;
[0023] The motion response data segments are divided into training set data segments, validation set data segments and test set data segments. The sliding window method is used to slide the training set data segments, validation set data segments and test set data segments into multiple historical time series and prediction time series to obtain motion response training samples.
[0024] A further technical solution is to determine the sea condition sets of each level including:
[0025] Various sea conditions for floating wind turbines are determined based on hydrological observation data of the target operating sea area, and the sea conditions include significant wave heights; various sea conditions are divided into different levels of sea condition sets based on the significant wave heights.
[0026] A further technical solution is to train a floating wind turbine motion response prediction model corresponding to any level of sea condition set, including:
[0027] Initialize the model parameters, divide the motion response training samples of the sea condition set into multiple groups of sub-datasets, and the number of motion response training samples in each sub-dataset is the same; use each group of sub-datasets to train the model in turn, calculate the prediction error of each model training and adjust the model parameters through the optimizer until the prediction error of the model training no longer decreases in a predetermined number of consecutive rounds of model training, or when the maximum number of rounds of model training is reached, train to obtain a candidate prediction model corresponding to the sea condition set, change the model hyperparameters and re-train the model to obtain multiple candidate prediction models corresponding to the sea condition set, and use the candidate prediction model with the smallest prediction error as the floating wind turbine motion response prediction model corresponding to the sea condition set; use the mean square error of the predicted motion response data and the predicted time series of the motion response training samples as the prediction error.
[0028] The beneficial technical effects of this application are:
[0029] This application proposes a real-time prediction method for the motion response of floating wind turbines based on time-series dynamic decomposition. Compared with the traditional integrated coupling simulation calculation method, which requires the construction of a complex integrated coupling numerical simulation model, resulting in a huge amount of calculation and serious time consumption. The method of this application is simple and flexible. During the actual operation of floating wind turbines at sea, it can achieve online real-time prediction of motion response while ensuring the accuracy of the prediction. The prediction results can provide advance information for the floating wind turbine motion suppression system and provide a basis for floating wind turbine motion status warning, which is conducive to improving the reliability and safety of the floating wind turbine system.
[0030] In view of the strong time dependence and complex frequency characteristics of floating wind turbine motion response data time series, the long-term evolution and short-term fluctuations of floating wind turbine motion response data time series are analyzed, and the time series dynamic decomposition method is used to decompose the historical time series into trend components and fluctuation components, which helps to separate and extract the intrinsic frequency characteristics of the time series. The one-dimensional time series data is further encoded into a two-dimensional image using the Gramian Angular Field (GAF) method. While retaining the temporal correlation of the original sequence data, the data features are enriched to a greater extent, which helps the neural network model better capture the patterns and data correlations in the complex floating wind turbine motion response data time series. Moreover, by constructing a neural network model that integrates CNN and GRU and setting up dual-channel parallel branches, the local feature information and long-term dependencies in the trend image and fluctuation image are captured respectively. Compared with the traditional cascade structure, the mutual influence of trend and fluctuation characteristics is avoided, effectively improving computational efficiency and prediction accuracy.
[0031] This application analyzes the actual sea conditions in the target operating area. Taking into account the significant differences in the motion responses of floating wind turbines under different sea conditions, the application divides the sea conditions into different levels of sea condition sets and constructs a floating wind turbine motion response prediction model for each sea condition set. This can further improve the accuracy of the prediction results. This provides greater flexibility and adaptability in predicting the motion response of floating wind turbines stimulated by time-varying environmental conditions at sea. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 It is a flow chart of the real-time prediction method of the motion response of a floating wind turbine.
[0033] Figure 2 This is a schematic diagram of the segmentation of motion response data.
[0034] Figure 3 This is the network architecture diagram of the floating wind turbine motion response prediction model.
[0035] Figure 4is a schematic diagram of encoding a fluctuation component to obtain a fluctuation image.
[0036] Figure 5 is a structure diagram of a CNN subnetwork.
[0037] Figure 6 is a structure diagram of a GRU unit.
[0038] Figure 7 is a comparison diagram of prediction results of an embodiment. DETAILED DESCRIPTION
[0039] The specific embodiments of the present application will be further described below in conjunction with the accompanying drawings.
[0040] The present application proposes a floating wind turbine motion response real-time prediction method based on time series dynamic decomposition, please refer to the flowchart shown in Figure 1 The specific steps of the method are as follows:
[0041] Step 1, determine the sea state condition set of each level of the target operation sea area, and obtain the motion response training sample of the floating wind turbine under different sea state condition sets; each sea state condition set includes multiple sea state conditions, and the motion response training sample of the floating wind turbine under each sea state condition includes a historical time series composed of N continuous time motion response data and a prediction time series composed of M continuous time motion response data after N continuous time; N and M are integer parameters and N>M.
[0042] The sea state conditions in the actual marine environment are complex and changeable. Based on the previous experimental research, it is found that the motion response of the floating wind turbine under different sea state conditions is quite different. If all the sea state conditions are integrated into one floating wind turbine motion response prediction model, the generalization performance of the model will be poor, resulting in low accuracy of the prediction results under each sea state condition. Therefore, in order to improve the generalization ability of the model, the present application divides the various sea state conditions of the target operation sea area into different sea state condition sets, and constructs a model respectively. In order to ensure the accuracy of the floating wind turbine motion response prediction model obtained by training, it is necessary to ensure that the motion response training sample in each sea state condition set obtained by division has similar features, so the accuracy of sea state condition set division determines the accuracy of the model.
[0043] In one embodiment, determining the sea state condition set of each level includes: determining various sea state conditions of the floating wind turbine according to the hydrological observation data of the target operation sea area, the sea state conditions including significant wave height; dividing the various sea state conditions into different sea state condition sets according to the significant wave height.
[0044] Since the changes in the sea conditions in the target operating sea area have certain regularities, the changes in the sea conditions in the target operating sea area can be obtained by searching the hydrological observation data of the target operating sea area. The sea conditions recorded in the hydrological observation data include various parameters of wind, wave and current, including significant wave height, environmental incident direction, wind speed, spectral peak period, spectral peak factor and flow rate, etc. Each sea condition corresponds to a set of wind, wave and current parameters. According to historical data and experience, it can be found that the significant wave height has a greater impact on the motion response of floating wind turbines. Therefore, this application classifies the sea conditions with significant wave heights in the range of 0-0.1m into the first-level sea condition set, the sea conditions with significant wave heights in the range of 0.1-0.5m into the second-level sea condition set, the sea conditions with significant wave heights in the range of 0.5-1.25m into the third-level sea condition set, and the sea conditions with significant wave heights in the range of 1.25-2.5m into the fourth-level sea condition set. The sea conditions with significant wave heights ranging from 2.5 to 4 m are classified into the fourth-level sea condition set, the sea conditions with significant wave heights ranging from 4 to 6 m are classified into the sixth-level sea condition set, the sea conditions with significant wave heights ranging from 6 to 9 m are classified into the seventh-level sea condition set, the sea conditions with significant wave heights ranging from 9 to 14 m are classified into the eighth-level sea condition set, and the sea conditions with significant wave heights exceeding 14 m are classified into the ninth-level sea condition set.
[0045] Based on the sea conditions in the sea condition set of each level, the motion response data of the floating wind turbine at each moment can be calculated, and then the motion response training samples can be constructed based on the motion response data at each moment. In one embodiment, obtaining the motion response training samples includes:
[0046] (1) According to the design scheme of the floating wind turbine, a time-domain integrated coupled numerical simulation model of the floating wind turbine consisting of a wind turbine group, a tower, a wind turbine platform and a mooring anchor system is established using numerical software such as OrcaFlex, OpenFAST, Bladed or SIMA. Each sea condition of each level of sea condition set is input into the time-domain integrated coupled numerical simulation model for numerical simulation calculation to obtain the motion response data of the floating wind turbine under the different levels of sea condition set; the numerical simulation duration of this application is 3 hours, and the motion response data of all sea conditions in each level of sea condition set constitutes the motion response database of the sea condition set of that level.
[0047] (2) The motion response data under each level of sea condition set are normalized, and the normalized motion response data are randomly cut into a predetermined length to obtain motion response data segments.
[0048] The motion response data in the motion response database of each sea state condition set is preprocessed. First, the motion response data is normalized. The original motion response data x = {x1, x2, …, xN} is normalized to the interval [-1, 1], and the normalization formula is: n
[0049]
[0050] where x i is the motion response data of the i-th calculation time step, is the normalized motion response data of the i-th calculation time step, and n is the maximum calculation time step of the numerical simulation.
[0051] Then, the motion response data corresponding to each sea state condition after normalization is cut and a data segment of a specified length is extracted. Since the occurrence of sea state conditions in the actual marine environment has a certain randomness, a certain randomness is also ensured when constructing the motion response training sample, wherein the specified length is selected according to the actual application requirement. Specifically, 9000 seconds of motion response data segments are randomly extracted from the motion response data obtained by 3-hour numerical simulation of each sea state condition for constructing the motion response training sample.
[0052] (3) The motion response data segments are divided into training set data segments, validation set data segments and test set data segments, and the training set data segments, the validation set data segments and the test set data segments are respectively divided into a plurality of historical time series and prediction time series by using the sliding window method, to obtain the motion response training sample.
[0053] According to the time sequence, the motion response data segments are divided into training set data segments, validation set data segments and test set data segments in the ratio of 8:1:1, which correspond to the training set, the validation set and the test set of the model training, respectively. The training set data segments correspond to the motion response data in the range of 1-7200 seconds, the validation set data segments correspond to the motion response data in the range of 7201-8100 seconds, and the test set data segments correspond to the motion response data in the range of 8101-9000 seconds. The specific division result is shown in Figure 2
[0054] Further, the training set data segments, the validation set data segments and the test set data segments divided are divided into a series of sub-time series with a length of N+M by using the sliding window method. Each sub-time series is composed of a historical time series with a length of N and a prediction time series with a length of M. Each sea state condition corresponds to a series of sub-time series. The sub-time series corresponding to the sea state conditions belonging to the same sea state condition set constitute the motion response training sample of the sea state condition set. The window size of the sliding window method is N+M. The values of N and M are set according to the actual application requirement. According to experience, N:M = 3:1 is selected, and the sliding window step is 1.
[0055] Step 2: sequentially cascade the temporal dynamic decomposition module, image encoding module, feature extraction module and prediction module to construct the network architecture of the floating wind turbine motion response prediction model. The historical temporal sequence of motion response training samples of each level of sea condition set is used as input, and the predicted temporal sequence of motion response training samples is used as output to train the floating wind turbine motion response prediction model corresponding to each level of sea condition set. The network architecture of the floating wind turbine motion response prediction model is as follows: Figure 3 As shown, after the historical time series of motion response training samples are input into the floating wind turbine motion response prediction model, the time series dynamic decomposition module decomposes the historical time series of motion response training samples into trend components and fluctuation components. The image encoding module encodes the trend component and fluctuation component to obtain trend images and fluctuation images respectively. The feature extraction module extracts the features of the trend image and fluctuation image respectively to obtain trend feature vectors and fluctuation feature vectors. The prediction module extracts the long-term time dependencies in the trend feature vectors and fluctuation feature vectors respectively and predicts the motion response data for the next M moments.
[0056] In the time domain, there is a long-term temporal correlation between the data at each moment of the historical time series of the motion response training sample, and the data at each moment also has a short-term fluctuation characteristic. The trend component obtained by decomposition represents the long-term evolution of the historical time series, and the fluctuation component represents the short-term fluctuation of the historical time series. Correspondingly, in the frequency domain, the trend component represents the low-frequency component of the historical time series, and the fluctuation component represents the high-frequency component of the historical time series. In order to fully extract the intrinsic frequency characteristics of the time series, the present application decomposes the historical time series into a trend component and a fluctuation component according to the time sequence of each motion response data in the historical time series of the motion response training sample.
[0057] In one embodiment, the specific method of decomposing the historical time series of motion response training samples into trend components and fluctuation components is as follows:
[0058] In order to ensure that the lengths of the decomposed trend component and the fluctuation component are consistent with the historical time series, so as to ensure that no data feature is lost, it is first necessary to calculate the historical time series of the motion response training samples. Perform data expansion, and then use the time series window to perform sliding average calculation on the historical time series after data expansion to obtain the trend component By calculating the average of the time series window, we can analyze the overall level and change trend of the time series data, which helps to extract features such as periodicity in the time series data.
[0059] In one embodiment, decomposing the historical time series of motion response training samples into trend components includes:
[0060] (1) Conventional data expansion methods usually add 0 at the beginning and end of the time series data. In order to improve the continuity and smoothness of the two ends of the time series, this application uses the values at the beginning and end for expansion. Specifically, in the historical time series The first data Previously, the predetermined number d1 was increased In historical time The last data Then increase the predetermined number d2 By expanding the data that is the same as the first and last data at the beginning and end of the historical time series, the edge information of the time series can be better preserved.
[0061] The number of data expanded at the beginning and end of the historical time series is determined by the size of the time series window. The first data Previously increased The number of d1=w-1-(w-1) / / 2; in the historical time series The last data Then increase The number of data expanded at the beginning and the end is 11.
[0062] Align the starting position of the time series window with the starting position of the historical time series after data expansion, calculate the average value of each data covered by the time series window to obtain the first data of the trend component, slide the time series window backward by one data, continue to calculate the average value of each data covered by the time series window to obtain the second data of the trend component, and continue to slide the time series window backward by one data until the end position of the time series window is aligned with the end position of the historical time series after data expansion, and sequentially organize each average value into a trend component of length N.
[0063] (2) Historical chronology and trend component The fluctuation component is obtained by subtraction Trend Component and the fluctuation component Length and historical sequence are equal in length, both are of length N, and
[0064] The image coding module is further used to encode the trend component and the fluctuation component respectively, and the one-dimensional time series data is encoded into a two-dimensional image, which further expands the characteristics of the original time series data, can better capture the correlation in the motion response data, and help improve the model's prediction performance for time series.
[0065] In one embodiment, the image encoding module uses the Gram angular field method to perform polar coordinate transformation on the trend component and the fluctuation component respectively and calculates the Gram matrix to obtain a trend image and a fluctuation image with a resolution of N×N.
[0066] Take the fluctuation component as an example to illustrate. First, the fluctuation component Each data is converted into polar coordinates. For any i-th data The polar coordinates obtained by conversion for:
[0067]
[0068] in, Indicates the i-th data Polar coordinate angle, r i Indicates the i-th data The polar coordinate radius of .
[0069] Then, based on the fluctuation component The polar coordinates of each data in the Gram matrix G, G∈R N×N , the Gram matrix G corresponds to the GAF image with a resolution of N×N in the two-dimensional plane, that is, the fluctuation image, and the fluctuation component is encoded to obtain the fluctuation image as follows Figure 4 As shown. Any matrix element G in the Gram matrix G ij The calculation formula is Indicates the jth data Polar coordinate angle of , integer parameter 1≤j≤N.
[0070] The trend image and the fluctuation image are input to the feature extraction module to extract their respective features. In order to achieve the separate processing of the trend image and the fluctuation image, a dual-channel structure needs to be designed. In one embodiment, the feature extraction module includes two CNN sub-networks with the same structure set in parallel. The two CNN sub-networks are independent of each other and correspond to Figure 3 The trend image and the fluctuation image are input into the two CNN sub-networks to obtain the trend feature vector z1 and the fluctuation feature vector z2; in the specific implementation, the trend image and the fluctuation image are stacked into a two-channel three-dimensional tensor T∈R 2×N×N , which can be input into the network. Each CNN sub-network includes a convolutional layer, a pooling layer, and an average pooling layer that are cascaded in sequence. The structure of each CNN sub-network is as follows: Figure 5As shown in the figure, the convolutional layer extracts features from the input image and mines the correlations between them to generate a feature map. The pooling layer downsamples the feature map to reduce its dimensionality and remove redundant information. The average pooling layer compresses the reduced feature map into a feature vector of length N. Each CNN subnetwork independently processes a channel in the three-dimensional tensor, avoiding the mutual influence of trend and fluctuation features and extracting more accurate features.
[0071] Similarly, the prediction module also needs to design a dual-channel structure to separately process the extracted trend feature vector and fluctuation feature vector. In one embodiment, the prediction module includes two hidden layers with the same structure set in parallel and an output layer. The trend feature vector and the fluctuation feature vector are respectively input into the two hidden layers to obtain the trend hidden state and the fluctuation hidden state. The trend hidden state and the fluctuation hidden state are passed to the output layer to obtain the motion response data of the next M moments. Each hidden layer includes multiple cascaded GRU units. The output layer uses a fully connected layer. The hidden layer is used to capture the long-term temporal dependency in the feature vector to obtain the hidden state. The output layer adds the two hidden states and maps them into the motion response data of the next M moments.
[0072] GRU is a type of recurrent neural network that has good processing capabilities for time series prediction problems. The structure of the GRU unit is as follows: Figure 6 As shown in Figure 2, the effective information in the sequence data is captured by updating the gate and resetting the gate. For any feature vector z, the relevant control equation is as follows:
[0073] u l =σ(W u ·h l-1 +V u ·z l +b u )
[0074] r l =σ(W r ·h l-1 +V r ·z l +b r )
[0075]
[0076] Among them, z l is the lth value of the eigenvector z, u l is the output of the update gate and r l is the output of the reset gate, W and V are weight matrices, and b is the bias vector. is a candidate hidden state, h lis the captured hidden state, σ(·) is the sigmoid activation function, represents element-wise multiplication, and tanh(·) is the hyperbolic tangent activation function.
[0077] Each hidden layer receives one of the feature vectors output by the feature extraction module and passes it backward layer by layer along the GRU unit. The GRU unit continuously extracts the long-term time dependency in the sequence data to obtain the trend hidden state h1 and the fluctuation hidden state h2. The output layer adds the trend hidden state h1 and the fluctuation hidden state h2 and then maps them to obtain the motion response data for the next M moments.
[0078] Based on the network architecture of the floating wind turbine motion response prediction model constructed above, the floating wind turbine motion response prediction model corresponding to any level of sea condition set is obtained by using motion response training samples, including:
[0079] Initializing model parameters, the motion response training samples from the sea condition set are divided into multiple mini-datasets (mini-baths), each containing the same number of motion response training samples. Model training is then performed sequentially using each mini-dataset, with the motion response training samples from each mini-dataset sequentially entering the temporal dynamic decomposition module, image encoding module, feature extraction module, and prediction module. This method of dividing training samples avoids requiring all motion response training samples from the training set to be trained at once. This allows the model to update its parameters multiple times during a single training cycle across all mini-datasets, thereby improving training efficiency.
[0080] After the model completes a round of training, the prediction error of each model training is calculated using the motion response training samples in the validation set and the model parameters are adjusted through the optimizer until the prediction error of the model training no longer decreases in the model training for a predetermined number of consecutive rounds, or when the maximum number of rounds of model training is reached, the candidate prediction model corresponding to the sea condition set is trained. This application uses the mean square error (MSE) of the predicted motion response data and the predicted time series of the motion response training samples as the prediction error. The predetermined number of rounds is set to 10 rounds, and the maximum number of rounds is 500. During the training process, when the prediction error on the validation set no longer decreases for 5 consecutive rounds, the learning rate will be multiplied by a decay factor of 0.9, and the Adam optimizer will be used to update the learnable parameters. Among them, the motion response training samples in the training set, validation set and test set are composed of sub-time series obtained by dividing the training set data segment, validation set data segment and test set data segment in step 1. The calculation method of the mean square error is a conventional calculation method, which will not be repeated in this application.
[0081] By changing the model hyperparameters and re-training the model, multiple candidate prediction models corresponding to the sea condition set are obtained, and the candidate prediction model with the smallest prediction error is used as the floating wind turbine motion response prediction model corresponding to the sea condition set. The hyperparameter combination of the candidate prediction models is used as the final parameter setting of the floating wind turbine motion response prediction model.
[0082] Step 3: Use the environmental monitoring system to measure the current sea condition of the floating wind turbine in real time, and determine the floating wind turbine motion response prediction model corresponding to the sea condition set of the current sea condition.
[0083] Using the model training method in step 2, we train the model on the motion response training samples for each set of sea conditions to obtain the corresponding floating wind turbine motion response prediction models for each set of sea conditions, forming a prediction model library. In actual application, the matching floating wind turbine motion response prediction model is selected from the model library based on the sea conditions obtained through real-time measurements.
[0084] It should be noted that, since the prediction model for simultaneously predicting the motion response data of multiple degrees of freedom has a huge number of parameters and high computational complexity, it will lead to low computational efficiency and cannot adapt to the real-time prediction of the motion response of floating wind turbines. The floating wind turbine motion response prediction model trained by the method of the present application only predicts the motion response data of a single degree of freedom. Compared with the prediction model that simultaneously predicts the motion response data of multiple degrees of freedom, it can further improve the real-time performance of the prediction. It is only necessary to use the motion response data of one degree of freedom when constructing the motion response training sample. If it is necessary to simultaneously predict the motion response data of multiple degrees of freedom, it is only necessary to select multiple corresponding prediction models from the model library to achieve simultaneous prediction of multiple degrees of freedom. However, in actual applications, it is not necessary to simultaneously predict the motion response of the six degrees of freedom of a floating wind turbine. Generally, only one or two degrees of freedom with the greatest impact are concerned, such as longitudinal sway, pitch, and other degrees of freedom that have a greater impact on the stability of the floating wind turbine.
[0085] Step 4: Obtain the motion response data of the floating wind turbine at the current moment and the N previous historical moments to form a historical time series to be predicted. Use the floating wind turbine motion response prediction model corresponding to the sea condition set where the current sea condition is located to predict the motion response data of the floating wind turbine at the next M moments based on the historical time series to be predicted.
[0086] In one embodiment, the trained floating wind turbine motion response prediction model is used to predict the motion response of the floating wind turbine in the pitch direction in the test set, and the comparison between the prediction result and the true value is as follows: Figure 7 As shown in the figure, the prediction results of the present application method are basically consistent with the true value, and have high prediction accuracy.
[0087] The above merely describes the preferred embodiments of the present application, and the present application is not limited to the above embodiments. It can be understood that other improvements and changes directly derived or thought by those skilled in the art without departing from the spirit and concept of the present application shall be considered to be included in the protection scope of the present application.
Claims
1. A real-time prediction method for floating wind turbine motion response based on time series dynamic decomposition, characterized in that: The real-time prediction method for the motion response of a floating wind turbine comprises: Determine a set of sea conditions at each level in the target operating sea area, and obtain training samples of the motion response of the floating wind turbine under each set of sea conditions at each level; each set of sea conditions includes multiple sea conditions, and the training samples of the motion response of the floating wind turbine under each sea condition include a historical time series consisting of motion response data at N consecutive moments and a predicted time series consisting of motion response data at M consecutive moments after the N consecutive moments; N and M are integer parameters and N>M; A network architecture of a floating wind turbine motion response prediction model is constructed by sequentially cascading a time series dynamic decomposition module, an image encoding module, a feature extraction module, and a prediction module. The historical time series of motion response training samples of each level of sea condition set is used as input, and the predicted time series of the motion response training samples is used as output to train a floating wind turbine motion response prediction model corresponding to each level of sea condition set; the time series dynamic decomposition module decomposes the historical time series of the motion response training samples into a trend component and a fluctuation component, the image encoding module encodes the trend component and the fluctuation component respectively to obtain a trend image and a fluctuation image, the feature extraction module extracts features of the trend image and the fluctuation image respectively to obtain a trend feature vector and a fluctuation feature vector, the prediction module extracts long-term time dependencies in the trend feature vector and the fluctuation feature vector respectively and predicts motion response data for M moments in the future; the trend component represents the long-term evolution of the historical time series, and the fluctuation component represents the short-term fluctuation of the historical time series; Using the environmental monitoring system to measure the current sea conditions of the floating wind turbine in real time, and determining the floating wind turbine motion response prediction model corresponding to the sea condition set of the current sea conditions; The motion response data of the floating wind turbine at the current moment and N previous historical moments are obtained to form a historical time series to be predicted. The motion response data of the floating wind turbine at the next M moments are predicted based on the historical time series to be predicted, using the floating wind turbine motion response prediction model corresponding to the sea condition set of the current sea condition.
2. The method for real-time prediction of motion response of a floating wind turbine according to claim 1, characterized in that: Decomposing the historical time series of the motion response training samples into a trend component and a fluctuation component includes: Historical time series of motion response training samples Perform data expansion and use the time series window to perform sliding average calculation on the historical time series after data expansion to obtain the trend component The historical time series and trend component The fluctuation component is obtained by subtraction The trend component and the fluctuation component The length of the historical sequence are of equal length.
3. The real-time prediction method for the motion response of a floating wind turbine according to claim 1, characterized in that: The feature extraction module includes two CNN sub-networks with the same structure arranged in parallel. The trend image and the fluctuation image are respectively input into the two CNN sub-networks to obtain a trend feature vector and a fluctuation feature vector; each CNN sub-network includes a convolution layer, a pooling layer and an average pooling layer cascaded in sequence. The convolution layer is used to extract the features of the input image and mine the correlation between the features of the input image to obtain a feature map. The pooling layer is used to downsample the feature map to reduce the dimension of the feature map. The average pooling layer is used to compress the feature map after dimensionality reduction into a feature vector with a length of N.
4. The method for real-time prediction of motion response of a floating wind turbine according to claim 1, characterized in that: The prediction module includes two hidden layers with the same structure arranged in parallel and an output layer. The trend feature vector and the fluctuation feature vector are respectively input into the two hidden layers to obtain the trend hidden state and the fluctuation hidden state. The trend hidden state and the fluctuation hidden state are transmitted to the output layer to obtain the motion response data of the next M moments; each hidden layer includes multiple cascaded GRU units, and the output layer adopts a fully connected layer. The hidden layer is used to capture the long-term time dependency in the feature vector to obtain the hidden state. The output layer adds the two hidden states and maps them into the motion response data of the next M moments.
5. The method for real-time prediction of motion response of a floating wind turbine according to claim 1, characterized in that: The image encoding module uses the Gram angular field method to perform polar coordinate conversion on the trend component and the fluctuation component respectively and calculates the Gram matrix to obtain a trend image and a fluctuation image with a resolution of N×N.
6. The method for real-time prediction of motion response of a floating wind turbine according to claim 2, characterized in that: Decomposing the historical time series of the motion response training samples into trend components includes: In the historical time series The first data Previously, the predetermined number d1 was increased In the historical time series The last data Then increase the predetermined number d2 Align the starting position of the time series window with the starting position of the historical time series after data expansion, calculate the average value of each data covered by the time series window to obtain the first data of the trend component, and slide the time series window backward by one data until the end position of the time series window is aligned with the end position of the historical time series after data expansion to obtain the trend component.
7. The method for real-time prediction of motion response of a floating wind turbine according to claim 6, characterized in that: The length of the time series window is w. The first data Previously increased The number of d1=w-1-(w-1) / / 2; in the historical time series The last data Then increase The number of d2 = (w-1) / / 2; / / represents rounding after division.
8. The method for real-time prediction of motion response of a floating wind turbine according to claim 1, characterized in that: Obtaining motion response training samples includes: Establishing a time-domain integrated coupled numerical simulation model of a floating wind turbine, inputting sea conditions of various levels of sea condition sets into the time-domain integrated coupled numerical simulation model for numerical simulation calculation, and obtaining motion response data of the floating wind turbine under the different levels of sea condition sets; Normalizing the motion response data under each level of sea condition set, and randomly cutting the normalized motion response data into predetermined lengths to obtain motion response data segments; The motion response data segments are divided into training set data segments, validation set data segments and test set data segments, and the training set data segments, validation set data segments and test set data segments are respectively divided into multiple historical time series and predicted time series using the sliding window method to obtain motion response training samples.
9. The method for real-time prediction of motion response of a floating wind turbine according to claim 1, characterized in that: The sea condition sets for determining each level include: Various sea conditions for floating wind turbines are determined based on hydrological observation data of the target operating sea area, wherein the sea conditions include significant wave heights; and the various sea conditions are divided into sea condition sets of different levels based on the significant wave heights.
10. The method for real-time prediction of motion response of a floating wind turbine according to claim 1, characterized in that: The floating wind turbine motion response prediction models corresponding to any level of sea condition set obtained by training include: Initialize the model parameters, divide the motion response training samples of the sea condition set into multiple groups of sub-data sets, and the number of motion response training samples in each sub-data set is the same; use each group of sub-data sets to perform model training in turn, calculate the prediction error of each model training and adjust the model parameters through the optimizer until the prediction error of the model training no longer decreases in a predetermined number of consecutive rounds of model training, or when the maximum number of rounds of model training is reached, train to obtain a candidate prediction model corresponding to the sea condition set, change the model hyperparameters and re-train the model to obtain multiple candidate prediction models corresponding to the sea condition set, and use the candidate prediction model with the smallest prediction error as the floating wind turbine motion response prediction model corresponding to the sea condition set; use the mean square error of the predicted motion response data and the predicted time series of the motion response training samples as the prediction error.