A method for offshore wind power prediction based on modal decomposition and deep learning
By combining modal decomposition and deep learning, high-frequency target modal components are screened and model parameters are optimized to construct a multi-scale feature prediction model. This solves the problems of insufficient prediction accuracy and stability in traditional methods and achieves efficient prediction of offshore wind power.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID SHANGHAI MUNICIPAL ELECTRIC POWER CO
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-14
AI Technical Summary
Traditional offshore wind power forecasting methods struggle to simultaneously improve data quality, characterize multi-scale features, and model time-series dependencies, resulting in insufficient forecast accuracy and stability, especially when faced with complex marine meteorological conditions and the impact of equipment operating status.
A method combining mode decomposition and deep learning is adopted. High-frequency target mode components are screened through complete ensemble empirical mode decomposition (CEEMDAN) and variational mode decomposition (VMD). The parameters of the deep learning model are optimized by particle swarm optimization algorithm. A branch prediction model composed of temporal convolutional network (TCN) and gated recurrent unit (GRU) is constructed to perform multi-scale feature extraction and prediction.
It improves the accuracy and stability of offshore wind power forecasting, reduces the impact of high-frequency disturbances on the forecast results, and enhances the reliability and robustness of the forecast results, making it suitable for offshore wind farm operation management and grid dispatch.
Smart Images
Figure CN122393904A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a power prediction method for offshore wind farms, specifically a power prediction method for offshore wind power based on mode decomposition and deep learning. Background Technology
[0002] Against the backdrop of "dual carbon" goals and energy structure transformation, the installed capacity of offshore wind power continues to grow, and wind power forecasting has become a crucial technical support for ensuring the safe operation of the power grid and improving the absorption capacity of new energy sources. Offshore wind power data is typically characterized by nonlinearity, non-stationarity, and multi-scale fluctuations due to the influence of complex marine meteorological conditions, random wind speed fluctuations, and equipment operating status. Furthermore, measured data often contains outliers, which adversely affect the training and inference processes of prediction models, thus limiting prediction accuracy and model stability. Traditional statistical forecasting methods or single deep learning models often struggle to simultaneously improve data quality, characterize multi-scale features, and model temporal dependencies when faced with these complex characteristics, leaving room for improvement in prediction performance. Therefore, it is necessary to adopt an offshore wind power forecasting technology that can balance data quality improvement with prediction accuracy enhancement. This method can provide reliable data support for wind farm operation management, power grid dispatching decisions, and the optimal allocation of new energy systems, thereby promoting the efficient utilization and sustainable development of offshore wind power. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for predicting offshore wind power based on mode decomposition and deep learning.
[0004] The above objectives are achieved through the following technical solutions:
[0005] A method for predicting offshore wind power based on mode decomposition and deep learning, comprising the following steps:
[0006] S1: Obtain historical operating data of offshore wind farms, including wind power data and corresponding time series information, and perform format unification and missing value processing on the historical operating data to form a raw wind power dataset to be cleaned;
[0007] S2: Preprocess the original dataset obtained in step S1 to obtain the cleaned wind power time series;
[0008] S3: The cleaned wind power time series obtained in step S2 is subjected to initial mode decomposition using the complete set empirical mode decomposition method to obtain multiple mode components that characterize different frequency scales.
[0009] S4: Calculate the sample entropy and center frequency of each modal component obtained in step S3, and set threshold rules based on the joint judgment result of sample entropy and center frequency to screen out high-frequency target modal components that meet the threshold conditions.
[0010] S5: Perform a secondary decomposition on the high-frequency target mode components obtained in step S4, and use the variational mode decomposition method to further decompose the high-frequency target mode components into multiple high-resolution sub-mode components.
[0011] S6: Using prediction error as the fitness evaluation index, the particle swarm algorithm is used to adaptively optimize the key parameters of variational mode decomposition in step S5 in order to obtain the target high-frequency submode components.
[0012] S7: Combine the low-frequency modal components that did not participate in the secondary decomposition in step S3 with the high-frequency submodal components obtained in step S6 to construct a multi-scale feature subsequence dataset.
[0013] S8: Establish a branch prediction model for the multi-scale feature subsequence dataset constructed in step S7. Input the high-frequency submodal components into the temporal convolutional network (TCN) branch and the low-frequency modal components into the gated recurrent unit (GRU) branch. Optimize the initial and training parameters of the branch prediction model using the particle swarm optimization algorithm to obtain the target model parameters.
[0014] S9: Input the multi-scale feature subsequences obtained in step S7 into the corresponding branches of the branch prediction model constructed using the target model parameters obtained in step S8, and obtain the prediction results corresponding to each modal component.
[0015] S10: According to the decomposition order of each modal component and sub-modal component, the prediction results of each component obtained in step S9 are summed and reconstructed item by item to output the final prediction result of offshore wind power.
[0016] Furthermore, the specific method for data preprocessing in step S1 is as follows:
[0017] Historical operational data within a preset time period is acquired through an offshore wind farm monitoring system. This historical operational data includes wind power data and corresponding time series information. The acquired historical operational data is preprocessed to unify data from different sources or with different sampling frequencies into the same time interval and data format, and then arranged chronologically. Simultaneously, missing data in the time series is detected, and when missing values exist, interpolation methods are used to fill in the missing values. In the formula, and These represent the data before and after the missing values, respectively. This indicates the time point corresponding to the missing data. Indicates the missing position The interpolation result at the location.
[0018] Furthermore, the specific method described in step S2 is as follows:
[0019] The original wind power time series obtained in step S1 is subjected to anomaly sample identification and correction. Specifically, the DBSCAN clustering method is used to perform cluster analysis on the wind power data, and the sample points that deviate from the normal power distribution in the clustering results are identified as anomaly sample points. Furthermore, the normal power data of the adjacent time before and after the anomaly sample point are selected as reference samples, and a ninth-order polynomial fitting model is constructed to smooth the power value at the corresponding time of the anomaly sample point.
[0020] The ninth-order polynomial fitting model can be expressed as: In the formula, For wind speed, To fit the output power value, Let the degree be a polynomial. The fitting coefficients are to be determined.
[0021] The corrected power values are used to replace the original power values of the outlier sample points to obtain the cleaned wind power time series.
[0022] Furthermore, the specific method described in step S3 includes:
[0023] The cleaned wind power time series obtained in step S2 is represented as follows:
[0024] The cleaned wind power time series was subjected to initial mode decomposition using the complete ensemble empirical mode decomposition method CEEMDAN to reduce mode aliasing and improve the decomposition stability of the non-stationary wind power series. After initial mode decomposition, multiple modal components and residual components characterizing different frequency scales were obtained, as follows: In the formula, Indicates the first One eigenmode function Indicates the total number of modal components. Indicates the residual component;
[0025] The multiple modal components respectively characterize the fluctuation characteristics of wind power time series at different frequency scales, wherein the high-frequency modal components reflect short-term fluctuation information, the low-frequency modal components reflect long-term trend information, and the residual components reflect the overall trend information of the series.
[0026] Furthermore, the specific method of step S4 includes:
[0027] The sample entropy and center frequency of each modal component obtained in step S3 are calculated respectively to characterize the complexity and frequency distribution characteristics of each modal component; wherein, the sample entropy is used to reflect the fluctuation complexity of the modal component sequence, and the center frequency is used to reflect the dominant frequency position of the modal component.
[0028] A joint judgment index is constructed based on the sample entropy and center frequency corresponding to each modal component. The joint judgment index can be expressed as: In the formula, Indicates the first The joint criterion for determining the modal components, Indicates the first The sample entropy of each modal component. Indicates the first The center frequency of each modal component and These are the weighting coefficients;
[0029] According to the preset threshold Discrimination is performed on each modal component, when When, the corresponding modal component is determined to be a high-frequency target modal component; when When this happens, the corresponding modal component is determined to be a non-target modal component.
[0030] Furthermore, the specific method of step S5 includes:
[0031] The high-frequency target mode components obtained in step S4 are subjected to secondary decomposition. Variational mode decomposition is used to decompose each high-frequency target mode component into multiple high-resolution sub-mode components with different center frequencies and frequency band characteristics, in order to further separate local fluctuation features in the high-frequency disturbance information. Let any high-frequency target mode component be... After variational mode decomposition, it can be expressed as: In the formula, Indicates the first One high-frequency target mode component Indicates the index of the high-frequency target modal component. Represents a time variable. Indicates the first The total number of sub-mode components obtained after secondary decomposition of the high-frequency target mode components. Indicates the first The first high-frequency target modal component corresponds to the th The index of each submodal component, Indicates the first The high-frequency target mode component is obtained after secondary decomposition. A high-resolution sub-mode component.
[0032] Furthermore, the specific method of step S6 includes:
[0033] The key parameters of variational mode decomposition in step S5 are used as the parameters to be optimized in the particle swarm optimization algorithm, wherein the key parameters include the number of mode decompositions. and penalty factor Variational mode decomposition is performed on the high-frequency target mode components corresponding to each set of particle parameters, and the obtained high-frequency sub-mode components are input into the prediction model for power prediction. The prediction error between the prediction result and the actual wind power value is used as the fitness evaluation index of the particle swarm algorithm, and the position and velocity of the particles are updated according to the fitness value. When the preset termination condition is met, the parameter combination corresponding to the optimal fitness is output as the target decomposition parameter, and the high-frequency target mode components are re-decomposed using the target decomposition parameter to obtain the target high-frequency sub-mode components.
[0034] Let the total number of samples be , No. The actual wind power value of each sample point is The predicted wind power value is Then the fitness function of the particle swarm optimization algorithm can be expressed as: In the formula, This represents the fitness function value of the particle swarm optimization algorithm. This represents the total number of sample points. Indicates the sample point number;
[0035] The particle's position and velocity updates can be represented as: In the formula, and They represent the first During the nth iteration The position and velocity of each particle, and They represent the first During the nth iteration The position and velocity of each particle, This represents the best position in the particle's history. Indicates the globally optimal position. Indicates inertia weight, Represents the learning factor. Represents a random number;
[0036] For different parameter combinations Perform the quadratic decomposition and prediction calculation separately to obtain the corresponding prediction error. And the parameter selection criterion is based on minimizing the prediction error, expressed as: In the formula, Indicates the number of target mode decompositions. Indicates the target penalty factor. Indicates parameter combination The corresponding prediction error, This represents the combination of parameters that minimizes the objective function.
[0037] In obtaining the target parameter combination Then, using the target parameter combination, the high-frequency target modal components obtained in step S4 are re-decomposed using variational mode decomposition to obtain the target high-frequency sub-modal components; then, in the target parameter combination... The following, after being re-decomposed, can be represented as: Indicates the first The total number of high-frequency sub-mode components obtained after re-decomposing the high-frequency target mode components. Indicates the first The first high-frequency target modal component corresponds to the th The sequence number of each high-frequency submode component. Indicates the first The high-frequency target mode component is obtained after re-decomposition. One high-frequency submode component.
[0038] Furthermore, the specific method of step S7 includes:
[0039] The low-frequency modal components obtained from the initial modal decomposition in step S3 that did not participate in the secondary decomposition are retained as low-frequency feature subsequences, and the target high-frequency submodal components obtained in step S6 are used as high-frequency feature subsequences. The multi-scale feature subsequence dataset is constructed by combining each modal component and submodal component according to the frequency level and decomposition order, and used as the input for the subsequent branch prediction model.
[0040] Furthermore, the specific method of step S8 includes:
[0041] For the multi-scale feature subsequence dataset constructed in step S7, a branch prediction model is established, consisting of a temporal convolutional network (TCN) branch and a gated recurrent unit (GRU) branch. Specifically, the target high-frequency sub-mode components are input into the TCN branch to extract local short-term fluctuation features in the wind power sequence; the low-frequency mode components are input into the GRU branch to extract long-term trend dependence features in the wind power sequence; and the feature results extracted by the two branches are then fused to obtain the branch model output for subsequent power prediction.
[0042] The input multi-scale feature subsequence is represented as: In the formula, Indicates time Multi-scale feature vectors;
[0043] For the TCN branch, causal convolution with dilation factor is used to extract local temporal features from the input sequence. Its one-dimensional dilated convolution calculation can be expressed as: In the formula, Indicates the kernel size. Indicates the coefficient of thermal expansion. Represents the convolution kernel weight parameters. Indicates the first convolution kernel The serial number of each position;
[0044] For the GRU branch, the low-frequency modal components are input into the gated recurrent unit, and the long-term dependency of the time series is modeled using update and reset gates. The calculation process is expressed as follows: In the formula, represents time. The input vector, This represents the hidden state vector from the previous time step. This indicates updating the gate vector. This represents the Sigmoid activation function. This represents the weight matrix input to the update gate. This represents the weight matrix from the hidden state to the update gate loop; In the formula, This indicates resetting the gate vector. This represents the weight matrix input to the reset gate. This represents the cyclic weight matrix from the hidden state to the reset gate; In the formula, Represents the candidate hidden state vector. This represents the hyperbolic tangent activation function. This represents element-wise multiplication; This represents the weight matrix input to the candidate hidden state. The cyclic weight matrix represents the transition from the hidden state to the candidate hidden state; In the formula, This represents the hidden state vector at the current moment;
[0045] The characteristic results from the TCN branch and the GRU branch are fused to form the branch model output for offshore wind power prediction.
[0046] Furthermore, the specific method of step S9 includes:
[0047] The initial and training parameters of the branch prediction model established in step S8 are optimized using the particle swarm optimization algorithm. The learning rate, convolutional kernel size, number of hidden layer nodes, and number of training iterations are selected as parameters to be optimized, and the prediction error is used as the fitness evaluation index. The position and velocity of the particles are updated iteratively, and the parameter combination corresponding to the optimal fitness is selected as the target model parameters. Then, the multi-scale feature subsequences obtained in step S7 are input into the corresponding branches of the branch prediction model optimized by the particle swarm optimization algorithm. The target high-frequency sub-mode components are input into the TCN branch, and the low-frequency mode components are input into the GRU branch. The power prediction of the corresponding mode components is completed, and the prediction results of each mode component are output.
[0048] Furthermore, the specific method of step S10 includes:
[0049] Extract the prediction results of each low-frequency mode component and each target high-frequency sub-mode component obtained in step S9, and align them according to the decomposition order corresponding to the initial mode decomposition and the secondary decomposition; sum and reconstruct each prediction result of the aligned components to obtain the final prediction result of offshore wind power.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] (1) By performing initial mode decomposition on the wind power sequence and combining sample entropy and center frequency to screen high-frequency target mode components, the ability to characterize complex non-stationary features can be improved.
[0052] (2) By performing secondary variational mode decomposition on the high-frequency target mode components and using particle swarm optimization algorithm to optimize the decomposition parameters, the accuracy of high-frequency disturbance information extraction can be improved.
[0053] (3) By constructing a branch prediction model consisting of TCN branch and GRU branch, and reconstructing the prediction results of each component, the accuracy and stability of offshore wind power prediction can be improved.
[0054] This invention provides a method for predicting offshore wind power based on mode decomposition and deep learning. Unlike traditional single prediction models or fixed-parameter prediction methods, this method addresses the characteristics of offshore wind power data, such as strong nonlinearity, significant non-stationarity, and prominent multi-scale fluctuations. It introduces mode decomposition to separate multi-scale features of the wind power signal and combines it with particle swarm optimization to adaptively optimize the structural parameters of the deep prediction model, thereby effectively improving prediction accuracy and model stability. Simultaneously, by specifically modeling features at different scales, the impact of high-frequency disturbances on the prediction results is reduced, improving the reliability and robustness of the prediction results. This is of great significance for the operation and management of offshore wind farms, grid dispatch, and the safe and stable operation of new energy systems. Attached Figure Description
[0055] Figure 1 This is a flowchart of a method for predicting offshore wind power based on mode decomposition and deep learning.
[0056] Figure 2 This is a two-stage cleaning process flowchart;
[0057] Figure 3 Here is a flowchart of the particle swarm optimization algorithm;
[0058] Figure 4 The following are the decomposition results of the wind speed sequence: (a) is the result of the first decomposition (only the first five components are shown), and (b) is the result of the second decomposition.
[0059] Figure 5 Here is a flowchart of the secondary decomposition and prediction process;
[0060] Figure 6 This is a graph showing the predicted power output of offshore wind power. Detailed Implementation
[0061] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.
[0062] It is worth mentioning that, unless otherwise stated, the technical terms used in this application should have the ordinary meaning as understood by those skilled in the art.
[0063] Please see Figure 1The present invention discloses a method for predicting offshore wind power based on mode decomposition and deep learning, the method comprising the following steps:
[0064] S1: Obtain historical operating data of offshore wind farms, including wind power data and corresponding time series information, and perform format unification and missing value processing on the historical operating data to form a raw wind power dataset to be cleaned;
[0065] S2: Preprocess the original dataset obtained in step S1 to obtain the cleaned wind power time series;
[0066] S3: The cleaned wind power time series obtained in step S2 is subjected to initial mode decomposition using the complete set empirical mode decomposition method to obtain multiple mode components that characterize different frequency scales.
[0067] S4: Calculate the sample entropy and center frequency of each modal component obtained in step S3, and set threshold rules based on the joint judgment result of sample entropy and center frequency to screen out high-frequency target modal components that meet the threshold conditions.
[0068] S5: Perform a secondary decomposition on the high-frequency target mode components obtained in step S4, and use the variational mode decomposition method to further decompose the high-frequency target mode components into multiple high-resolution sub-mode components.
[0069] S6: Using prediction error as the fitness evaluation index, the particle swarm algorithm is used to adaptively optimize the key parameters of variational mode decomposition in step S5 in order to obtain the target high-frequency submode components.
[0070] S7: Combine the low-frequency modal components that did not participate in the secondary decomposition in step S3 with the high-frequency submodal components obtained in step S6 to construct a multi-scale feature subsequence dataset.
[0071] S8: Establish a branch prediction model for the multi-scale feature subsequence dataset constructed in step S7. Input the high-frequency submodal components into the temporal convolutional network (TCN) branch and the low-frequency modal components into the gated recurrent unit (GRU) branch. Optimize the initial and training parameters of the branch prediction model using the particle swarm optimization algorithm to obtain the target model parameters.
[0072] S9: Input the multi-scale feature subsequences obtained in step S7 into the corresponding branches of the branch prediction model constructed using the target model parameters obtained in step S8, and obtain the prediction results corresponding to each modal component.
[0073] S10: According to the decomposition order of each modal component and sub-modal component, the prediction results of each component obtained in step S9 are summed and reconstructed item by item to output the final prediction result of offshore wind power.
[0074] Furthermore, the specific method for data preprocessing in step S1 is as follows:
[0075] Historical operational data within a preset time period is acquired through an offshore wind farm monitoring system. The historical operational data includes wind power data and corresponding time series information. The acquired historical operational data is preprocessed to unify data from different sources or with different sampling frequencies into the same time interval and data format, and then arranged in chronological order.
[0076] We selected the measured operation dataset of a 48MW offshore wind farm (site f1) in a certain city from January 2022 to January 2023 for analysis, processing, and algorithm verification. The data was recorded with SCADA data at 15-minute intervals. After detecting and imputing missing data in the time series, a total of 37,163 data points were obtained, including variables such as wind speed and power.
[0077] Furthermore, the specific method described in step S2 is as follows:
[0078] Two-stage cleaning process diagram as follows Figure 2 As shown in the diagram, the raw data is first preprocessed to remove invalid values and normalize it, reducing the impact of dimensional differences on subsequent calculations. In the first stage, DBSCAN is used to perform density clustering on the wind power data to identify random, isolated outliers. After cleaning, inverse normalization is performed. In the second stage, a ninth-order polynomial regression is used to fit the relationship between power and wind speed, correcting for clustered structural outliers, and the fitting effect is evaluated using the coefficient of determination. After these two stages of cleaning, local anomalies in the data are suppressed, and the overall trend remains reasonable, providing reliable input for subsequent prediction models.
[0079] The Spearman correlation coefficient was used to evaluate the cleaning effect, and the results are shown in Table 1. The Spearman correlation coefficient of the wind speed-power data after cleaning using the DBSCAN and multinomial regression methods was improved to some extent compared to the DBSCAN method.
[0080] Table 1. Spearman correlation coefficients before and after data cleaning: Cleaning method Spearman correlation coefficient Raw data 0.7977 DBSCAN Method 0.85140.9062
[0081] Furthermore, the specific method described in step S3 includes:
[0082] The cleaned wind power time series obtained in step S2 is represented as follows:
[0083] The cleaned wind power time series was subjected to initial mode decomposition using the complete ensemble empirical mode decomposition method CEEMDAN to reduce mode aliasing and improve the decomposition stability of the non-stationary wind power series. After initial mode decomposition, multiple modal components and residual components characterizing different frequency scales were obtained, as follows: In the formula, Indicates the first One eigenmode function Indicates the total number of modal components. Indicates the residual component;
[0084] The multiple modal components respectively characterize the fluctuation characteristics of wind power time series at different frequency scales, wherein the high-frequency modal components reflect short-term fluctuation information, the low-frequency modal components reflect long-term trend information, and the residual components reflect the overall trend information of the series.
[0085] Furthermore, the specific method described in step S4 includes:
[0086] The sample entropy and center frequency of each modal component obtained in step S3 are calculated respectively to characterize the complexity and frequency distribution characteristics of each modal component; wherein, the sample entropy is used to reflect the fluctuation complexity of the modal component sequence, and the center frequency is used to reflect the dominant frequency position of the modal component.
[0087] A joint judgment index is constructed based on the sample entropy and center frequency corresponding to each modal component. The joint judgment index can be expressed as: In the formula, Indicates the first The joint criterion for determining the modal components, Indicates the first The sample entropy of each modal component. Indicates the first The center frequency of each modal component and These are the weighting coefficients;
[0088] According to the preset threshold Discrimination is performed on each modal component, when When, the corresponding modal component is determined to be a high-frequency target modal component; when When this happens, the corresponding modal component is determined to be a non-target modal component.
[0089] Furthermore, the specific methods described in steps S5, S6, and S7 include:
[0090] After obtaining the optimal VMD decomposition parameters through iterative search using the particle swarm optimization algorithm, the optimal parameters are then used... The target signal is decomposed twice to obtain several high-frequency sub-mode components. The remaining intrinsic mode functions obtained from the first decomposition are then combined with the sub-mode components obtained from the second decomposition in chronological order to construct a multi-scale feature dataset for subsequent prediction model input. The flowchart of the particle swarm optimization algorithm is as follows: Figure 3 As shown, after obtaining the optimal VMD decomposition parameters through iterative search using the particle swarm optimization algorithm, the target signal is decomposed a second time using the optimal parameters to obtain several high-frequency submode components.
[0091] In this embodiment, the offshore wind speed sequence is first decomposed using CEEMDAN to obtain 10 IMF components. IMF1, as the key high-frequency component, is then further decomposed using VMD to obtain 5 sub-modes; the remaining IMFs 2-10 remain unchanged. After feature similarity combination and reconstruction, 14 wind speed components are finally obtained. Correspondingly, the CEEMDAN and VMD decompositions of the offshore wind power sequence output 10 and 4 components respectively, which are combined and reconstructed to obtain 13 wind power components. The wind speed sequence decomposition results are as follows: Figure 4 As shown, (a) is the result of the first decomposition, and (b) is the result of the second decomposition.
[0092] Furthermore, the specific method of step S8 includes:
[0093] For the multi-scale feature subsequence dataset constructed in step S7, a branch prediction model is established, consisting of a temporal convolutional network (TCN) branch and a gated recurrent unit (GRU) branch. Specifically, the target high-frequency sub-mode components are input into the TCN branch to extract local short-term fluctuation features in the wind power sequence; the low-frequency mode components are input into the GRU branch to extract long-term trend dependence features in the wind power sequence; and the feature results extracted by the two branches are then fused to obtain the branch model output for subsequent power prediction.
[0094] The characteristic results from the TCN branch and the GRU branch are fused to form the branch model output for offshore wind power prediction.
[0095] Furthermore, the specific method of step S9 includes:
[0096] The particle swarm optimization algorithm is used to optimize the parameters of the branch prediction model constructed in step S8. The learning rate, convolutional kernel size, number of hidden layer units, and number of training iterations are used as variables to be optimized, and the prediction error is used as the fitness evaluation criterion. The optimal parameter combination is determined by continuously updating the position and velocity of the particles to obtain the target prediction model. Subsequently, the multi-scale feature subsequences formed in step S7 are input into the corresponding branches of the optimized branch prediction model. The target high-frequency sub-mode components are input into the TCN branch, and the low-frequency mode components are input into the GRU branch to complete the power prediction of each mode component and obtain the corresponding prediction results.
[0097] Furthermore, the specific method of step S10 includes:
[0098] Extract the prediction results of each low-frequency mode component and each target high-frequency sub-mode component obtained in step S9, and align them according to the decomposition order corresponding to the initial mode decomposition and the secondary decomposition; sum and reconstruct each prediction result of the aligned components to obtain the final prediction result of offshore wind power.
[0099] The flowchart of the secondary decomposition and prediction is as follows: Figure 5 As shown.
[0100] This invention uses mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean relative error percentage (MAPE) as evaluation indicators. The smaller the value of these evaluation indicators, the smaller the prediction error and the higher the accuracy. The calculation method is shown in Table 2.
[0101] Table 2 Main evaluation indicators:
[0102] In this embodiment, historical operating data of offshore wind farms are selected as experimental samples, and data cleaning, feature construction, and power prediction are performed sequentially according to the method described in this invention. Simulation results are as follows: Figure 6 As shown in Table 3. From Figure 6 As shown in Table 3, the method of the present invention exhibits good prediction accuracy and stability in typical offshore wind power prediction scenarios, and can effectively improve prediction performance, indicating that the method has good feasibility and engineering application value.
[0103] Table 3 Performance comparison with traditional algorithms: Prediction methods MSE / MW RMSE / MW MAE / MW MAPE / % Traditional Algorithm 3.34 3.41 2.14 5.93 This article's method 0.75 0.87 0.63 2.52
[0104] In summary, based on the attached figures and Tables 1 and 3, the algorithm proposed in this embodiment can significantly improve the quality of data cleaning and effectively enhance prediction performance.
[0105] This embodiment proposes a method for predicting offshore wind power based on mode decomposition and deep learning. Addressing the challenges of offshore wind power data being significantly influenced by the marine environment and exhibiting strong nonlinearity, non-stationarity, and significant multi-scale fluctuations in time series, the proposed method performs mode decomposition on the wind power series to effectively extract features at different time scales. Furthermore, it utilizes a particle swarm optimization algorithm to adaptively optimize the structural parameters of the deep prediction model, thereby enhancing the model's ability to represent complex time series characteristics. By modeling features at each scale separately and synthesizing the prediction results, the adverse effects of high-frequency random disturbances on prediction performance are reduced, enhancing the stability and reliability of the prediction results. This method has significant engineering application value and can provide strong support for the operation and management of offshore wind farms and grid dispatch.
Claims
1. A method for predicting offshore wind power based on mode decomposition and deep learning, characterized in that: The method includes the following steps: S1: Obtain the raw wind power dataset of the offshore wind farm; S2: Preprocess the raw wind power dataset obtained in step S1 to obtain a cleaned wind power time series; S3: Perform initial mode decomposition on the wind power time series using the complete ensemble empirical mode decomposition method to obtain multiple mode components representing features at different frequency scales; S4: Filter out high-frequency target mode components that meet the threshold conditions from each mode component obtained in step S3; S5: Perform secondary decomposition on the high-frequency target mode components obtained in step S4 into multiple high-resolution sub-mode components; S6: The key to variational mode decomposition in step S5... S7: Adaptively optimize the parameters to obtain the target high-frequency sub-mode components; S8: Construct a multi-scale feature sub-sequence dataset; S9: Build a branch prediction model for the multi-scale feature sub-sequence dataset constructed in step S7 to obtain the target model parameters; S10: Input each multi-scale feature sub-sequence obtained in step S7 into the corresponding branch of the branch prediction model constructed using the target model parameters obtained in step S8 to obtain the prediction results corresponding to each mode component; S11: According to the decomposition order corresponding to each mode component and sub-mode component, sum and reconstruct the prediction results of each component obtained in step S9 to output the final prediction result of offshore wind power.
2. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method described in step S1 is as follows: Historical operational data within a preset time period is acquired through an offshore wind farm monitoring system. This historical operational data includes wind power data and corresponding time series information. The acquired historical operational data is preprocessed to unify data from different sources or with different sampling frequencies into the same time interval and data format, and then arranged chronologically. Simultaneously, missing data in the time series is detected, and when missing values exist, interpolation methods are used to fill in the missing values. In the formula, and These represent the data before and after the missing values, respectively. This indicates the time point corresponding to the missing data. Indicates the missing position The interpolation result at the location.
3. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method described in step S2 is as follows: The original wind power time series obtained in step S1 is subjected to anomaly sample identification and correction. Specifically, the DBSCAN clustering method is used to perform cluster analysis on the wind power data, and the sample points that deviate from the normal power distribution in the clustering results are identified as anomaly sample points. Furthermore, the normal power data of the adjacent time before and after the anomaly sample point are selected as reference samples, and a ninth-order polynomial fitting model is constructed to smooth the power value at the corresponding time of the anomaly sample point. The ninth-order polynomial fitting model can be expressed as: In the formula, For wind speed, To fit the output power value, Let the degree be a polynomial. The fitting coefficients are to be determined. The corrected power values are used to replace the original power values of the outlier sample points to obtain the cleaned wind power time series.
4. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method described in step S3 includes: The cleaned wind power time series obtained in step S2 is represented as follows: ; The cleaned wind power time series was subjected to initial mode decomposition using the complete ensemble empirical mode decomposition method CEEMDAN to reduce mode aliasing and improve the decomposition stability of the non-stationary wind power series. After initial mode decomposition, multiple modal components and residual components characterizing different frequency scales were obtained, as follows: In the formula, Indicates the first One eigenmode function Indicates the total number of modal components. Indicates the residual component; The multiple modal components respectively characterize the fluctuation characteristics of wind power time series at different frequency scales, wherein the high-frequency modal components reflect short-term fluctuation information, the low-frequency modal components reflect long-term trend information, and the residual components reflect the overall trend information of the series.
5. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method of step S4 includes: The sample entropy and center frequency of each modal component obtained in step S3 are calculated respectively to characterize the complexity and frequency distribution characteristics of each modal component; wherein, the sample entropy is used to reflect the fluctuation complexity of the modal component sequence, and the center frequency is used to reflect the dominant frequency position of the modal component. A joint judgment index is constructed based on the sample entropy and center frequency corresponding to each modal component. The joint judgment index can be expressed as: In the formula, Indicates the first The joint criterion for determining the modal components, Indicates the first The sample entropy of each modal component. Indicates the first The center frequency of each modal component and These are the weighting coefficients; According to the preset threshold Discrimination is performed on each modal component, when When, the corresponding modal component is determined to be a high-frequency target modal component; when When this happens, the corresponding modal component is determined to be a non-target modal component.
6. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method of step S5 includes: The high-frequency target mode components obtained in step S4 are subjected to secondary decomposition. Variational mode decomposition is used to decompose each high-frequency target mode component into multiple high-resolution sub-mode components with different center frequencies and frequency band characteristics, in order to further separate local fluctuation features in the high-frequency disturbance information. Let any high-frequency target mode component be... After variational mode decomposition, it can be expressed as: In the formula, Indicates the first One high-frequency target mode component Indicates the index of the high-frequency target modal component. Represents a time variable. Indicates the first The total number of sub-mode components obtained after secondary decomposition of the high-frequency target mode components. Indicates the first The first high-frequency target modal component corresponds to the th The index of each submodal component, Indicates the first The high-frequency target mode component is obtained after secondary decomposition. A high-resolution sub-mode component.
7. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method of step S6 includes: The key parameters of variational mode decomposition in step S5 are used as the parameters to be optimized in the particle swarm optimization algorithm, wherein the key parameters include the number of mode decompositions. and penalty factor Variational mode decomposition is performed on the high-frequency target mode components corresponding to each set of particle parameters, and the obtained high-frequency sub-mode components are input into the prediction model for power prediction. The prediction error between the prediction result and the actual wind power value is used as the fitness evaluation index of the particle swarm algorithm, and the position and velocity of the particles are updated according to the fitness value. When the preset termination condition is met, the parameter combination corresponding to the optimal fitness is output as the target decomposition parameter, and the high-frequency target mode components are re-decomposed using the target decomposition parameter to obtain the target high-frequency sub-mode components. Let the total number of samples be , No. The actual wind power value of each sample point is The predicted wind power value is Then the fitness function of the particle swarm optimization algorithm can be expressed as: In the formula, This represents the fitness function value of the particle swarm optimization algorithm. This represents the total number of sample points. Indicates the sample point number; The particle's position and velocity updates can be represented as: In the formula, and They represent the first During the nth iteration The position and velocity of each particle, and They represent the first During the nth iteration The position and velocity of each particle, This represents the best position in the particle's history. Indicates the globally optimal position. Indicates inertia weight, Represents the learning factor. Represents a random number; For different parameter combinations Perform the quadratic decomposition and prediction calculation separately to obtain the corresponding prediction error. And the parameter selection criterion is based on minimizing the prediction error, expressed as: In the formula, Indicates the number of target mode decompositions. Indicates the target penalty factor. Indicates parameter combination The corresponding prediction error, This represents the combination of parameters that minimizes the objective function. In obtaining the target parameter combination Then, using the target parameter combination, the high-frequency target modal components obtained in step S4 are re-decomposed using variational mode decomposition to obtain the target high-frequency sub-modal components; then, in the target parameter combination... The following, after being re-decomposed, can be represented as: Indicates the first The total number of high-frequency sub-mode components obtained after re-decomposing the high-frequency target mode components. Indicates the first The first high-frequency target modal component corresponds to the th The sequence number of each high-frequency submode component. Indicates the first The high-frequency target mode component is obtained after re-decomposition. One high-frequency submode component.
8. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method of step S7 includes: The low-frequency modal components obtained from the initial modal decomposition in step S3 that did not participate in the secondary decomposition are retained as low-frequency feature subsequences, and the target high-frequency submodal components obtained in step S6 are used as high-frequency feature subsequences. The multi-scale feature subsequence dataset is constructed by combining each modal component and submodal component according to the frequency level and decomposition order, and used as the input for the subsequent branch prediction model.
9. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific method of step S8 includes: For the multi-scale feature subsequence dataset constructed in step S7, a branch prediction model is established, consisting of a temporal convolutional network (TCN) branch and a gated recurrent unit (GRU) branch. Specifically, the target high-frequency sub-mode components are input into the TCN branch to extract local short-term fluctuation features in the wind power sequence; the low-frequency mode components are input into the GRU branch to extract long-term trend dependence features in the wind power sequence; and the feature results extracted by the two branches are then fused to obtain the branch model output for subsequent power prediction. The input multi-scale feature subsequence is represented as: In the formula, Indicates time Multi-scale feature vectors; For the TCN branch, causal convolution with dilation factor is used to extract local temporal features from the input sequence. Its one-dimensional dilated convolution calculation can be expressed as: In the formula, Indicates the kernel size. Indicates the coefficient of thermal expansion. Represents the convolution kernel weight parameters. Indicates the first convolution kernel The serial number of each position; For the GRU branch, the low-frequency modal components are input into the gated recurrent unit, and the long-term dependency of the time series is modeled using update and reset gates. The calculation process is expressed as follows: In the formula, represents time. The input vector, This represents the hidden state vector from the previous time step. This indicates updating the gate vector. This represents the Sigmoid activation function. This represents the weight matrix input to the update gate. This represents the weight matrix from the hidden state to the update gate loop; In the formula, This indicates resetting the gate vector. This represents the weight matrix input to the reset gate. This represents the cyclic weight matrix from the hidden state to the reset gate; In the formula, Represents the candidate hidden state vector. This represents the hyperbolic tangent activation function. This represents element-wise multiplication; This represents the weight matrix input to the candidate hidden state. The cyclic weight matrix represents the transition from the hidden state to the candidate hidden state; In the formula, This represents the hidden state vector at the current moment; The characteristic results from the TCN branch and the GRU branch are fused to form the branch model output for offshore wind power prediction.
10. The offshore wind power prediction method based on mode decomposition and deep learning according to claim 1, characterized in that: The specific methods for steps S9-S10 include: The initial and training parameters of the branch prediction model established in step S8 are optimized using the particle swarm optimization algorithm. The learning rate, convolutional kernel size, number of hidden layer nodes, and number of training iterations are selected as the parameters to be optimized, and the prediction error is used as the fitness evaluation index. The position and velocity of the particles are updated iteratively, and the parameter combination corresponding to the optimal fitness is selected as the target model parameters. The multi-scale feature subsequences obtained in step S7 are input into the corresponding branches of the branch prediction model optimized by the particle swarm optimization algorithm. The target high-frequency sub-mode components are input into the TCN branch, and the low-frequency mode components are input into the GRU branch. The power prediction of the corresponding mode components is completed, and the prediction results of each mode component are output. Extract the prediction results of each low-frequency mode component and each target high-frequency sub-mode component obtained in step S9, and align them according to the decomposition order corresponding to the initial mode decomposition and the secondary decomposition; sum and reconstruct each prediction result of the aligned components to obtain the final prediction result of offshore wind power.