A wind turbine power prediction method and system for an offshore wind farm

Abstract

The application provides a wind turbine power prediction method and system for offshore wind farms. A numerical model of wave-structure interaction is used to generate a simulation dataset of the interaction between sea waves and structures for a target wind turbine type of an offshore wind farm. Then, a proxy model is trained using the simulation dataset. The proxy model is embedded in an atmospheric surface layer parameterization module of a sea-air-wave coupled model, so that at each time step, the proxy model predicts the attenuated significant wave height. Based on the wave height, the atmospheric surface layer parameterization module updates the atmospheric data and calculates the inflow wind speed. Finally, the wind turbine power is calculated using the inflow wind speed. In this way, the dissipation effect of the wind turbine foundation structure on the waves and its feedback mechanism on the atmospheric wind field are systematically introduced into the sea-air-wave coupled model, improving the accuracy of the predicted wind turbine power in the coupled model.

Description

Technical Field

[0001] This invention relates to the technical field of wind farm prediction, and particularly to a method and system for predicting the power of wind turbines in offshore wind farms. Background Technology

[0002] To achieve efficient site selection, accurate power generation prediction, and environmental impact assessment of offshore wind farms, mesoscale ocean-atmosphere-wave coupled models (such as COAWST) are commonly used for integrated simulation. Existing models, when parameterizing wind farms, mainly focus on the wake effect generated by the wind turbine blades extracting momentum from the atmospheric boundary layer, thereby capturing the unidirectional impact of wind farms on the atmosphere.

[0003] However, in practical applications, due to the lack of consideration for the backward coupling mechanism of wind farms, the models often underestimate or overestimate the downstream wind speed, resulting in deviations in the power curve of wind turbines and the prediction of overall power generation. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for predicting the power of wind turbines in offshore wind farms, which can improve the accuracy of predicting wind turbine power in the sea-air-wave coupling model.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for predicting the power output of offshore wind farm turbines, comprising: For the target wind turbine type in offshore wind farms, a simulated dataset of wind turbines and waves is generated using a wave structure interaction numerical model. The simulated dataset is then used to train a machine learning model. The prediction logic of the machine learning model is expressed in a complete analytical expression without any third-party runtime dependencies. The machine learning model is used as a proxy model for wave height attenuation effect. The proxy model is embedded into the near-surface layer parameterization module of the atmospheric component of the ocean-atmosphere-wave coupling model, and the near-surface layer parameterization module calls the proxy model to predict the weakened effective wave height under the current sea state. Based on the weakened effective wave height, the atmospheric data is updated in the near-surface layer parameterization module, and the inflow wind speed is calculated based on the atmospheric data. The wind turbine power of the wind farm is predicted based on the inflow wind speed.

[0006] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A wind turbine power prediction system for an offshore wind farm includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the various steps in the aforementioned wind turbine power prediction method for an offshore wind farm.

[0007] The beneficial effects of this invention are as follows: A numerical model of wave-structure interaction is used to generate a simulated dataset of wave-structure interaction for the target wind turbine type in offshore wind farms. This simulated dataset is then used to train a surrogate model. This surrogate model is embedded in the near-surface layer parameterization module of the ocean-air-wave coupled model, allowing the near-surface layer parameterization module to predict the weakened effective wave height at each time step. Based on this wave height, the near-surface layer parameterization module updates the atmospheric data and calculates the inflow wind speed. Finally, the wind turbine power is calculated using the inflow wind speed. In this way, the dissipation effect of wind turbine foundation structure on waves and its feedback mechanism on atmospheric wind field are systematically introduced into the ocean-air-wave coupled model, improving the accuracy of wind turbine power prediction in the coupled model. This scheme uses a machine learning model with prediction logic expressed in a complete analytical form and without third-party operational dependencies as a proxy model for wave height attenuation effect. Through "offline pre-computation + online hard coding", the model is seamlessly embedded in the near-surface layer parameterization module of the atmospheric model. While ensuring the accuracy of wave height simulation, it is fully adapted to the real-time operation requirements of the coupled mode, balancing prediction accuracy and computational efficiency, and has strong engineering applicability. Attached Figure Description

[0008] Figure 1 This is a flowchart of a wind turbine power prediction method for an offshore wind farm according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the implementation process of the wind farm parameterization scheme in the COAWST model according to an embodiment of the present invention; Figure 3 This is a scatter plot showing the effective wave heights of the training data from this embodiment of the invention compared to four typical machine learning regression models. Figure 4 This is a schematic diagram of a wind turbine power prediction system for an offshore wind farm according to an embodiment of the present invention. Detailed Implementation

[0009] Definitions:

[0010] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0011] In existing technologies, the site selection and power generation prediction of offshore wind farms generally employ mesoscale ocean-atmosphere-wave coupled models (such as COAWST) for comprehensive simulation. These models, when parameterizing wind farms, primarily focus on the wake effect generated by the wind turbine blades extracting momentum from the atmospheric boundary layer, thus capturing the unidirectional (upwind) impact of the wind farm on the atmosphere. However, in practical applications, existing parameterization methods cannot fully reflect the changes in local sea surface wind speed, sea surface temperature, wave height, and ocean currents caused by the wind farm, leading to biased risk assessments of the marine environment. Due to the lack of consideration for the backward coupling (backward feedback) mechanism of the wind farm, the models often underestimate or overestimate downstream wind speed, resulting in deviations in the power curves of the wind turbines and overall power generation predictions. Therefore, existing ocean-atmosphere-wave coupled models are mainly limited to unidirectional energy extraction in wind farm parameterization, failing to accurately assess the actual impact of wind farm construction on local marine meteorology, thus introducing errors in predicting downstream wind speed and power output.

[0012] To at least address the aforementioned issues, a Gaussian process surrogate model is generated through numerical simulation of wave structure interaction. This model is then embedded with near-surface layer parameterization of the sea-air-wave coupling system. The weakened wave height and atmospheric data are updated in real time, thereby accurately correcting the wind speed. In other words, a feedback mechanism of "wind turbine structure-wave-atmosphere-wind farm" is established to improve the power prediction accuracy of offshore wind farms.

[0013] The following details a method for predicting the power output of offshore wind farms according to the present invention. Please refer to [link / reference]. Figure 1 The method 100 includes steps 101 to 104: Step 101: For the target wind turbine type of offshore wind farm, generate a simulated dataset of wind turbine and waves using a wave structure interaction numerical model, train a machine learning model using the simulated dataset, and use the prediction logic of the machine learning model to express the prediction logic in a complete analytical expression without any third-party runtime dependency, and use the machine learning model as a proxy model for wave height attenuation effect.

[0014] The machine learning models that satisfy the requirement of having a prediction logic expressed through a complete analytical expression and without third-party runtime dependencies include, but are not limited to: linear regression models, multinomial regression models, support vector regression (SVM) models, and Gaussian process regression (GPR) models. All of the above models meet the requirements of a fully analytical prediction process, no third-party dependencies, and hard-codeable embedding. It should be noted that this embodiment excludes deep learning neural networks and ensemble learning models (such as random forests and gradient boosting trees). Although these models also fall under the category of machine learning, their prediction process cannot be expressed through a single mathematical analytical expression, or they rely on specific inference frameworks and runtime environments, and cannot be directly embedded into the near-ground parameterization module of the WRF pattern, thus not meeting the engineering implementation requirements of this embodiment.

[0015] Specifically, in the design and operation of offshore wind farms, the interaction between the wind turbine structure and ocean waves leads to wave height attenuation. To accurately quantify the wave height attenuation effect, a high-precision numerical model of wave structure interaction is first used for the target wind turbine type (e.g., stationary or floating). This model takes seabed depth, incident significant wave height, and spectral peak period as inputs, and outputs the significant wave height attenuated by the structure. By sampling different parameter combinations under typical sea states, a training dataset covering a wide range of sea states is formed.

[0016] In some embodiments, a machine learning model expressed by a complete mathematical analytical expression is trained using a simulated dataset; the machine learning model is used to fit a functional relationship between input features and weakened effective wave height.

[0017] Specifically, the core selection principle for machine learning models used to fit the wave dissipation effect of piles is that the model's prediction logic must be expressed through a complete mathematical analytical expression, without relying on a third-party inference framework, and can be directly hard-coded into the near-surface layer parameterization module of the WRF mode through a compiled programming language.

[0018] This is because the core control equations of the WRF model are solved and the physical process parameters are compiled using the Fortran language, which only supports analytical calculations that can be directly encoded and cannot be compatible with black-box machine learning models that depend on external operating environments. At the same time, it requires the model to have excellent fitting accuracy and generalization ability for small-sample, highly nonlinear physical processes such as the interaction between offshore wind turbine piles and waves.

[0019] Based on the above principles, the Gaussian process regression (GPR) model is the optimal implementation method in this embodiment, with the following core advantages: 1) The prediction process is completely analytical: The prediction logic of the GPR model can be fully expressed through the kernel function analytical expression, the Gaussian distribution joint probability density formula, and the prediction mean formula. There are no black box links. It can be directly hard-coded into the WRF mode through the Fortran language, perfectly adapting to the compilation and running requirements of the mode. 2) Excellent small-sample nonlinear fitting ability: For the limited simulation dataset of marine pile-wave interaction, the GPR model can accurately capture the nonlinear relationship between input parameters and wave height attenuation effect, with strong generalization ability, avoiding the risk of overfitting, and prediction accuracy is far superior to traditional models such as linear regression and multinomial regression. 3) High degree of matching with physical processes: In this embodiment, the Matern 5 / 2 kernel function is used as the covariance function of the GPR model. This kernel function has the best fitting effect on the continuously changing and non-stationary physical processes such as ocean waves. It can accurately characterize the damping dissipation law of the pile to the waves under different sea conditions and is highly consistent with the actual physical process.

[0020] The analytical expression for the Matern 5 / 2 kernel function is:

[0021] In the formula, and This represents two sets of input vectors with the same dimension and physical meaning, but different values. Let Euclidean distance be the distance between the input vectors. The standard deviation of the signal. The feature scale length.

[0022] Besides the GPR model, all machine learning models that satisfy the prediction logic can be expressed in a complete analytical formula and have no third-party runtime dependency (such as support vector regression, linear regression, multinomial regression, etc.) can be adapted to the model embedding requirements of this embodiment. This embodiment uses the GPR model as the optimal embodiment.

[0023] These data were then used to train a Gaussian process regression (GPR) machine learning model, achieving an extremely low root mean square error (0.0033m) using the Matern5 / 2 kernel function. Due to its nonparametric nature, estimable uncertainty, and excellent performance with small samples and nonlinear relationships, GPR became the optimal "surrogate model for wave height attenuation." This surrogate model can be invoked in real-time within the sea-air-wave coupling model to quickly provide the attenuated wave height.

[0024] Step 102: Embed the surrogate model into the near-surface layer parameterization module of the atmospheric component of the ocean-atmosphere-wave coupling model, and have the near-surface layer parameterization module call the surrogate model to predict the weakened effective wave height under the current sea state.

[0025] The core principle of this step is that the near-surface layer parameterization module of the WRF mode is the core calculation link for momentum and energy exchange at the sea-air interface. By embedding the GPR proxy model into this module, the dissipation effect of wind turbine piles on waves can be quantified in real time during the iteration process of each time step of the coupled mode, and fed back to the calculation links of sea surface roughness and wind field, realizing the full-chain bidirectional coupling of "wind turbine-wave-atmosphere".

[0026] In some embodiments, offline training of the surrogate model is completed based on a simulated dataset. The inverse of the kernel matrix, the input features of the training set, and the optimal hyperparameters of the kernel function required for model prediction are pre-calculated, and the pre-calculated data is stored in the form of a text file. The analytical expression of the surrogate model's prediction is hard-coded into the near-surface layer parameterization module of the atmospheric component of the ocean-atmosphere-wave coupled model using the Fortran language. When the model starts, the pre-calculated text file data is read into memory all at once.

[0027] Specifically, within the ocean-atmosphere-wave coupling framework, the ocean component (SWAN) outputs wave parameters for each wind turbine location in real time; these wave parameters are then passed to the near-surface layer parameterization module of the atmospheric component (WRF). This module embeds a trained Gaussian process regression (GPR) surrogate model, which can calculate the effective wave height after dissipation by the wind turbine structure.

[0028] To adapt to the WRF-based compilation and execution framework, and to balance prediction accuracy and computational efficiency, the GPR proxy model is embedded using an "offline training and pre-computation + online hard-coded invocation" approach. The specific implementation is as follows: (1) Offline pre-calculation and data file generation: Based on the wind turbine-wave interaction simulation dataset generated in step 101, offline training of the GPR model and pre-calculation of the core matrix are completed before the model runs: The training set is composed of the input features of the dataset (water depth, incident significant wave height, and spectral peak period). The training set is the output label (the weakened effective wave height). After completing the GPR model training, the optimal hyperparameters of the Matern 5 / 2 kernel function, namely the signal standard deviation, were obtained. Feature scale length ; Based on training set The kernel matrix of the training set is pre-computed using the optimal kernel function hyperparameters. And complete the matrix inversion to obtain ; Inverse the pre-calculated kernel matrix Training set input features The optimal hyperparameters of the kernel function are stored as two-dimensional text arrays (.txt format).

[0029] (2) Data file deployment and mode startup reading: Deploy the three pre-calculated data files to the specified working directory of WRF mode; During WRF mode operation, read the contents of the three data files into the mode memory through the file reading interface of the Fortran language, allocate them to the global array variable for storage, and use them for real-time prediction calls at each subsequent time step without needing to read or solve them again during mode operation.

[0030] (3) Online real-time prediction and coupled feedback: In the iteration process of each time step in the coupled mode, real-time prediction is completed according to the following process: The near-surface parameterization module receives real-time data from the SWAN ocean component of the COAWST coupled model, including water depth, incident significant wave height, and spectral peak period at the wind turbine location, and uses this data to form the input vector to be predicted. ; The training set is pre-stored in memory in the calling mode. Using the kernel function hyperparameters and the Matern 5 / 2 kernel function analytical expression, the cross-kernel matrix between the point to be predicted and the training set is calculated. ; Substituting the values ​​into the GPR model's prediction mean formula, and performing real-time matrix multiplication, we obtain the weakened effective wave height prediction value under the current sea state. The prediction formula is as follows: ; The predicted weakened effective wave height is directly input into the sea surface roughness length and friction velocity calculation stage of the near-surface layer parameterization module to complete the real-time feedback coupling of wind turbine pile-wave dissipation-atmospheric wind field, providing accurate wind field input for subsequent wind turbine power prediction.

[0031] Step 103: Based on the weakened effective wave height, update the atmospheric data in the near-surface layer parameterization module, and calculate the inflow wind speed based on the atmospheric data.

[0032] Specifically, the weakened wave height directly affects atmospheric data. Substituting the updated atmospheric data into the WRF's near-surface layer formula yields the corrected three-dimensional wind speed profile. In this way, by accurately reflecting the impact of wind turbines on wave energy dissipation at the near-surface layer, the model can adjust boundary layer friction and wind speed profiles in real time, helping to improve the accuracy of offshore wind power prediction.

[0033] Step 104: Predict the wind turbine power of the wind farm based on the inflow wind speed.

[0034] Specifically, the inflow wind speed is used as the input to the wind farm parameterization module to recalculate the momentum extraction, turbulence generation, and power generation of the wind farm, thereby completing a complete feedback loop from "structure influencing waves" to "waves influencing the atmosphere" and then to "atmosphere influencing the wind farm".

[0035] As described above, this embodiment establishes a two-way feedback mechanism of "wind turbine structure-wave-atmosphere-wind farm". By quantifying the dissipation effect of wind turbine piles on waves and dynamically feeding it back to the atmospheric wind field and wind farm calculations, it overcomes the physical mechanism deficiency of "one-way influence" in existing technologies. This enables the coupled model to simulate the secondary modulation of atmospheric boundary layer wind speed caused by changes in sea surface roughness, improving the accuracy of wind farm resource assessment, extreme load analysis, and regional environmental impact assessment. It solves the problems of existing technologies being unable to accurately assess the impact of wind farms on local marine meteorology and having biases in power prediction. Furthermore, the established GPR surrogate model avoids the high computational cost of directly solving wave-object interactions while ensuring the accuracy of wave height attenuation simulation, achieving a balance between accuracy and efficiency.

[0036] In one embodiment of this application, step 101 uses a wave structure interaction numerical model to generate a simulation dataset of wind turbines and waves, including steps 1011 to 1012.

[0037] Step 1011: For the target wind turbine type of the offshore wind farm, the water depth, incident significant wave height, and spectral peak period are used as input features. The input features are input into the wave structure interaction numerical model, and the weakened significant wave height is output.

[0038] Specifically, the wave structure interaction numerical model used in this invention is built upon the industry-standard SWAN third-generation wave model. Its core functionality is achieved through secondary development of the model's built-in vegetation module: an inertial force term is introduced to supplement the original drag term in the vegetation module, thereby accurately parameterizing the damping and dissipation effect of the offshore wind turbine pile structure on waves and fully reproducing the physical process of the pile-wave interaction. This model supports flexible adjustment of the core input parameters across the entire range, including the target sea area's static water depth, incident significant wave height, spectral peak period, and wind turbine pile diameter, comprehensively covering typical operating conditions across all scenarios of offshore wind farm construction and operation.

[0039] In the numerical calculation process, the model's computational grid size is set to match the diameter of the target wind turbine pile. Simultaneously, structural parameters (inertia coefficient, drag coefficient, and pile cylinder diameter) are determined for the selected wind turbine type (fixed pile or floating platform). Subsequently, typical sea state data for the region are collected: water depth, incident significant wave height, and spectral peak period. These three sea state data are used as input to drive the aforementioned high-precision wave-structure interaction numerical model. This model quantifies the energy dissipation effect of the pile structure on waves by solving the dynamic response of ocean waves near the wind turbine structure, outputting the weakened significant wave height under the corresponding operating conditions. This process is repeated under different sea state combinations, generating a large amount of effective sample data on wind turbine-wave coupling.

[0040] For example, to cover typical sea conditions of offshore wind farms in my country's nearshore waters, the parameter gradient can be set as follows: significant wave height range of 2 m to 4 m, sampling interval of 0.1 m; spectral peak period divided into four typical intervals: 7.4 s to 7.6 s, 8.4 s to 8.6 s, 9.6 s to 9.8 s, and 11.0 s to 11.2 s, sampling interval of 0.1 s; water depth range of 5 m to 50 m, sampling interval of 5 m; a total of 2520 sets (21×12×10) of numerical simulation experiments were carried out using the above parameter combinations. Each set of experiments outputs the significant wave height weakened by the dissipation of the wind turbine pile under the corresponding working condition, and finally forms a complete simulation dataset of wind turbine-wave interaction, which is used for subsequent training and verification of machine learning models.

[0041] Step 1012: Combine multiple sets of input features for the target wind turbine type and the corresponding weakened effective wave height to generate a simulated dataset of wind turbines and waves.

[0042] Specifically, for the selected target wind turbine type, the multiple sets of input features generated in step 1011 (i.e., different water depths, incident significant wave heights, and spectral peak periods) are summarized together with the corresponding attenuated significant wave heights. Each wave structure interaction numerical simulation generates a complete input-output record. All records are stored in tabular or database form to form a wind turbine-wave simulation dataset covering common sea conditions in the preset sea area.

[0043] As described above, by combining the wave structure interaction model of the target wind turbine type with three major input features—sea depth, incident significant wave height, and spectral peak period—the energy dissipation of the wind turbine structure on waves can be accurately calculated, yielding the weakened significant wave height. Subsequently, multiple input-output pairs are aggregated to form a complete simulation dataset covering typical sea conditions. This dataset provides high-quality training samples for subsequent Gaussian process regression (GPR) or other machine learning techniques, ultimately achieving rapid and interpretable surrogate modeling of the wind turbine-wave coupling effect.

[0044] In step 102, the near-surface layer parameterization module calls the surrogate model to predict the weakened effective wave height under the current sea state, including steps 1021 to 1023.

[0045] Step 1021: The near-surface layer parameterization module receives real-time data on the incident significant wave height, spectral peak period, and static water depth at the wind turbine location from the ocean component of the ocean-air-wave coupling model.

[0046] Specifically, within the ocean-air-wave coupled framework, the ocean component (SWAN) simulates the wave spectrum of each wind turbine location in real time. The incident significant wave height, spectral peak period, and static water depth output by the SWAN are transmitted to the near-surface layer parameterization module of the atmospheric component (WRF) via a coupling interface. This module receives and decodes this data at each time step, forming a feature vector corresponding to the wind turbine location, which is then passed to the surrogate model for dissipation prediction.

[0047] Step 1022: Input the real-time data of the ocean component into the proxy model to predict the weakened effective wave height under the current sea state.

[0048] Specifically, the feature vector is passed to a trained Gaussian process regression (GPR) surrogate model in real time. The surrogate model uses a pre-saved covariance matrix and kernel function (e.g., Matern5 / 2) to quickly predict the input vector, directly obtaining the effective wave height after dissipation by the wind turbine structure. This prediction process requires only one matrix-vector multiplication, with extremely low computational latency, and can update the effective wave height in a timely manner at each time step. This provides timely wave dissipation information to the near-surface layer parameterization module, thereby accurately correcting atmospheric data in the atmospheric boundary layer.

[0049] Step 1023: Based on the weakened effective wave height, update the sea surface roughness length and friction velocity in the atmospheric data in the near-surface layer parameterization module.

[0050] There are multiple schemes for calculating the roughness length in WRF. The specific scheme expression is as follows:

[0051] In the formula, Indicates the roughness length. Indicates the significant wave height. Indicates the wavelength of the spectral peak. Indicates the viscosity coefficient. This indicates the friction speed.

[0052] The formula for calculating friction speed is:

[0053] In the formula, This represents the von Karman constant (0.4). Indicates different heights Wind speed at the location, A function representing atmospheric stability.

[0054] Specifically, the near-surface layer parameterization module first uses the weakened effective wave height obtained from the surrogate model, updates the sea surface roughness length in real time using empirical formulas, and quickly solves for the friction velocity based on the friction velocity relationship. This update ensures that the atmospheric model can reflect the real-time impact of wave dissipation on surface roughness and friction stress at each time step, laying the foundation for subsequent wind speed field correction.

[0055] As described above, the near-surface layer parameterization module instantly acquires the incident significant wave height, spectral peak period, and static water depth from the SWAN output, and uses these ocean features as input to the surrogate model to calculate the significant wave height after dissipation by the wind turbine structure in real time. In this way, efficient and interpretable real-time wave attenuation prediction can be achieved, providing a more accurate physical basis for offshore wind power estimation, wind farm layout, and operation scheduling.

[0056] In one embodiment of this application, step 103, which involves incorporating wind speed into the atmospheric data, includes: calculating a wind speed vector based on the updated sea surface roughness length and the friction velocity, wherein the wind speed vector includes meridional wind. and latitude wind Decomposing wind speed into orthogonal zonal and meridional components for calculation is a native design requirement of the WRF mesoscale numerical model. The core reason is that the WRF model solves the original atmospheric equations, and its horizontal momentum equation naturally uses zonal and meridional momentum as independent solution variables, without a single "wind speed scalar" governing equation. All terms are solved discretely for each of the two components, and the component form is the foundation of the model's numerical calculations.

[0057] Specifically, the WRF model performs basic calculations of each component of the three-dimensional wind speed vector using the classic near-surface logarithmic law formula, and then combines numerical discretization methods to complete the final iterative solution. The core calculation formula is as follows:

[0058]

[0059] In the formula, Indicates the azimuth angle of the wind field.

[0060] The wind speed vector is passed to the wind farm parameterization module as the inflow wind speed, and the grid cells are calculated ( , , Zonal and meridional winds at location )

[0061]

[0062] The formula for calculating the trend of wind speed change is:

[0063] In the formula, Represents grid cells ( , , ) zonal wind, Represents grid cells ( , , Meridional wind at ) Represents grid cells ( , The number of wind turbines in ) Represents grid cells ( , , The new inflow velocity Indicates thrust coefficient Represents grid cells ( , , The area of ​​the wind turbine in ) This represents the thickness difference between two adjacent mode layers in the vertical direction.

[0064] The revised zonal and meridional wind trends are not only used for local power calculations, but are also transmitted back to the planetary boundary layer module of the WRF model in real time as the core boundary conditions for solving the atmospheric motion control equations in the next time step. This enables dynamic feedback of the wind farm's momentum extraction effect on the atmospheric boundary layer, allowing the model to self-consistently simulate the propagation and evolution of wind turbine wakes in the atmosphere.

[0065] As described above, directly reflecting wave energy dissipation in the near-surface wind speed field improves the spatiotemporal resolution and physical consistency of the wind field, thereby enhancing the accuracy of wind turbine power prediction, optimizing wind field layout and operation decisions, and providing reliable dynamic boundary conditions for environmental assessment and climate coupling of offshore wind power.

[0066] In one embodiment of this application, the calculation formula for step 104 is:

[0067] In the formula, Represents grid cells ( , , The power output of the wind farm at that location This represents the power factor.

[0068] In some embodiments, grid cells can also be predicted ( , , Turbulent kinetic energy at ) :

[0069] In the formula, Indicates the turbulent kinetic energy coefficient. The turbulent kinetic energy term will be fed back to the turbulence parameterization module of the WRF model to correct the turbulent diffusion coefficient and vertical mixing intensity within the boundary layer, further improving the simulation of the impact of wind farm wake turbulence on the atmospheric boundary layer structure, and ensuring the self-consistency of the model's entire physical process and the stability of long-term integration.

[0070] Please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating the implementation process of the wind farm parameterization scheme in the COAWST model. The implementation of the parameterization scheme follows the core principles of "offline training, online coupling, and dynamic feedback." Specifically, the implementation process mainly includes the following stages: 1. A regression learning model (GPR model) was trained through extensive numerical simulation experiments. After determining the inertial force coefficient, drag coefficient, and cylinder diameter, the energy dissipation caused by wave force is only related to water depth, incident significant wave height, and mean wave period (or spectral peak period). To cover common sea states in typical wind farm construction areas, the experimental parameter range was determined based on long-term wave reanalysis data for the region, and a series of ideal experiments with different water depths, incident significant wave heights, and mean wave periods were designed. Significant wave heights ranged from 2 m to 4 m, with intervals of 0.1 m. Spectral peak periods were 7.4 s to 7.6 s, 8.4 s to 8.6 s, 9.6 s to 9.8 s, and 11.0 s to 11.2 s, with intervals of 0.1 s, and water depths ranged from 5 m to 50 m, with intervals of 5 m (the above data selection was based on a pre-defined wave reanalysis dataset for the sea area). A total of 2520 experiments (21 × 12 × 10) were conducted. Then, simulated data at depths of 30 m, 40 m, and 50 m were selected to train several machine learning (regression) models, as these three depths would be used as validation data. The input variables were incident significant wave height, water depth, and spectral peak period, and the output variable was the attenuated significant wave height. These machine learning (regression) models can be categorized into four main types: linear regression models, regression tree models, support vector machines (SVM), and Gaussian process regression (GPR).

[0071] The performance of each model was evaluated using a five-fold cross-validation method, with root mean square error (RMSE) as the core evaluation metric. Figure 3 The figure shows a scatter plot comparing the fitted effective wave height and actual effective wave height of four machine learning models on the training set. Here, a represents stepwise linear regression, b represents fine-grained tree regression, c represents medium support vector machine regression, and d represents Matern 5 / 2 Gaussian process regression. The diagonal line in the figure represents the perfect fit between the fitted and actual values; the closer the scatter points are to the diagonal, the stronger the model's ability to fit the training data and the more accurately it captures the physical laws. Among them, the GPR model with the Matern 5 / 2 kernel function (covariance) is the best, with the smallest RMSE of 0.0033 m. This is because the GPR model's advantage lies mainly in handling nonlinearity and small sample data. Validation data was also used to analyze the advantages and disadvantages of these four machine learning models and to prevent overfitting during model training. The Matern 5 / 2 GPR model still performs best, while the stepwise linear regression model still performs worst. On the other hand, in some cases, the medium SVM model is not necessarily superior to the fine-grained tree model. By comprehensively comparing the RMSE, computational stability, and overfitting risk of each model, the GPR model with the Matern 5 / 2 kernel function was selected as the final proxy model for the wave height reduction effect.

[0072] 2. Embed the GPR model into the atmospheric component model (WRF).

[0073] Given the water depth, incident significant wave height, and spectral peak period at the wind turbine location, the effective wave height weakened by the wind turbine structure can be predicted using a trained GPR model. To enable the mesoscale model (WRF) to account for this process, the GPR model is implemented within the WRF.

[0074] A major advantage of using the GPR model is its interpretability. Since it involves matrix operations, the matrix data calculated through training only needs to be stored in a designated directory in text format, and then a GPR expression can be written using a programming language to read the matrix data from the specified path. Specifically, the GPR model is written into the Surface Layer Parameterization module of WRF, as this module is responsible for calculating roughness length and friction velocity. Because the input of the GPR model depends on the wave field simulated in real time by the SWAN model, this improved WRF must be coupled with SWAN within the COAWST framework. Static water depth data needs to be transferred to the WRF side.

[0075] 3. The near-surface layer parameterization module calculates the roughness length and friction velocity, among other variables, after the significant wave height changes. These variables are then passed to the planetary boundary layer (i.e., atmospheric boundary layer) parameterization module to calculate the affected three-dimensional wind speed, mainly zonal and meridional winds. These two winds are then passed to the wind farm parameterization module as the new inflow wind speed. Based on the updated inflow wind speed, the tendency terms of each variable in the wind farm parameterization module need to be updated synchronously.

[0076] To further verify the effect of the bidirectional coupling mechanism of this invention on the wind power prediction results, a controlled variable ideal experiment was designed: except for whether or not the "wind turbine foundation-wave-atmosphere" feedback module of this invention was introduced, the initial conditions, boundary conditions, physical process parameterization schemes, and grid settings of the two sets of experiments were completely identical. Comparing the computational domain settings of this ideal experiment with the wind power deviation, it can be seen that: the original scheme ignored the dissipation effect of the wind turbine foundation on waves, systematically overestimated the sea surface roughness length, and thus underestimated the near-surface wind speed and wind turbine power; the scheme of this embodiment, by introducing a bidirectional coupling mechanism, accurately quantifies the weakening effect of the wind turbine pile on waves and its feedback on the atmospheric wind field, corrects the systematic underestimation deviation of the original scheme, and makes the wind power prediction results more consistent with the actual physical process.

[0077] Through the above steps, the wave height reduction effect of wind turbine piles can be integrated into the COAWST model by solving the chain of dynamic roughness updates and wind field corrections within the model through the GPR proxy model. When the model is running, the secondary modulation of atmospheric boundary layer wind speed caused by changes in sea surface roughness can be simulated in the downstream region of the wind turbine, thereby realizing a fully coupled feedback simulation of "structure-wave-atmosphere-wind farm".

[0078] According to another aspect of the invention, Figure 4 This is a schematic diagram illustrating a wind turbine power prediction system 200 for an offshore wind farm according to an embodiment of the present invention. (Refer to...) Figure 4 The wind turbine power prediction system 200 includes a memory 201, a processor 202, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of a wind turbine power prediction method for an offshore wind farm as described above.

[0079] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting the power output of wind turbines in an offshore wind farm, characterized in that, include: For the target wind turbine type in offshore wind farms, a simulated dataset of wind turbines and waves is generated using a wave structure interaction numerical model. The simulated dataset is then used to train a machine learning model. The prediction logic of the machine learning model is expressed in a complete analytical expression without any third-party runtime dependencies. The machine learning model is used as a proxy model for wave height attenuation effect. The proxy model is embedded into the near-surface layer parameterization module of the atmospheric component of the ocean-atmosphere-wave coupling model, and the near-surface layer parameterization module calls the proxy model to predict the weakened effective wave height under the current sea state. Based on the weakened effective wave height, atmospheric data is updated in the near-surface layer parameterization module, and inflow wind speed is calculated based on the atmospheric data. The wind turbine power of the wind farm is predicted based on the inflow wind speed; A simulation dataset of wind turbines and waves was generated using a numerical model of wave structure interactions, including: For the target wind turbine type in offshore wind farms, water depth, incident significant wave height, and spectral peak period are used as input features. These input features are then input into a wave structure interaction numerical model, and the weakened significant wave height is output. By combining multiple sets of input features for the target wind turbine type and the corresponding weakened significant wave height, a simulation dataset of wind turbine and wave interaction is generated. The near-surface layer parameterization module calls the surrogate model to predict the weakened effective wave height under the current sea state, including: The near-surface layer parameterization module receives real-time data on the incident significant wave height, spectral peak period, and static water depth at the wind turbine location from the ocean component of the ocean-air-wave coupled model; the real-time data from the ocean component is input into the surrogate model to predict the weakened significant wave height under the current sea state; Based on the weakened effective wave height, the sea surface roughness length and friction velocity in the atmospheric data are updated in the near-surface layer parameterization module.

2. The method for predicting wind turbine power in offshore wind farms according to claim 1, characterized in that, Training a machine learning model using the simulated dataset includes: The simulated dataset was used to train a machine learning model expressed by a complete mathematical analytical expression; The machine learning model is used to fit the functional relationship between the input features and the weakened effective wave height.

3. The method for predicting wind turbine power in offshore wind farms according to claim 2, characterized in that, The machine learning model can be any one of the following: Gaussian process regression model, support vector regression model, linear regression model, or multinomial regression model.

4. The method for predicting wind turbine power in offshore wind farms according to claim 3, characterized in that, If a Gaussian process regression model with a Matern 5 / 2 kernel function is selected, the analytical expression of the Matern 5 / 2 kernel function is: In the formula, and Represents two sets of input vectors, Let Euclidean distance be the distance between the input vectors. The standard deviation of the signal. The feature scale length.

5. The method for predicting wind turbine power in offshore wind farms according to claim 1, characterized in that, Embedding the surrogate model into the near-surface layer parameterization module of the atmospheric component of the ocean-atmosphere-wave coupled model includes: The proxy model is trained offline based on the simulated dataset. The inverse kernel matrix, training set input features, and optimal hyperparameters of the kernel function required for model prediction are pre-calculated, and the pre-calculated data is stored in the form of a text file. The predictive analytical formula of the proxy model is hard-coded into the near-surface layer parameterization module of the atmospheric component of the ocean-atmosphere-wave coupled model using the Fortran language. When the model starts, the pre-calculated text file data is read into memory all at once.

6. The method for predicting wind turbine power in offshore wind farms according to claim 1, characterized in that, The inflow wind speed is calculated based on the atmospheric data, including: The wind speed vector is calculated based on the updated sea surface roughness length and the friction velocity, and the wind speed vector includes meridional wind and zonal wind. The wind speed vector is passed to the wind farm parameterization module as the inflow wind speed, and the grid cells are calculated ( , , Zonal and meridional winds at location ) The formula for calculating the trend of wind speed change is: In the formula, Represents grid cells ( , , ) zonal wind, Represents grid cells ( , , The direction of the wind at that location, Represents the wind speed vector. Represents grid cells ( , The number of wind turbines in ) Represents grid cells ( , , The new inflow wind speed in the ) Indicates the thrust coefficient. Represents grid cells ( , , The area of ​​the wind turbine in ) This represents the thickness difference between two adjacent mode layers in the vertical direction.

7. The method for predicting wind turbine power in offshore wind farms according to claim 6, characterized in that, Predicting the wind turbine power of a wind farm based on the inflow wind speed includes: In the formula, Represents grid cells ( , , The power output of the wind farm at that location This represents the power factor.

8. The method for predicting wind turbine power in offshore wind farms according to claim 7, characterized in that, Predicting the wind turbine power of a wind farm based on the inflow wind speed also includes: Predictive grid cells ( , , Turbulent kinetic energy at ) : In the formula, Indicates the turbulent kinetic energy coefficient. .

9. A wind turbine power prediction system for an offshore wind farm, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the wind turbine power prediction method for an offshore wind farm as described in any one of claims 1 to 8.

Images