Sea floating type fan hybrid model test method and device and readable storage medium

Abstract

The invention provides an offshore floating fan hybrid model test method and device and a readable storage medium, for each preset working condition in a working condition set, a physical model is driven by using the preset working condition, so that the physical model generates initial motion. And the electronic equipment generates standard data with continuous time according to the motion data of the physical model performing the initial motion, and inputs the standard data of the corresponding time point into the AI proxy model when the AI proxy model calls the motion data every time, so that the AI proxy model at least outputs the predicted aerodynamic load. The electronic equipment drives the physical model according to the aerodynamic load, so that the physical model generates motion response; and updating the standard data with continuous time according to the motion response until the preset working condition ends. By adopting the scheme, the AI agent model is used for replacing high-frequency iteration, so that high-frequency iterative calculation is bypassed, the delay is reduced to a great extent, and the aim of improving the accuracy of a test result of a hybrid model test is fulfilled.

Description

Hybrid model test method, equipment and readable storage medium for offshore floating wind turbines Technical Field

[0001] This application relates to the field of offshore floating wind turbine technology, and in particular to a hybrid model test method, equipment and readable storage medium for offshore floating wind turbines. Background Technology

[0002] Compared to onshore wind power, offshore wind energy has advantages such as high quality, high wind speed, and no land occupation, making offshore wind power generation more efficient than onshore wind power. Model testing of offshore floating wind turbines is crucial to ensuring their safe and stable operation at sea.

[0003] Common model tests include physical model tests and hybrid model tests. Hybrid model tests combine numerical simulation and physical models, applying simulation outputs to the physical model through a "hardware-in-the-loop" approach. This enables data transfer between the physical model and the numerical simulation, thereby improving the efficiency and reliability of hybrid model tests.

[0004] However, hybrid model experiments rely on the construction of numerical subsystems, which require high-frequency iterative nonlinear equations, resulting in large delays and accumulated errors, thus leading to inaccurate experimental results from hybrid model experiments. Summary of the Invention

[0005] This application provides a hybrid model test method, equipment, and readable storage medium for offshore floating wind turbines. By utilizing an AI proxy model to replace high-frequency iteration, the high-frequency iteration calculation is bypassed, greatly reducing latency and improving the accuracy of the test results of the hybrid model test.

[0006] In a first aspect, this application provides a hybrid model testing method for offshore floating wind turbines, comprising: for each preset working condition in a set of working conditions, driving a physical model using the preset working condition to induce initial motion in the physical model, wherein the physical model is a scaled-down test model of the offshore floating wind turbine; generating time-continuous standard data based on the motion data of the physical model undergoing the initial motion, wherein the time points of the time-continuous standard data correspond one-to-one with the times when an AI agent model calls the motion data; inputting the standard data at the corresponding time point into the AI ​​agent model each time the AI ​​agent model calls the motion data, so that the AI ​​agent model outputs at least the predicted aerodynamic load; driving the physical model according to the aerodynamic load to induce a motion response in the physical model; updating the time-continuous standard data according to the motion response, until the preset working condition ends and the AI ​​agent model outputs the test results of the preset working condition.

[0007] Secondly, this application provides a hybrid model test device for offshore floating wind turbines, comprising: a driving module, used to drive a physical model for each preset working condition in a set of working conditions, so that the physical model generates initial motion, the physical model being a scaled-down test model of an offshore floating wind turbine; a generation module, used to generate time-continuous standard data based on the motion data of the physical model undergoing the initial motion, wherein the time points of the time-continuous standard data correspond one-to-one with the times when the AI ​​agent model calls the motion data; a processing module, used to input the standard data at the corresponding time point into the AI ​​agent model each time the AI ​​agent model calls the motion data, so that the AI ​​agent model outputs at least the predicted aerodynamic load; the driving module is further used to drive the physical model according to the aerodynamic load, so that the physical model generates a motion response; and an updating module, used to update the time-continuous standard data according to the motion response, until the preset working condition ends and the AI ​​agent model outputs the test results of the preset working condition.

[0008] Thirdly, this application provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the electronic device implements the method described in the first aspect or various possible implementations of the first aspect.

[0009] Fourthly, this application provides a computer-readable storage medium storing computer instructions that, when executed by a processor, are used to implement the method described in the first aspect or various possible implementations of the first aspect.

[0010] Fifthly, this application provides a computer program product comprising a computing program, which, when executed by a processor, implements the method described in the first aspect or various possible implementations of the first aspect.

[0011] This application provides a method, equipment, and readable storage medium for hybrid model testing of offshore floating wind turbines. For each preset working condition within a set of working conditions, the preset working condition drives the physical model, causing the physical model to generate initial motion. Subsequently, electronic equipment generates time-continuous standard data based on the motion data of the physical model undergoing initial motion. Each time the AI ​​proxy model calls the motion data, the standard data at the corresponding time point is input into the AI ​​proxy model, ensuring that the AI ​​proxy model outputs at least the predicted aerodynamic load. After outputting the aerodynamic load, the electronic equipment drives the physical model based on the aerodynamic load, causing the physical model to generate a motion response; the time-continuous standard data is updated based on the motion response until the preset working condition ends and the AI ​​proxy model outputs the test results for the preset working condition. This approach, by using an AI proxy model instead of high-frequency iteration, bypasses high-frequency iterative calculations, significantly reducing latency and improving the accuracy of the hybrid model test results. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 is a schematic diagram of the network architecture of the hybrid model test method for offshore floating wind turbines provided in this application; Figure 2 is a flowchart of the hybrid model test method for offshore floating wind turbines provided in this application; Figure 3 is another network architecture diagram of the hybrid test model for offshore floating wind turbines provided in this application; Figure 4 is a schematic diagram of the model structure of the AI ​​agent model in the hybrid model test method for offshore floating wind turbines provided in this application; Figure 5 is a schematic diagram of the hybrid model test device for offshore floating wind turbines provided in this application; Figure 6 is a schematic diagram of the structure of the electronic equipment provided in this application. Detailed Implementation

[0014] Offshore wind power is a crucial renewable energy source, playing a vital role in optimizing energy structure and mitigating climate change. With advancements in technology and the economy, global wind power equipment is increasingly trending towards larger scale and offshore deployment. For the development and utilization of wind energy in deep-sea areas, deploying large offshore floating wind turbines is one of the best solutions, offering both social and economic benefits. However, the marine environment is complex and variable. To better ensure the structural integrity and operational safety of offshore wind turbines throughout their service life, it is necessary to strengthen research on the structural performance and coupling mechanisms of offshore wind turbines in complex marine environments. Common research methods include physical model testing and hybrid model testing.

[0015] Physical model testing refers to directly simulating all physical processes in a test environment without relying on numerical calculations. Hybrid model testing refers to integrating numerical simulation with a physical model, applying the numerical simulation results to the physical model through "hardware-in-the-loop" technology, enabling data transfer between the physical model and the numerical simulation, and greatly improving the efficiency and reliability of hybrid model testing.

[0016] In traditional hybrid model testing, the floating wind turbine system is decoupled into a numerical substructure and a physical substructure. The numerical substructure simulates the aerodynamic loads on the turbine blades in real time and converts these loads into control commands, which then drive the blade simulation devices in the physical substructure. The calculations in the numerical substructure, for example, are based on numerical solutions of Blade Element Momentum (BEM) theory. This numerical solution requires iterating nonlinear equations, resulting in a computational delay greater than 10 milliseconds. This leads to a discrepancy between the physical and numerical interfaces, causing accumulated motion feedback errors. For example, the tower velocity phase shift may exceed 15 degrees. Furthermore, due to the limitations of BEM iteration efficiency, complex operating conditions are prone to divergence.

[0017] In the above scheme, the BEM-based aerodynamic model is the core model for numerical simulation of hybrid model tests, mainly used to calculate aerodynamic loads, etc. To reduce latency, a common method is to simplify the aerodynamic model. However, the simplified aerodynamic model suffers a 20% drop in accuracy due to neglecting higher-order efficiencies such as dynamic stall. Clearly, traditional hybrid model testing methods conflict between real-time performance and accuracy, failing to achieve both simultaneously.

[0018] Based on this, this application provides a hybrid model test method, equipment and readable storage medium for offshore floating wind turbines. By using an AI proxy model to replace high-frequency iteration, the high-frequency iteration calculation is bypassed, which greatly reduces the latency and achieves the goal of improving the accuracy of the test results of the hybrid model test.

[0019] This application applies to the field of offshore floating wind turbine engineering, and relates to the verification of the dynamic performance of offshore floating wind turbines in complex marine environments. Wind-wave-current coupling is a core dynamic characteristic of complex marine environments. Figure 1 is a schematic diagram of the network architecture of the hybrid model test method for offshore floating wind turbines provided in this application. Referring to Figure 1, the network architecture includes electronic devices 11, a physical model 12, and sensors 13. A pre-trained AI agent model is deployed on the electronic devices 11, and the sensors 13 are set around the physical model 12 or physical module 12. The physical model 12 is a scaled-down test model of the offshore floating wind turbine, used to reproduce the dynamic characteristics of the offshore floating wind turbine, etc. The physical model 12 includes a floating body, tower, blade simulation device, and test pool, etc. The sensors 13 are used to collect motion data of the physical model 12, such as tower top velocity and floating body attitude.

[0020] In this application, the electronic device 11 possesses computing and storage capabilities, and can be a mobile phone, computer, server, edge node, or central node in a Content Delivery Network (CDN). During the hybrid model experiment, the physical model 12 undergoes initial motion driven by a preset operating condition, and each sensor 13 collects motion data and stores it in a circular data buffer. When the AI ​​agent model on the electronic device 11 calls data, the electronic device performs linear interpolation to fill in missing time points using a sliding window mechanism to obtain standard data. The time points of this standard data correspond one-to-one with the time when the AI ​​agent model calls the motion data. The electronic device 11 uses the standard data as input to the AI ​​agent model, ensuring that the AI ​​agent model outputs at least the aerodynamic load at the corresponding time point. The aerodynamic load is used to drive the physical model to generate a motion response. The sensors collect motion data, which is then interpolated and filled in as standard data, and this data is continued to be provided to the AI ​​agent model, causing the AI ​​agent model to output aerodynamic loads, etc., continuously cycling until the preset operating condition ends. For example, if the preset operating condition lasts for 10 minutes, the hybrid model experiment ends after 10 minutes. The electronic device outputs experimental results such as the aerodynamic load curve within 10 minutes. In addition to aerodynamic loads, the AI ​​proxy model also outputs platform position coordinates, mooring tension, etc.; correspondingly, the test results include platform motion trajectory, mooring tension characteristics, etc.

[0021] After obtaining the test results for each preset operating condition, the electronic device 11 determines whether the test results enable the offshore floating wind turbine to operate safely and stably. If the test results for one or more preset operating conditions indicate that the offshore floating wind turbine cannot operate safely and stably under the corresponding operating conditions, then it is necessary to optimize the structure of the offshore floating wind turbine, etc.

[0022] The hybrid model test method for offshore floating wind turbines of this application will be described in detail below, based on the implementation environment shown in Figure 1. For example, please refer to Figure 2, which is a flowchart of the hybrid model test method for offshore floating wind turbines provided in this application. This application is applied to electronic devices, and this embodiment includes: 201. For each preset working condition in the set of working conditions, the preset working condition is used to drive the physical model, so that the physical model generates initial motion. The physical model is a scaled-down test model of the offshore floating wind turbine.

[0023] In this application, the physical model is a scaled-down test model of an offshore floating wind turbine, used to reproduce the dynamic characteristics of the offshore floating wind turbine. The physical model, also known as the physical layer, includes a floating body, tower, blade simulation device, and test pool, etc., with the test pool used to generate waves, etc.

[0024] This application pre-sets multiple preset operating conditions, which cover as many possible operating conditions that offshore floating wind turbines may encounter, such as normal, complex, and rare conditions. For example, the normal operating conditions meet the following criteria: wind speed range of 3-15 meters per second (m / s), significant wave height range of 1-5 meters, and turbulence intensity of 5%-15%. Under normal operating conditions, the blades rotate normally to generate electricity, the platform sways slightly, and there is no dynamic stall.

[0025] The complex operating conditions are as follows: wind speed range of 15-25 m / s, effective wave height range of 5-8 meters, and turbulence intensity of 15%-30%. Under these conditions, the blades experience dynamic stall, and the platform sway intensifies.

[0026] Rare operating conditions, also known as extreme operating conditions, are characterized by wind speeds of 35-50 m / s, significant wave heights of 10-15 meters, and turbulence intensity of 30%-40%. Under these conditions, the wind turbine may shut down to avoid the wind, and the platform may pitch or yaw significantly.

[0027] 202. Generate time-continuous standard data based on the motion data of the physical model performing the initial motion, wherein the time points of the time-continuous standard data correspond one-to-one with the times when the AI ​​agent model calls the motion data.

[0028] In this application, motion data includes tower top velocity, floating body attitude, etc. Tower top velocity is the six-degree-of-freedom velocity component collected by the sensor, and floating body attitude includes the six-degree-of-freedom displacement, angular velocity, angular displacement, etc., collected by the sensor. Due to the different collection frequencies of motion data by the sensor and the call frequencies of motion data by the AI ​​agent model, data gaps or time misalignments exist. For example, the sensor collects motion data every 10 milliseconds, while the AI ​​agent model calls motion data every 55 milliseconds, resulting in some call times lacking corresponding motion data. To address this, the motion data collected by the sensor is stored in a circular data buffer, and the electronic device generates time-continuous standard data based on the buffered motion data. Time-continuous standard data means that a standard data queue is generated and continuously updated according to the call times of the AI ​​agent model. Each standard data in the standard data queue corresponds to a call time, that is, there is a standard data every 55 milliseconds, such as 55 milliseconds, 110 milliseconds, 165 milliseconds, etc. The standard data at the 55th millisecond is generated by interpolation using the motion data collected by the sensor at the 50th and 60th milliseconds, and the standard data at the 110th millisecond is the motion data collected by the sensor at the 110th millisecond.

[0029] As can be seen from the above, in this application, when the AI ​​agent model calls motion data, linear interpolation is performed to fill in missing time points based on the sliding window mechanism to ensure the temporal continuity of the AI ​​agent model's input. This can solve the problem of temporal mismatch between the motion data collected by the sensor and the call of the AI ​​agent model, greatly ensuring real-time performance and thus ensuring the reliability of the experimental results of the hybrid model experiment.

[0030] 203. Each time the AI ​​agent model calls the motion data, the standard data at the corresponding time point is input into the AI ​​agent model so that the AI ​​agent model outputs at least the predicted aerodynamic load.

[0031] In this application, the AI ​​agent model periodically calls motion data. At each call, standard data corresponding to the time point is input into the AI ​​agent model, enabling the AI ​​agent model to output load-type data, predicted motion data, etc. Load-type data includes at least aerodynamic loads; optionally, it also includes hydrodynamic loads, mooring system loads, etc. Hydrodynamic loads refer to the changes in wave excitation force and tidal drag force on the floating body over time, while mooring system loads include the tension change curve of the mooring cable, etc. Predicted motion data includes the displacement and velocity of the floating body, and the displacement of the tower top, etc.

[0032] In this application, the AI ​​proxy model is an end-to-end model, and its structure includes, but is not limited to, feedforward neural networks, convolutional neural networks, recurrent neural networks, or any combination thereof. The AI ​​proxy model directly maps inputs such as platform motion velocity and wave spectrum to outputs such as aerodynamic loads, bypassing the BEM iteration process. In other words, the AI ​​proxy model directly maps standard data to load-type data.

[0033] 204. Drive the physical model according to the aerodynamic load so that the physical model generates a motion response.

[0034] In this application, the output of the AI ​​agent model includes at least aerodynamic loads, which are applied to the physical model by an electronic device, causing the physical model to generate a motion response.

[0035] 205. Update the time-continuous standard data according to the motion response until the preset working condition ends and the AI ​​agent model outputs the test results of the preset working condition.

[0036] In this application, since the preset working condition corresponds to a duration, such as 10 minutes or half an hour, during the duration of the preset working condition, the sensor continuously updates the motion data in the annular data buffer based on the first-in-first-out principle. Correspondingly, the time-continuous standard data is also continuously updated, so that the AI ​​agent model continuously outputs aerodynamic loads and drives the physical model until the preset working condition ends. The electronic device then obtains a test result of the preset working condition, such as the aerodynamic load change curve, the platform motion trajectory, and the mooring tension characteristics.

[0037] After completing a mixed test model for one preset working condition, the process continues to the next preset working condition until every preset working condition in the set of working conditions has been traversed. The number of preset working conditions in the set of working conditions can be several, dozens, hundreds, etc., and this application does not limit this number.

[0038] After traversing all the preset operating conditions, if the test results of one or more preset operating conditions indicate that the offshore floating wind turbine cannot operate safely and stably under that preset operating condition, then the offshore floating wind turbine needs to be optimized, the physical model updated based on the optimized offshore floating wind turbine, and the hybrid model test conducted again until the offshore floating wind turbine can operate safely and stably under each preset operating condition.

[0039] The hybrid model testing method for offshore floating wind turbines provided in this application uses each preset working condition within a set of working conditions to drive the physical model, causing the physical model to generate initial motion. Then, the electronic equipment generates time-continuous standard data based on the motion data of the physical model undergoing initial motion. Each time the AI ​​proxy model calls for motion data, the standard data at the corresponding time point is input into the AI ​​proxy model, ensuring that the AI ​​proxy model outputs at least the predicted aerodynamic load. After outputting the aerodynamic load, the electronic equipment drives the physical model based on the aerodynamic load, causing the physical model to generate a motion response; the time-continuous standard data is updated based on the motion response until the preset working condition ends and the AI ​​proxy model outputs the test results for the preset working condition. This approach, by using an AI proxy model to replace high-frequency iteration, bypasses high-frequency iterative calculations, significantly reducing latency and improving the accuracy of the hybrid model test results.

[0040] Figure 3 is another network architecture diagram of the hybrid test model of the offshore floating wind turbine provided in this application. Referring to Figure 3, the area above the dashed line represents the usage process of the AI ​​agent model, i.e., the process of conducting hybrid model experiments based on the AI ​​agent model, and the area below the dashed line represents the training process of the AI ​​agent model. For the usage process, the network architecture diagram includes a physical layer, an interaction layer, and an intelligence layer. Among them, the physical layer is the aforementioned physical model, including the floating body, tower, blade simulation device, and test pool, which is used to generate waves.

[0041] The interaction layer comprises various sensors that transmit motion data to the intelligent layer based on an information exchange protocol. This motion data includes, but is not limited to, platform attitude and tower top velocity. The interaction layer addresses the timing mismatch between the motion data collected by the sensors and the AI ​​agent model calls in the intelligent layer through an adaptive time step alignment protocol. Based on this protocol, real-time motion data collected by the sensors is stored in a circular data buffer, which can be located in either the interaction layer or the intelligent layer.

[0042] When the AI ​​agent model in the intelligent layer retrieves motion data, it performs linear interpolation to fill in missing time points using a sliding window mechanism, obtaining time-continuous standard data to ensure the temporal continuity of the AI ​​agent model's input. At each invocation time, the intelligent layer retrieves the standard data and outputs aerodynamic loads, etc. This aerodynamic load is converted from analog to digital to obtain a load command, which drives the blade simulation device, causing the physical layer to generate a motion response. Sensors in the interaction layer continue to collect motion data, updating the motion data within the annular buffer. When the electronic device retrieves motion data, it performs linear interpolation to continuously update the time-continuous standard data… until the preset operating condition ends.

[0043] Optionally, in addition to outputting aerodynamic loads, the AI ​​proxy model also outputs predicted motion data from the physical model. Simultaneously, the electronic device uses sensors to acquire actual motion data from the physical model, compares the predicted motion data with the actual motion data, and obtains a comparison result. Based on the comparison result, the electronic device determines whether the error of the predicted motion data exceeds a preset error. If the comparison result indicates that the error of the predicted motion data exceeds the preset error, the AI ​​proxy model is optimized; if the comparison result indicates that the error of the predicted motion data does not exceed the preset error, the aerodynamic load, etc., are output.

[0044] For example, the AI ​​agent model can also output aerodynamic loads based on standard data such as the input platform attitude and tower top velocity. Optionally, the AI ​​agent model can also obtain predicted motion data, such as the moving speed and acceleration of the float and tower, based on the input standard data such as the platform attitude and tower top velocity. The electronic device compares the predicted motion data with the actual motion data measured by the sensors. If the error between the two is large, online learning is triggered to optimize the AI ​​agent model; if the error between the predicted motion data and the actual motion data is small, the aerodynamic loads obtained in this experiment are output.

[0045] Using this approach, electronic devices compare the predicted motion data output by the AI ​​agent model with the actual motion data measured by sensors. Based on the comparison results, they decide whether to optimize the AI ​​agent model. By dynamically updating the AI ​​agent model, it adapts to environmental changes and improves the accuracy of hybrid model experiments.

[0046] In this application, an AI agent model in the intelligent layer is used to replace the iterative calculation based on BEM in the traditional hybrid experimental model, thereby improving the response speed by 10-100 times. The training and usage processes of the AI ​​agent model are described in detail below.

[0047] First, the training process of the AI ​​agent model.

[0048] In this application, the training phase includes a pre-training phase and a fine-tuning phase. Correspondingly, the training dataset includes a numerical simulation dataset for the pre-training phase and a historical experimental dataset for the fine-tuning phase, as shown in Figure 3. The electronic device generates various operating conditions based on wind power equipment design data and external environmental data, including rare conditions, through numerical simulation. For each operating condition, the electronic device extracts sample motion data and sample loads to obtain multiple sample pairs, which form the numerical simulation dataset. Then, the electronic device uses multiple sample pairs from the numerical simulation dataset to train an initial model to obtain an intermediate model. After training the intermediate model, the electronic device uses historical pairs from the historical experimental dataset to fine-tune the intermediate model to obtain an AI proxy model. The electronic device integrates numerical simulation and physical experimental data using a multi-fidelity fusion strategy to cover all operating conditions.

[0049] In the process of generating numerical simulation datasets for electronic equipment, several operating conditions are calculated using numerical simulation software, covering wind speeds of 3-50 m / s, significant wave heights of 1-15 m, and turbulence intensities of 5%-40%. Based on numerical simulation, nonlinear aerodynamic effects such as dynamic stall and vortex shedding, as well as extreme operating conditions such as typhoons, are accurately captured.

[0050] In the process of creating the historical test dataset, the floating platform and the top of the tower were collected under various working conditions by using a water tank test and deploying a six-component force sensor and an optical tracking system. The data included several sets of historical motion data and corresponding historical loads.

[0051] Using this approach, electronic devices generate various operating conditions based on numerical simulation, and generate sample pairs based on the various operating conditions obtained from the simulation. This approach is compatible with nonlinear and extreme operating conditions, significantly reduces the investment in physical experiments, lowers the cost of training samples, and simultaneously improves the training speed of AI agent models.

[0052] Optionally, in the above embodiments, after the electronic device generates multiple operating conditions based on numerical simulation, it also labels the rare and normal operating conditions among the multiple operating conditions to obtain a labeled dataset. Then, the electronic device trains an adversarial network based on the labeled dataset and uses the trained adversarial network to generate a typhoon-distorted wave joint operating condition.

[0053] In this application, to overcome the bottleneck of scarce data for rare operating conditions, a generative adversarial network (GAN) is introduced to synthesize typhoon-abnormal wave joint operating conditions. During the synthesis of the typhoon-abnormal wave joint operating conditions, the electronic device extracts core parameters such as wind speed, turbulence intensity, and wave height from normal and rare operating conditions, and labels each core parameter; for example, 0 represents normal operating conditions, 1 represents typhoon operating conditions, and 2 represents abnormal wave operating conditions. Then, the electronic device builds an adversarial network based on the labeled data, which includes a generator and a discriminator. After building the adversarial network, the electronic device iteratively optimizes the adversarial network to train a mature adversarial network. After training the adversarial network, the typhoon-abnormal wave joint operating conditions are generated using the trained adversarial network.

[0054] By adopting this approach, electronic devices utilize adversarial networks to generate joint typhoon-distorted wave operating conditions, overcoming the bottleneck of scarce rare operating condition data and increasing the amount of extreme operating condition data by 5 times, which greatly improves the accuracy of AI agent models.

[0055] Optionally, in the above embodiments, after the electronic device creates a numerical simulation dataset and a historical test dataset, it also injects ±0.5% Gaussian noise using stochastic differential equations (SDE) to simulate sensor errors and enhance the robustness of the AI ​​proxy model to measured disturbances.

[0056] Figure 4 is a schematic diagram of the AI ​​proxy model in the hybrid model test method for offshore floating wind turbines provided in this application. Referring to Figure 4, the AI ​​proxy model includes an input layer, an encoding layer, a physical calculation layer, and an output layer.

[0057] During training, the physical model moves under the influence of sample conditions. Parameters of these sample conditions include, but are not limited to, wind field information and wave information. Wind field information includes turbulence intensity, wind speed, and wind frequency, while wave information includes, but is not limited to, wave height and wave period. Each sample pair includes sample motion data and sample loads. Sample motion data includes tower top velocity and floating body attitude. Tower top velocity is the six-degree-of-freedom velocity component acquired by sensors, and floating body attitude includes six-degree-of-freedom displacement, angular velocity, and angular displacement acquired by sensors. The sample motion data is passed from the input layer to the encoding layer.

[0058] The encoding layer contains several fully connected layers, such as FC1 to FCn in the figure. The sample motion data passes through multiple fully connected layers to obtain feature vectors. These feature vectors are then passed to the output layer. The output layer contains multiple fully connected layers and regression layers. After processing by each fully connected layer and regression layer, the feature vectors generate a preliminary payload vector.

[0059] The sample motion data from the input layer is also fed into the physics computation layer. After the physics cells in the physics computation layer calculate the sample residuals, the multiplication unit multiplies these residuals by a weight λ to obtain a correction term, which is then passed to the output layer for correction. In the output layer, the initial load vector output from the regression layer is added to the correction term to obtain the corrected load vector. Based on this corrected load vector, the corrected aerodynamic loads, etc., are derived. The electronic device continuously adjusts the parameters of the initial model based on the difference between the corrected load and the sample load corresponding to the sample motion data, ensuring that the corrected load output by the initial model approximates the sample load as closely as possible, until an intermediate model is trained.

[0060] Referring to Figure 4, this application introduces sample residuals during the model training stage, that is, constructing motion correction equations based on the sample motion data; constructing the residual function of the physical computing layer based on leaf element momentum theory and the motion correction equations; and determining the sample residuals based on the residual function.

[0061] For example, the electronic device embeds the leaf element momentum theory and the motion correction equations in a differentiable form into the AI ​​agent model, and the residual function is shown in Equation (1): Formula (1) where, This represents the residual; during model training, it is the sample residual, and during model usage, it is the physical residual. The axial induction factor represents the impeller's rotating surface. The tangential induction factor represents the impeller's rotating surface. Indicates leaf density. Indicates the normal force coefficient of the impeller plane. Indicates the impeller plane tangential force coefficient. Indicates the inflow wind speed. Indicates the velocity at the top of the tower. This indicates the relative inflow velocity.

[0062] It can be determined by the following formula (2): Formula (2) where, Indicates the impeller speed. This represents the radius of the impeller's rotating surface.

[0063] Please refer to formula (1). and Based on the leaf element momentum theory, This is the motion correction equation. After calculating the sample residuals, multiplying the sample residuals by the weighting coefficient λ yields the correction term. The correction term is added to the initial load vector to compensate for the force loop hysteresis caused by dynamic stall.

[0064] This approach, by introducing residuals, forces the output to satisfy momentum-energy conservation while retaining the flexibility of data-driven approaches, thus avoiding the physical paradoxes of black-box models and improving the accuracy of AI proxy models.

[0065] Below, based on the model architecture shown in Figure 4, the training process of the AI ​​agent model will be explained in detail.

[0066] This application employs a phased transfer learning and target loss optimization strategy during training. Phased transfer learning refers to the training process of the AI ​​agent model including a pre-training phase and a fine-tuning phase. The pre-training phase yields an intermediate model, and the fine-tuning phase fine-tunes the intermediate model to obtain the AI ​​agent model. During the pre-training phase, the electronic device inputs the sample motion data from the sample pairs into the initial model to obtain a sample load vector. This allows the physical computation layer of the initial model to determine the sample residuals. A loss function is constructed based on the sample load vector, the sample loads in the sample pairs, and the sample residuals. The parameters of the initial model are then adjusted according to the loss function to obtain the intermediate model.

[0067] In this application, the electronic device constructs a loss function based on sample pairs in the numerical simulation dataset, using mean-square error (MSE) and physically conserved residuals. As shown in the following formula (3): Formula (3) where, , , Indicates the weighting coefficient. This indicates that the sample motion data is input into the initial model to obtain the sample load vector. That is, the initial load vector in Figure 4 represents the sample load vector during the model training process. This represents the simulation vector corresponding to the sample load obtained from the simulation of sample motion data during the numerical simulation process. This represents the sample residuals calculated by the physical computation layer of the initial model when the input is sample motion data. This represents the gradient of the network parameters. In formula (3), Indicates data matching items, Represents physical constraint terms. This represents the gradient smoothing term. The physical meaning of the data matching term is: Forced approximation of the simulation vector corresponding to the sample motion data The physical meaning of the physical constraint term is to ensure that the output satisfies the BEM equation. The physical meaning of the gradient constraint term is to avoid the optimization getting stuck in a saddle point.

[0068] This approach utilizes sample load vectors, sample loads in sample pairs, and sample residuals to construct a loss function. This loss function includes data matching terms, physical constraint terms, and gradient constraint terms, thus balancing the model's data fitting accuracy, physical rationality, and practicality. It avoids overfitting, violation of physical laws, and instability caused by a single loss, thereby improving the accuracy of intermediate models.

[0069] Furthermore, the numerical simulation dataset in this application includes sample motion data for various operating conditions, including normal, complex, and rare conditions. Under normal operating conditions, the wind and waves are stable and the turbulence intensity is low; under complex operating conditions, the turbulence intensity is moderate; and under rare operating conditions, such as typhoon conditions, the turbulence intensity is high. During the pre-training phase, sample motion data for normal, complex, and rare operating conditions are sequentially input into the initial model. The nonlinear dynamic intensity is increased at each stage, allowing the model to progressively adapt to extreme scenarios.

[0070] Optionally, in the above embodiments, during the process of fine-tuning the intermediate model to obtain the AI ​​agent model, historical motion data from the historical test dataset is input into the intermediate model to obtain historical results, and a fine-tuning function is constructed based on the historical results, the historical load corresponding to the historical motion data, and the loss function. Then, the electronic device adjusts the parameters of the intermediate model according to the fine-tuning function to obtain the AI ​​agent model.

[0071] For example, during the fine-tuning phase, the electronic device incorporates historical motion data from the historical test dataset and uses this historical motion data to construct a fine-tuning function, as shown in the following formula (4): Formula (4) where, Indicates the weighting coefficient. This indicates the historical output of the initial model when the input is historical motion data; This means that when the input to the initial model is historical motion data, the data obtained by the sensor from the motion response of the physical model, that is, the historical load corresponding to the historical motion data, is the actual measured data of the sensor. This represents the robust term of the measured torque, which is an L1 norm. The L1 norm can reduce the interference of sensor outliers.

[0072] By adopting this approach, the electronic device constructs a fine-tuning function based on the loss function, historical results, and historical loads, which can reduce the interference of sensor outliers and improve the accuracy of the AI ​​agent model.

[0073] In this application, an AI agent model is obtained through a pre-training stage and a fine-tuning stage. Furthermore, an electronic device performs precision quantization and channel pruning on the AI ​​agent model to compress the model volume and reduce latency.

[0074] Secondly, we conducted hybrid model experiments using AI agent models.

[0075] In this application, for offshore floating wind turbines requiring hybrid model testing, the electronic equipment pre-acquires a set of operating conditions, which includes multiple preset operating conditions, each corresponding to parameters such as wind speed, wave height, and wave spectrum. The electronic equipment loads a pre-trained AI agent model and initializes the data interface. The physical layer initiates the wave generation system to reproduce the marine environment, while the blade simulation device remains in standby mode.

[0076] The electronic device uses preset operating conditions to drive the physical model, causing the physical model to generate initial motion, and the sensors capture the motion data in real time. The interaction layer performs time-series alignment and interpolation processing on the motion data collected by the sensors to generate time-continuous standard data.

[0077] The AI ​​agent model in the intelligent layer calls standard data. Each time the AI ​​agent model calls standard data, the electronic device inputs the standard data at the corresponding time point into the AI ​​agent model so that the AI ​​agent model can output at least the predicted aerodynamic load. This process is the application stage of the AI ​​agent model. In this stage, the physical calculation layer of the AI ​​agent model generates physical residuals, thereby automatically correcting nonlinear effects such as dynamic stall. The aerodynamic load drives the blade simulation device, causing the blade simulation device and other components to make motion responses. Sensors collect motion data, which is interpolated into standard data and then provided to the AI ​​agent model, causing the AI ​​agent model to output aerodynamic loads, etc., continuously cycling until the preset operating condition ends.

[0078] Optionally, for each standard data point in the time-continuous standard data, the electronic device inputs the standard data into the encoding layer of the AI ​​proxy model to obtain a feature vector, and then inputs the feature vector into the output layer of the AI ​​proxy model so that the output layer outputs an initial load vector. Simultaneously, the electronic device inputs the standard data into the physical calculation layer of the AI ​​proxy model so that the physical calculation layer outputs a physical residual, and determines the aerodynamic load based on the physical residual and the initial load vector. The physical residual is used to compensate for deviations in the aerodynamic characteristics of the blades in the physical model under unsteady airflow conditions, as well as deviations in the aerodynamic load caused by force loop hysteresis. Force loop hysteresis refers to the deviation of the aerodynamic load throughout the entire chain from the input of the standard data to the generation of the motion response by the physical model.

[0079] Referring to Figure 4, the encoding layer outputs a feature vector. This feature vector is input to the fully connected layer of the output layer and simultaneously connected directly to the adder in the output layer. In this way, the adder sums the feature vector, the output of the regression layer, and the physical residual from the physical computation layer to obtain the initial load vector. The corrected aerodynamic load can then be obtained from this initial load vector. This is done to prevent gradient decay that may occur in the AI ​​proxy model, and to prevent the AI ​​proxy model from forgetting or diluting the key operating condition feature information extracted by the encoder. The feature vector is a high-order feature, integrating the input environmental and motion parameters. Motion parameters include, but are not limited to, tower top velocity, inflow wind speed, wave height, spectral peak period, and pitch angle, forming a deep encoding of the coupled dynamic state of the floating wind turbine system. Furthermore, the connection of the feature vector ensures that the abrupt changes in the large-scale motion of the floating platform under extreme conditions such as typhoons can be directly used for the final decision, without being lost due to network depth.

[0080] In this approach, during the hybrid model test, the physical computation layer of the AI ​​proxy model is used to calculate the physical residuals to compensate for the force loop hysteresis caused by dynamic stall, thereby improving the quality of the hybrid test model.

[0081] Optionally, in the above embodiments, during the hybrid model test, when encountering sudden events such as a surge in wind speed, the intelligent layer predicts the load change trend based on real-time motion data and generates anti-overturning control commands to adjust the blade simulation device. Simultaneously, an online learning mechanism is triggered to dynamically update model parameters to adapt to environmental changes.

[0082] After traversing all preset working conditions in the set of working conditions, the loading of the preset working conditions is terminated. The intelligent layer outputs a full-time motion parameter-load coupled dataset to obtain the experimental results. The experimental results include, but are not limited to, aerodynamic load curves, platform motion trajectories, mooring tension characteristics, etc., thus providing a basis for the design verification of offshore floating wind turbines.

[0083] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0084] Figure 5 is a schematic diagram of the hybrid model test device for offshore floating wind turbines provided in this application. The hybrid model test device 500 for offshore floating wind turbines includes: a drive module 51, a generation module 52, a processing module 53, and an update module 54.

[0085] The driving module 51 is used to drive a physical model for each preset working condition in the set of working conditions, so that the physical model generates initial motion. The physical model is a scaled-down test model of an offshore floating wind turbine. The generation module 52 is used to generate time-continuous standard data based on the motion data of the physical model that generates the initial motion. The time points of the time-continuous standard data correspond one-to-one with the times when the AI ​​agent model calls the motion data. The processing module 53 is used to input the standard data at the corresponding time point into the AI ​​agent model each time the AI ​​agent model calls the motion data, so that the AI ​​agent model outputs at least the predicted aerodynamic load. The driving module 51 is also used to drive the physical model according to the aerodynamic load, so that the physical model generates a motion response. The updating module 54 is used to update the time-continuous standard data according to the motion response until the preset working condition ends and the AI ​​agent model outputs the test results of the preset working condition.

[0086] In one feasible implementation, the processing module 53 is configured to: input the standard data into the encoding layer of the AI ​​proxy model for each standard data in the time-continuous standard data to obtain a feature vector; input the feature vector into the output layer of the AI ​​proxy model to output an initial load vector; input the standard data into the physical calculation layer of the AI ​​proxy model to output a physical residual, the physical residual being used to compensate for the deviation of the aerodynamic characteristics of the blades on the physical model from the ideal state under unsteady airflow, and the deviation of the aerodynamic load caused by force loop hysteresis, the force loop hysteresis being the deviation of the aerodynamic load in the entire link from the input of the standard data to the physical model generating the motion response; and determine the aerodynamic load based on the physical residual and the initial load vector.

[0087] In one feasible implementation, each time the AI ​​proxy model calls the motion data, the processing module 53 inputs standard data at the corresponding time point into the AI ​​proxy model, so that before the AI ​​proxy model outputs at least the predicted aerodynamic load, it is also used to generate multiple working conditions based on numerical simulation, the multiple working conditions including rare working conditions; extract sample motion data and sample load of each working condition from the multiple working conditions to obtain multiple sample pairs; use the multiple sample pairs to train an initial model to obtain an intermediate model; and fine-tune the intermediate model to obtain the AI ​​proxy model.

[0088] In one feasible implementation, when the processing module 53 trains an initial model using the multiple sample pairs to obtain an intermediate model, it is used to input the sample motion data in the sample pairs into the initial model to obtain a sample load vector; input the sample motion data in the sample pairs into the initial model so that the physical computing layer of the initial model determines the sample residuals; construct a loss function based on the sample load vector, the sample loads in the sample pairs, and the sample residuals; and adjust the parameters of the initial model according to the loss function to obtain the intermediate model.

[0089] In one feasible implementation, during the process of fine-tuning the intermediate model to obtain the AI ​​agent model, the processing module 53 inputs historical motion data from the historical test dataset into the intermediate model to obtain historical results; constructs a fine-tuning function based on the historical results, the historical load corresponding to the historical motion data, and the loss function; and adjusts the parameters of the intermediate model according to the fine-tuning function to obtain the AI ​​agent model.

[0090] In one feasible implementation, when the processing module 53 inputs the sample motion data from the sample pair into the initial model to obtain the sample load vector and sample residual, it is used to construct a motion correction equation based on the sample motion data; construct the residual function of the physical calculation layer based on the leaf element momentum theory and the motion correction equation; and determine the sample residual based on the residual function.

[0091] In one feasible implementation, after the processing module 53 generates multiple working conditions based on numerical simulation, it is further used to label rare working conditions and normal working conditions among the multiple working conditions to obtain a labeled dataset; train an adversarial network based on the labeled dataset; and use the trained adversarial network to generate a typhoon-abnormal wave joint working condition.

[0092] In one feasible implementation, when the AI ​​agent model outputs the predicted motion data of the physical model, the processing module 53 acquires the actual motion data of the physical model; compares the predicted motion data and the actual motion data to obtain a comparison result; and when the comparison result indicates that the error of the predicted motion data exceeds a preset error, optimizes the AI ​​agent model.

[0093] The hybrid model test device for offshore floating wind turbines provided in this application can perform the actions of the electronic devices in the above embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0094] Figure 6 is a schematic diagram of the structure of the electronic device provided in this application. The electronic device 600 includes: a processor 61 and a memory 62; the memory 62 stores computer instructions and test data; the processor 61 executes the computer instructions stored in the memory 62, causing the processor 61 to perform the hybrid model test method for offshore floating wind turbines as described above.

[0095] The specific implementation process of processor 61 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0096] Optionally, the electronic device 600 also includes a communication component 63. The processor 61, memory 62, and communication component 63 can be connected via a bus 64.

[0097] This application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, are used to implement the hybrid model test method for offshore floating wind turbines as described above.

[0098] This application also provides a computer program product comprising a computer program that, when executed by a processor, implements the hybrid model test method for offshore floating wind turbines as described above.

[0099] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0100] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A hybrid model test method for offshore floating wind turbines, characterized in that, include: For each preset working condition in the set of working conditions, the preset working condition drives the physical model to generate initial motion. The physical model is a scaled-down test model of an offshore floating wind turbine. Time-continuous standard data is generated based on the motion data of the physical model undergoing the initial motion. The time points of the time-continuous standard data correspond one-to-one with the times when the AI ​​agent model calls the motion data. Each time the AI ​​agent model calls the motion data, the standard data at the corresponding time point is input into the AI ​​agent model so that the AI ​​agent model outputs at least the predicted aerodynamic load. The physical model is driven according to the aerodynamic load to generate a motion response. The time-continuous standard data is updated according to the motion response until the preset working condition ends and the AI ​​agent model outputs the test results of the preset working condition.

2. The method according to claim 1, characterized in that, The step of inputting standard data at the corresponding time point into the AI ​​proxy model each time the motion data is invoked, so that the AI ​​proxy model outputs at least the predicted aerodynamic load, includes: for each standard data in the time-continuous standard data, inputting the standard data into the encoding layer of the AI ​​proxy model to obtain a feature vector; inputting the feature vector into the output layer of the AI ​​proxy model so that the output layer outputs an initial load vector; inputting the standard data into the physical calculation layer of the AI ​​proxy model so that the physical calculation layer outputs a physical residual, the physical residual being used to compensate for the deviation of the aerodynamic characteristics of the blades on the physical model from the ideal state under unsteady airflow, and the deviation of the aerodynamic load caused by force loop hysteresis, the force loop hysteresis being the deviation of the aerodynamic load in the entire link from the input of the standard data to the physical model generating the motion response; and determining the aerodynamic load based on the physical residual and the initial load vector.

3. The method according to claim 1, characterized in that, Before the AI ​​agent model inputs standard data at the corresponding time point into the AI ​​agent model each time it calls the motion data, so that the AI ​​agent model outputs at least the predicted aerodynamic load, the method further includes: generating multiple working conditions based on numerical simulation, the multiple working conditions including rare working conditions; extracting sample motion data and sample load for each working condition from the multiple working conditions to obtain multiple sample pairs; training an initial model using the multiple sample pairs to obtain an intermediate model; and fine-tuning the intermediate model to obtain the AI ​​agent model.

4. The method according to claim 3, characterized in that, The step of training an initial model using the multiple sample pairs to obtain an intermediate model includes: inputting sample motion data from the sample pairs into the initial model to obtain a sample load vector; inputting sample motion data from the sample pairs into the initial model so that the physical computation layer of the initial model determines the sample residuals; constructing a loss function based on the sample load vector, the sample loads in the sample pairs, and the sample residuals; and adjusting the parameters of the initial model based on the loss function to obtain the intermediate model.

5. The method according to claim 4, characterized in that, The step of fine-tuning the intermediate model to obtain the AI ​​agent model includes: inputting historical motion data from the historical test dataset into the intermediate model to obtain historical results; constructing a fine-tuning function based on the historical results, the historical load corresponding to the historical motion data, and the loss function; and adjusting the parameters of the intermediate model according to the fine-tuning function to obtain the AI ​​agent model.

6. The method according to claim 4, characterized in that, The step of inputting the sample motion data from the sample pair into the initial model to obtain the sample load vector and sample residual includes: constructing a motion correction equation based on the sample motion data; constructing the residual function of the physical computing layer based on leaf element momentum theory and the motion correction equation; and determining the sample residual based on the residual function.

7. The method according to claim 4, characterized in that, After generating multiple operating conditions based on numerical simulation, the process further includes: labeling rare and normal operating conditions among the multiple operating conditions to obtain a labeled dataset; training an adversarial network based on the labeled dataset; and using the trained adversarial network to generate a typhoon-abnormal wave joint operating condition.

8. The method according to any one of claims 1 to 3, characterized in that, Also includes: When the AI ​​agent model outputs the predicted motion data of the physical model, the actual motion data of the physical model is obtained; the predicted motion data and the actual motion data are compared to obtain a comparison result; when the comparison result indicates that the error of the predicted motion data exceeds a preset error, the AI ​​agent model is optimized.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it causes the electronic device to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images