A wind turbine blade icing prediction method based on deep learning

Abstract

This invention relates to the field of wind power generation technology and discloses a deep learning-based method for predicting icing on wind turbine blades. The method includes acquiring two-dimensional icing morphology data of wind turbine blades under different operating conditions, along with corresponding multi-dimensional operating parameters such as ambient temperature and relative humidity; performing normalization and rasterization preprocessing on the data; constructing a conditional generative adversarial network (GAN) containing a generator and a discriminator; using the operating parameters as constraints, alternately updating the network parameters through adversarial training until Nash equilibrium is reached; and using the trained generator to directly output a two-dimensional icing morphology matrix based on the input operating conditions, followed by post-processing to reconstruct the physical icing curve. This invention establishes a nonlinear mapping between operating parameters and icing morphology by introducing physical constraints such as ambient relative humidity, ambient temperature, incoming wind speed, and liquid water content, achieving efficient end-to-end prediction, improving the physical consistency of the prediction results, and meeting real-time monitoring requirements.

Description

Technical Field

[0001] This invention relates to the field of wind power operation and maintenance and intelligent prediction technology, specifically a deep learning-based method for predicting icing on wind turbine blades. Background Technology

[0002] Wind power generation, as an important component of clean energy, is increasingly widely used in cold and high-altitude regions. However, wind turbine blades are prone to icing in low-temperature and high-humidity environments. Icing alters the aerodynamic shape of the blade airfoil, leading to decreased lift, increased drag, and consequently reduced power generation efficiency. Furthermore, uneven icing distribution can cause impeller mass imbalance, increasing fatigue load on the turbine and, in severe cases, even causing blade breakage or tower collapse. Therefore, accurately predicting blade icing morphology under different meteorological conditions is crucial for assessing turbine performance degradation, developing appropriate anti-icing and de-icing strategies, and ensuring safe turbine operation.

[0003] Currently, wind turbine blade icing prediction mainly relies on icing wind tunnel experiments and numerical simulation methods. While icing wind tunnel experiments can provide relatively realistic icing data, they are limited by high experimental costs, site conditions, and model scaling effects, making it difficult to cover the complex and variable meteorological conditions encountered in actual wind turbine operation. Numerical simulation methods based on computational fluid dynamics (CFD) calculate ice shape by solving airflow fields, water droplet impact characteristics, and thermodynamic phase transition equations. Although these methods offer high accuracy, their calculation process involves coupled iterations of multiple physics fields, resulting in high computational complexity and long simulation times, making it difficult to meet the engineering requirements for real-time online monitoring and rapid assessment of icing conditions in wind turbine generators.

[0004] With the development of artificial intelligence technology, data-driven icing prediction methods have gradually become a research hotspot. Existing data-driven methods typically utilize algorithms such as support vector machines and artificial neural networks to establish a mapping relationship between meteorological parameters and icing characteristics. However, existing research often focuses on predicting single indicators such as maximum icing thickness and icing area, or only performs binary classification judgments on the presence or absence of icing. This fails to intuitively and completely reconstruct the two-dimensional icing geometry of the blade leading edge, making it difficult to support refined aerodynamic performance analysis. Although some studies have attempted to predict icing patterns using image generation techniques, the lack of effective constraints on physical boundary conditions during model construction often results in generated icing morphologies with blurred edges, missing texture details, or inconsistent physical evolution. Furthermore, existing prediction models often neglect the crucial variable of relative humidity in their input feature construction, typically assuming a saturated environment and failing to fully consider the impact of water evaporation or sublimation endothermic effects on the blade surface thermal balance under unsaturated humidity conditions. This leads to insufficient generalization ability of the model under complex meteorological conditions and poor physical consistency of the prediction results. Summary of the Invention

[0005] To address the problem of insufficient generalization ability and poor physical consistency of prediction results under complex meteorological conditions due to the failure to fully consider the impact of water evaporation or sublimation endothermic effects on the blade surface heat balance under unsaturated humidity conditions, this invention achieves this through the following technical solution: A deep learning-based method for predicting icing on wind turbine blades, comprising the following steps: Step S1: Obtain two-dimensional icing morphology data and corresponding operating parameters of wind turbine blades under different operating conditions; the two-dimensional icing morphology data characterizes the icing profile or icing thickness distribution on the airfoil section of the blade. Step S2: Normalize the working condition parameters and map the two-dimensional icing morphology data into a two-dimensional matrix with a preset size to construct training sample pairs; Step S3: Construct a generative adversarial network model, which includes a generator network and a discriminator network that are in conflict with each other; Step S4: Input the processed real two-dimensional icing morphology data and the corresponding real working condition parameters into the generative adversarial network model; according to the preset loss function, alternately update the parameters of the generator network and the discriminator network until the generative adversarial network model reaches Nash equilibrium. Step S5: Receive the operating condition parameters to be predicted and input them into the trained generator network. The generator network outputs the two-dimensional icing morphology prediction results of the wind turbine blades under the corresponding operating conditions.

[0006] Furthermore, in step S1, the operating parameters constitute a multi-dimensional vector, the dimensions of which include at least ambient temperature, incoming wind speed, ambient relative humidity, liquid water content, and median volume diameter of water droplets; in the multi-dimensional vector, the ambient relative humidity is used as an independent input variable to characterize the influence of water vapor content in the air on the icing process.

[0007] In step S2, the normalization process uses the maximum and minimum values ​​obtained from the training dataset to map the operating parameters to a preset numerical range; the processing of the two-dimensional icing morphology data includes coordinate alignment, size unification and rasterization operations to convert the continuous airfoil leading edge icing profile into the two-dimensional matrix.

[0008] Further, in step S3, the generator network and the discriminator network are configured as a conditional generative adversarial network topology. The generator network is configured to receive the normalized operating parameters as conditional inputs and map the input operating parameters to predicted icing morphology samples.

[0009] The discriminator network is configured to receive a two-dimensional icing morphology matrix as input, the input including real icing morphology samples and predicted icing morphology samples; the discriminator network also receives normalized operating condition parameters as constraints, and outputs a probability scalar representing that the input sample is real data.

[0010] Furthermore, the generator network includes a parameter embedding module, a feature expansion module, and a shape generation module connected in sequence.

[0011] The parameter embedding module uses a fully connected layer to extract features from the input operating parameters, mapping low-dimensional physical parameters into high-dimensional feature vectors.

[0012] The feature expansion module uses a fully connected layer to expand the dimension of the high-dimensional feature vector and transforms it into a seed feature map with a three-dimensional tensor structure through a reshaping operation.

[0013] The morphology generation module consists of multiple cascaded transposed convolutional layers. The seed feature map is upsampled through transposed convolution operations to gradually improve the spatial resolution of the feature map and reduce the number of channels. Finally, the predicted icing morphology sample is generated through the output layer.

[0014] Furthermore, the discriminator network includes feature fusion operations and multiple convolutional modules stacked in series.

[0015] In the input layer or feature extraction stage, the discriminator network expands the spatial dimension of the input working condition parameters to form a feature mask with the same spatial resolution as the input two-dimensional icing morphology matrix. The feature mask and the two-dimensional icing morphology matrix are then concatenated in the channel dimension to form a composite input tensor.

[0016] The convolution module includes convolutional layers, normalization layers, and nonlinear activation layers. It extracts the geometric texture and spatial distribution features of the input through convolution operations, and outputs the probability scalar through fully connected layers and activation functions.

[0017] Furthermore, in step S4, the loss function includes a discriminator loss function and a generator loss function, and the alternating update process of the parameters solves the minimax game problem.

[0018] The discriminator loss function is constructed based on the log-expected value of the probability that a real sample is judged as real and the log-expected value of the probability that a generated sample is judged as fake; the optimization objective is to maximize the accuracy of the discriminator network in distinguishing between real samples and generated samples.

[0019] The generator loss function is constructed based on the log-expectation of the probability that a generated sample is judged as real; the optimization objective is to maximize the probability that the discriminator network judges the generated sample as real.

[0020] Furthermore, in step S4, the alternating update of the parameters of the generator network and the discriminator network includes cyclically performing the following steps: Freeze the weight parameters of the generator network, and use the generator network to generate fake icing morphology samples based on the operating parameters of the current batch; The discriminator network is input with real sample pairs and generated sample pairs, the discriminator loss function is calculated, and the parameters of the discriminator network are updated using an adaptive moment estimation optimizer. Freeze the weight parameters of the discriminator network, use the generator network to generate fake icing morphology samples and input them into the discriminator network; The generator loss function is calculated and the parameters of the generator network are updated using an adaptive moment estimation optimizer.

[0021] Further, in step S5, the prediction includes: acquiring the target working condition data to be predicted, and using the statistical parameters stored in the training phase to perform consistency normalization processing on the target working condition data to obtain a normalized vector. The normalized vector is fed as a conditional input to the trained generator network, which performs forward inference computation and directly outputs a normalized two-dimensional gray matrix as the prediction result.

[0022] Furthermore, after step S5, a post-processing step is included for the prediction result. The post-processing includes: using the maximum ice thickness parameter recorded during the training phase to perform an inverse denormalization transformation on the two-dimensional gray matrix, restoring the dimensionless gray values ​​to physical ice thickness values, and generating a physical thickness distribution matrix. The physical thickness distribution matrix is ​​truncated by applying a preset thickness determination threshold, and elements smaller than the thickness determination threshold are set to zero to eliminate background noise and distinguish between icing areas and air areas.

[0023] Furthermore, the post-processing also includes: mapping the non-zero elements in the truncated physical thickness distribution matrix back to the airfoil coordinate system according to the preset grid physical size resolution and the positioning coordinates of the airfoil leading edge stationary point; By connecting discrete outer edge points of the ice shape using a curve fitting algorithm, a continuous and closed icing airfoil geometry curve is reconstructed. Based on the reconstructed icing airfoil geometry curve, the maximum icing thickness and icing cross-sectional area are calculated to quantitatively assess the severity of icing.

[0024] This invention provides a deep learning-based method for predicting icing on wind turbine blades. It offers the following advantages: 1. This invention constructs a conditional generative adversarial network model, which uses relative humidity, ambient temperature, incoming wind speed, liquid water content, and median volume diameter of water droplets to form a multidimensional working condition parameter vector, and uses it as a strong constraint condition for the generator and discriminator. This setting enables the model to capture the nonlinear influence of air water vapor content on the icing thermodynamic process, and solves the problem of large prediction deviation in traditional methods under unsaturated humidity conditions, thereby maintaining the consistency between prediction results and physical laws under various meteorological combinations.

[0025] 2. This invention utilizes a trained generator network to perform forward inference, achieving end-to-end prediction from low-dimensional operating condition numerical values ​​to a two-dimensional icing morphology matrix. Compared to traditional computational fluid dynamics numerical simulation methods that require iterative solutions to the flow field and droplet trajectory, this invention avoids heavy computational loads and can generate prediction results at millisecond speeds, thereby reducing system computational latency and meeting the needs of wind turbine generators for real-time monitoring and rapid response to icing conditions during operation.

[0026] 3. This invention combines a post-processing mechanism including denormalization, threshold truncation, and geometric reconstruction to transform the abstract grayscale matrix output by the neural network into an icing airfoil curve with clear physical coordinates. This processing flow not only eliminates the background noise interference inherent in the generated model, but also directly quantifies and calculates engineering indicators such as maximum icing thickness and icing cross-sectional area, providing directly callable quantitative data support for the wind turbine main control system to judge the severity of icing and trigger anti-icing and de-icing strategies. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the generative adversarial network of the present invention; Figure 3 This is a flowchart of the training process for the generative adversarial network model of the present invention.

[0028] The module includes: 101. Data acquisition module; 102. Data preprocessing module; 103. Generative adversarial network construction module; 104. Model training module; and 105. Two-dimensional icing morphology prediction module. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] See attached document Figure 1This invention provides a deep learning-based method for predicting icing on wind turbine blades. The method operates on an electronic computing device, which includes a memory, a processor, and a graphics processing unit. The memory stores computer-executable instructions, and the processor executes these instructions to implement the wind turbine blade icing prediction method. The method establishes a mapping relationship between multi-dimensional operating parameters and two-dimensional icing morphology by constructing and training a generative adversarial network, thereby predicting the icing morphology of the blades.

[0031] The deep learning-based wind turbine blade icing prediction method is mainly executed collaboratively by a data acquisition module 101, a data preprocessing module 102, a generative adversarial network construction module 103, a model training module 104, and a two-dimensional icing morphology prediction module 105.

[0032] First, the data acquisition module 101 executes step S1, which involves acquiring two-dimensional icing morphology data of the wind turbine blades under different operating conditions and the corresponding operating parameters. The operating parameters constitute a multi-dimensional vector, with dimensions including at least ambient temperature, incoming wind speed, ambient humidity, liquid water content, and median droplet volume diameter. Ambient humidity, as an independent input variable, characterizes the physical influence of air vapor content on the icing phase change process and the final morphology. The two-dimensional icing morphology data consists of icing profile data or icing thickness distribution data on the blade airfoil cross-section.

[0033] The data preprocessing module 102 is connected to the data acquisition module 101 and is used to execute step S2. The data preprocessing module 102 normalizes the operating parameters, mapping parameters with different physical dimensions to a unified numerical range. Simultaneously, the data preprocessing module 102 performs coordinate alignment, size unification, and rasterization on the two-dimensional icing morphology data. Through rasterization, the continuous airfoil leading-edge icing profile is converted into a two-dimensional matrix or grayscale image with a fixed size, thereby constructing training sample pairs containing the operating parameter vector and the icing morphology matrix.

[0034] The Generative Adversarial Network (GAN) construction module 103 is used to execute step S3. The GAN construction module 103 constructs a GAN model in the memory of the computing device. This model includes two sub-networks: a generator network and a discriminator network. The generator network is configured to receive a normalized vector of operating conditions as input and output the predicted two-dimensional icing morphology under the corresponding operating conditions. The discriminator network is configured to receive two-dimensional icing morphology data as input, distinguish whether the input two-dimensional icing morphology data originates from a real dataset or is generated by the generator network, and output the corresponding discrimination probability value.

[0035] The model training module 104 is used to execute step S4. During the training phase, the model training module 104 inputs real two-dimensional icing morphology data and corresponding real-world operating parameters into the generative adversarial network (GAN). The model training module 104 updates the parameters of the generator network and the discriminator network alternately according to a preset loss function. During the parameter update process, the generator network learns the coupling relationship between operating parameters such as environmental humidity and liquid water content and icing morphology features, so that the generated two-dimensional icing morphology gradually approximates the real icing morphology in terms of spatial distribution characteristics and geometric shape, until the network reaches Nash equilibrium.

[0036] The two-dimensional icing morphology prediction module 105 is used to execute step S5. After the generative adversarial network model is trained, the two-dimensional icing morphology prediction module 105 receives the operating condition parameters to be predicted and inputs them into the trained generator network. Based on the learned mapping relationship, the generator network directly outputs the two-dimensional icing morphology prediction result of the wind turbine blade under the corresponding operating condition. This prediction result represents the icing thickness distribution at each coordinate point on the airfoil leading edge under specific ambient temperature, wind speed, and humidity conditions in the form of a two-dimensional matrix or image.

[0037] By executing computer program instructions in a processor, end-to-end prediction from environmental condition parameter input to two-dimensional icing morphology output is achieved. The prediction method specifically includes the following steps: Step S1: Acquisition of 2D icing morphology data and operating parameters. Two-dimensional icing morphology data and corresponding operating parameter data of the wind turbine blades under various historical operating conditions are acquired. The operating parameter data is used to construct the input operating parameter vector. This vector contains physical quantities in at least five dimensions, represented as: ; in, For ambient temperature, For the incoming wind speed, For ambient relative humidity, Liquid water content, The median volume diameter of the water droplet. Represents the matrix transpose symbol. Ambient relative humidity. The feature dimension configured as an independent input is used to quantify the nonlinear influence of air moisture content on icing rate and ice shape growth in subsequent model training. The two-dimensional icing morphology data are obtained from wind tunnel experimental measurement data or numerical simulation data based on computational fluid dynamics (CFD), which characterize the ice layer geometry of the blade airfoil leading edge under corresponding operating conditions.

[0038] Step S2, Data Preprocessing. Data preprocessing includes two sub-processes: normalization of operating parameters and rasterization of icing morphology. For the operating parameters, to eliminate the influence of differences in physical dimensions and numerical ranges on the convergence of the neural network gradient descent, the maximum and minimum normalization methods are used to normalize the operating parameter vector. Each component in the mapping is mapped to The numerical range of the operating condition parameter vector. The first in Parameters Normalized values The calculation formula is as follows: ; in, and These represent the maximum and minimum values ​​of the operating condition parameter in the historical dataset, respectively. The normalized operating condition parameter vector is denoted as... For two-dimensional icing morphology data, coordinate alignment, size unification, and rasterization operations are performed. The original icing contour coordinates or unstructured mesh data are mapped to a fixed Cartesian mesh system, generating a mesh with a size of [missing information]. Two-dimensional matrix : ; Elements in the matrix Corresponding to grid nodes The icing state value or normalized thickness value at that location. Represents the set of real numbers. The meaning is to belong to, Represents the dimension of the matrix. The processed matrix. This is a sample of the actual morphology of ice formation. .

[0039] Step S3, Construction of the Generative Adversarial Network (GAN) Model. Construct a GAN model comprising a generator network and a discriminator network. Generator Network Configured to learn conditional probability distribution Its input terminal receives a normalized chemical condition parameter vector. After internal parameter embedding, feature expansion, and deconvolution operations, the output is a predicted icing morphology matrix. : ; Discriminator Network Configured as a binary classifier, its input receives a two-dimensional icing morphology matrix. After multi-layer convolutional feature extraction, it outputs the probability value that the input sample belongs to the real data. : ; in, The mapping function representing the discriminator network has an output value approaching 1, indicating that the discriminator determines the input to be real data, and approaching 0, indicating that it determines it to be generated data. For the specific convolutional layers, activation functions, and connection methods within the generator and discriminator networks, those skilled in the art can construct them according to general deep neural network construction principles, or adopt the specific structures detailed in subsequent embodiments of this invention.

[0040] Step S4, Model Training. The generator and discriminator networks are trained adversarially using an alternating iterative approach until they reach Nash equilibrium. The training process aims to optimize the generator parameters to minimize the difference between the generated and real distributions, while simultaneously optimizing the discriminator parameters to maximize its ability to distinguish between real and generated samples. The loss function of the discriminator is defined. Its mathematical expression is: ; in, Represents the mathematical expectation. Represents the distribution of real data. This indicates the distribution of operating parameters. This represents a sample of the actual morphology of icing. It takes the form of cross-entropy loss. During the discriminator training step, the generator parameters are fixed, and the discriminator parameters are updated using gradient descent to minimize... Define the loss function for the generator. Its mathematical expression is: ; In the generator training step, the discriminator parameters are fixed, and the generator parameters are updated using the gradient descent algorithm to minimize the discriminator parameters. Through the aforementioned alternating update process, the generator network gradually learns the relative humidity of the environment. The nonlinear mapping law between other operating parameters and two-dimensional icing morphology.

[0041] Step S5, Two-dimensional icing morphology prediction. After model training, the weight parameters of the solidification generator network are determined. For a given operating condition to be predicted, the corresponding ambient temperature, incoming wind speed, ambient humidity, liquid water content, and median droplet volume diameter are obtained, and processed according to the normalization parameters in step S2 to obtain the input vector of the operating condition to be predicted. .Will Input the trained generator network and perform the forward propagation operation: ; Output This is the predicted result of the two-dimensional icing morphology of the wind turbine blades under this operating condition. The predicted result can be further processed by denormalization and coordinate transformation to restore the icing thickness distribution curve or image in physical space, which can be used for subsequent aerodynamic performance evaluation or de-icing control strategy formulation.

[0042] The operating condition parameter system is not merely a collection of numerical values, but rather serves as boundary conditions describing the physical process of icing on wind turbine blades, directly determining the input dimension and physical constraint capability of the prediction model.

[0043] Construct input operating condition parameter vector This vector is configured as a conditional input to the generator network, controlling the geometric features of the generated two-dimensional icing morphology. (Operating condition parameter vector) It contains five components, and its mathematical expression is as follows: ; The components correspond to the following physical meanings and technical characteristics: This indicates the ambient temperature, usually expressed in degrees Celsius (°C) or Kelvin (K). This parameter determines the thermodynamic state of the airflow field and whether supercooled water droplets possess the thermodynamic conditions for freezing, directly affecting the ice formation rate and ice type (such as clear ice, frost ice, or mixed ice).

[0044] The incoming airflow velocity is typically expressed in meters per second (m / s). The incoming airflow velocity determines the aerodynamic heating effect of the airflow on the blade surface and the momentum of water droplets impacting the blade surface. Higher airflow velocities alter the impact limit of water droplets and the local water collection coefficient, thus changing the extent of icing coverage on the airfoil surface.

[0045] The ambient relative humidity is typically expressed as a percentage (%). In this embodiment, ambient humidity is configured as an independent model input variable, rather than being passively assumed to be saturated or negligible. The ambient humidity parameter is directly related to the mass transfer processes (such as evaporation or sublimation) between the air and the water film on the blade surface, as well as the accompanying latent heat exchange. In unsaturated humidity environments, evaporative cooling alters the thermal equilibrium state of the blade surface, thus affecting the freezing coefficient and the final ice thickness. By... Explicitly include input operating condition parameter vector The model is able to learn and characterize the nonlinear coupling effect of humidity on the icing process.

[0046] This parameter represents the liquid water content, typically expressed in grams per cubic meter (g / m³). It characterizes the mass of liquid water contained in a unit volume of air and directly determines the total mass of water impacting the blade surface per unit time. It is a primary mass source parameter affecting the icing growth rate and icing thickness.

[0047] The median volume diameter of a water droplet is typically expressed in micrometers. This parameter characterizes the size distribution of supercooled water droplets in the air. The droplet diameter determines its inertial characteristics, which in turn affects its trajectory in the flow field and its impact point on the airfoil surface. Different droplet diameters result in different local water collection coefficient distributions, thus affecting the chordal distribution range of the icing morphology.

[0048] The five physical quantities mentioned above together constitute a complete operating condition space describing the icing phenomenon of wind turbine blades. During the data construction phase, for each discrete time point or experimental batch, the values ​​of these five parameters are simultaneously collected and encapsulated into an operating condition parameter vector. Operating condition parameter vector In subsequent processing, the samples are mapped one-to-one with the corresponding two-dimensional icing morphology data to form sample pairs for supervised learning. For the measurement and acquisition of operating parameters, those skilled in the art can use standard measuring instruments from meteorological station sensors, airborne measuring equipment, or wind tunnel testing systems. These are conventional technical methods in the field, and specific hardware measurement circuits or sensor models will not be described in detail here.

[0049] In this embodiment of the invention, two-dimensional icing morphology data is defined as a digital structure characterizing the geometric profile or thickness distribution of ice layer in the airfoil section of a wind turbine blade, particularly in the leading edge region. This data serves as the ground truth input to the discriminator in a generative adversarial network and the optimization objective of the generator. Its geometric features directly reflect the icing physics results under the coupled effects of specific environmental parameters such as temperature, wind speed, humidity, and liquid water content.

[0050] To adapt to the processing requirements of convolutional neural networks for structured data, the continuous distribution of airfoil icing morphology in physical space is discretized and mapped into a two-dimensional matrix. Two-dimensional matrix The mathematical definition of is as follows: ; in, Indicates the number of rows (height) of the matrix. Indicates the number of columns (width) of the matrix. Each element in the matrix... Corresponding to the local region of the airfoil leading edge at the discretized mesh nodes The freezing state attribute at the location. In one specific implementation, a two-dimensional matrix. Constructed as a single-channel grayscale image matrix, elements The numerical value represents the normalized ice thickness or ice density at that spatial location, and the numerical range is typically [value missing]. or The value 0 represents an ice-free area, while non-zero values ​​represent the presence of ice and its relative thickness. In another specific implementation, a two-dimensional matrix... Constructed as a binary geometric mask, elements Only 0 or 1 is used to precisely define the geometric outer contour boundary of the ice shape, thereby distinguishing the ice layer area from the background air area.

[0051] The main methods for obtaining two-dimensional icing morphology data include numerical simulation calculations based on physical mechanisms and experimental measurements based on controlled environments.

[0052] For data acquisition through numerical simulation, a computational fluid dynamics (CFD) solver is used to calculate the airflow field around the wind turbine airfoil, obtaining the velocity and pressure distribution. The trajectory and impact limit of supercooled water droplets in the flow field are calculated based on the Lagrange or Eulerian method. Furthermore, the mass and energy conservation equations are coupled to calculate the phase change freezing process of water droplets on the airfoil surface, thus generating simulated ice layer geometry data over time. The sequence of ice shape coordinates generated by this simulation process constitutes the original two-dimensional icing morphology data.

[0053] For data acquisition through experimental measurements, the airfoil model of the blade under test is fixed in the icing wind tunnel test section, and the set icing meteorological conditions are reproduced by controlling the wind tunnel spray system and refrigeration system. After the icing experiment, the actual ice accumulation profile of the airfoil leading edge is recorded using a 3D laser scanner, high-resolution camera photogrammetry, or ice shape depiction method. The collected physical ice shape contour coordinates are digitally extracted to obtain real two-dimensional icing morphology data.

[0054] When constructing the training dataset, each two-dimensional icing morphology sample obtained above is associated with its corresponding working condition parameter vector under the same generation conditions. A unique and rigorous index mapping is established. This ensures that the generative adversarial network can learn the causal logic between a given input condition and the output shape, thereby accurately reconstructing the corresponding two-dimensional ice distribution based on the input condition parameters during the prediction phase. For the raw, irregular ice shape coordinate data, subsequent data preprocessing steps are required to transform it into the aforementioned standard matrix form.

[0055] To eliminate the impact of differences in the dimensions and numerical magnitudes of different physical parameters on the training stability of generative adversarial network models, and to construct a standardized data structure suitable for convolutional neural network operations, this embodiment of the invention performs standardization and preprocessing operations on the originally acquired dataset. This process mainly includes normalization of operating parameters and geometric correction and rasterization of two-dimensional icing morphology data.

[0056] Operating parameter vector at the input end Because ambient temperature, incoming wind speed, ambient humidity, liquid water content, and median droplet diameter have completely different physical dimensions and vastly different dynamic ranges (e.g., droplet diameter is on the order of micrometers, while wind speed is on the order of meters per second), directly inputting these parameters into the network can lead to difficulties in loss function convergence or unstable gradient updates. Therefore, a max-min normalization strategy is employed to linearly map the physical parameters of each dimension to... The dimensionless interval.

[0057] For the operating condition parameter vector The first in physical components Its normalized value Calculated using the following formula: ; in, This represents the maximum value statistically obtained for this physical parameter in the training dataset. This represents the minimum value statistically obtained for this physical parameter in the training dataset. Through this transformation, a normalized vector of operating condition parameters is constructed. This ensures that all input features are numerically of the same order of magnitude, thereby balancing the weight contributions of various physical conditions in the feature extraction process.

[0058] For the two-dimensional icing morphology data at the output end, the raw data is usually a set of discrete coordinate points along the airfoil surface. It may exist in the form of unstructured mesh nodes. In order to transform it into a tensor format that can be processed by generator and discriminator networks, spatial geometric transformation and discretization mapping need to be performed.

[0059] Perform coordinate system alignment. Establish a unified local Cartesian coordinate system and translate the stationary point or geometric leading edge point of the airfoil to the origin. The airfoil chord direction is aligned with the horizontal axis. By using rotation and translation transformation matrices, the reference geometric position of the airfoil for all samples is unified, ensuring that the neural network only focuses on the morphological changes caused by icing, rather than the spatial pose differences of the airfoil itself.

[0060] Perform sizing and rasterization. Define a fixed physical extent covering the critical icing region of the airfoil's leading edge, and divide this extent into... A regular grid is used. The original irregular ice-shaped contour data is mapped onto this regular grid using linear interpolation or spline interpolation algorithms. For each pixel in the grid... Calculate its corresponding physical spatial location and assign it the icing characteristic value at that location.

[0061] In a specific implementation, the icing characteristic values ​​are assigned using a thickness mapping method. The ice thickness perpendicular to the airfoil surface at each grid node is calculated. To accommodate the activation function range of the network output layer (typically Tanh or Sigmoid), this thickness value is also normalized to generate the final image matrix elements. : ; in, This represents the threshold for the maximum ice thickness observed in the dataset. After the above processing, the original physical field data is transformed into a normalized grayscale image matrix, where the pixel grayscale value directly represents the relative ice thickness at that location. The normalized grayscale image matrix and the corresponding normalized conditioned vector... Together, they form standard training sample pairs, which are used for supervised training of the generative adversarial network. For image data augmentation, small random noise or small-amplitude grid translation perturbations can be introduced during training to improve the model's robustness to measurement errors, while maintaining the physical topology.

[0062] See attached document Figure 2 In this embodiment of the invention, the constructed wind turbine blade icing prediction model adopts a conditional generative adversarial network (GAN) topology. The overall architecture of this model mainly consists of two adversarial deep neural networks that are synchronously updated through backpropagation: a generator network (denoted as...). ) and discriminator network (denoted as The generator network and discriminator network are configured to engage in a zero-sum game within a multidimensional feature space to achieve a nonlinear mapping from a low-dimensional operating parameter space to a high-dimensional icing morphology data space.

[0063] Generator Networks It acts as a reconstruction engine for icing morphology, and its function is to learn the conditional probability distribution of icing morphology under given operating conditions. The generator network receives a preprocessed and normalized vector of operating parameters at its input. In specific implementations, to increase the diversity of generated results or prevent pattern collapse, the input of the generator network can also be coupled in parallel with a noise vector sampled from a standard normal or uniform distribution. The generator network, through its internal neural network layers, maps the input vector into a two-dimensional matrix with the same size as the actual icing data, thus generating the predicted icing morphology. The mathematical expression of this process is as follows: ; in, It is a five-dimensional normalized vector that includes ambient temperature, incoming wind speed, ambient humidity, liquid water content, and median volume diameter of water droplets; For noise variables in the potential space; This represents the nonlinear mapping function defined by the generator network.

[0064] Discriminator Network The discriminator network acts as a discrimination engine for icing morphology, distinguishing whether the input two-dimensional icing morphology data comes from a real physical observation dataset or is fabricated data synthesized by the generator network. The input to the discriminator network is configured to receive a two-dimensional icing morphology matrix. This input matrix can be either a real icing morphology sample provided by the data acquisition module 101 or... It can also be the predicted icing morphology sample output by the generator network. The discriminator network extracts the geometric texture and spatial distribution features of the input image and outputs a normalized probability scalar. This is used to characterize the confidence level that the input sample is real data. The mathematical expression of this process is as follows: ; in, This represents the nonlinear mapping function defined by the discriminator network. When the output value... When the value approaches 1, it indicates that the discriminator determines the input sample to be a true icing morphology; when the output value... When the value approaches 0, it indicates that the discriminator determines that the input sample is a fake sample generated by the generator.

[0065] In the overall architecture of this embodiment, the generator network With discriminator network There is a clear interaction between data flow and gradient flow. During the forward propagation phase, the normalized operating condition parameter vector... Input generator gets Subsequently Compared with real samples The discriminator is fed into the data to obtain the discrimination results. During the backpropagation phase, an adversarial loss function is constructed to backpropagate the discriminator's error gradient to update the generator's weight parameters.

[0066] Unlike traditional unsupervised generative adversarial networks, the architecture in this embodiment explicitly introduces a normalized condition parameter vector. This constraint not only applies to the generator network to guide ice shape generation, but is also fed into the discriminator network during the training phase. This ensures that the discriminator not only judges the realism of the shape, but also its consistency with the operating conditions. This constraint mechanism guarantees the accuracy of the model's final ice formation prediction. Strictly controlled by input environmental humidity and other meteorological parameters, it achieves physically consistent targeted forecasts. The entire architecture is built on a differentiable deep neural network, supporting end-to-end gradient descent training without the need for manually designed complex feature extraction operators.

[0067] In this embodiment of the invention, the generator network serves as the core inference unit for predicting icing morphology, and its internal topology is designed as a deconvolutional neural network architecture or a deep decoder architecture. This network aims to deconstruct and reconstruct low-dimensional numerical vectors containing information such as ambient temperature, wind speed, and humidity layer by layer, ultimately mapping them into a two-dimensional matrix with rich spatial geometric details through nonlinear transformation. The specific processing flow and internal structural components of the generator network sequentially include a parameter embedding module, a feature expansion module, and a morphology generation module.

[0068] The parameter embedding module is configured as the entry layer of the generator network to receive normalized operating condition parameter vectors. In an embodiment that introduces random perturbations to increase generation diversity, the module first performs a vector concatenation operation, combining the normalized operating condition parameter vectors... With random noise vectors sampled from the latent space (The vectors typically follow a standard normal distribution N(0,1)) and their dimensions are concatenated to form a combined input vector. Subsequently, at least one fully connected layer is used to extract features from the input vector, mapping it from a low-dimensional physical parameter space to a high-dimensional latent feature space. The mathematical expression for this process is as follows: ; in, This represents the weight matrix of the parameter embedding layer. This represents the bias vector. This represents a non-linear activation function (preferably ReLU or Leaky ReLU). The output is a high-dimensional feature vector, which contains complex nonlinear coupling characteristics between physical parameters such as environmental humidity and liquid water content, such as the potential impact of the synergistic effect of humidity and temperature on the freezing phase change process.

[0069] The feature expansion module is connected to the output of the parameter embedding module. Its main function is to bridge the gap between feature vectors and the two-dimensional spatial structure. This is achieved by configuring a second-level fully connected layer to extend the high-dimensional feature vectors... The dimension is further expanded so that the total number of its elements equals the preset initial feature map tensor. The total number of elements, i.e. ,in , , These represent the number of channels, height, and width of the initial feature map, respectively. Next, a reshape operation is performed, transforming the one-dimensional vector into a three-dimensional tensor structure. The mathematical expression for this process is as follows: ; in, and These represent the weights and biases of the feature expansion layer, respectively. This is a tensor dimension transformation operator. The output is... This is a seed feature map with a preliminary spatial structure.

[0070] The topography generation module consists of multiple cascaded transposed convolutional layers. This module is configured to progressively increase the spatial resolution of the feature maps and refine the geometric texture of the ice-covered topography. For the first... The layer-based generation structure takes the feature map from the previous layer as its input. By applying learnable transposed convolution kernels for upsampling, higher resolution feature maps are output. This hierarchical recursive operation is expressed as: ; in, This indicates the transpose convolution operation. Indicates the first The convolution kernel parameter tensor of the layer, For bias terms, This refers to the inter-layer activation function. In specific network configurations, it is often used in conjunction with batch normalization layers to accelerate training convergence and prevent gradient vanishing. As the number of layers increases, the spatial size of the feature maps... The number of channels increases exponentially with each layer. Reduce layer by layer.

[0071] At the end of the topography generation module, an output layer is set to generate the final two-dimensional icing topography matrix. This output layer uses... Convolution, or a transposed convolution of a specific size, compresses the number of feature channels to 1 (corresponding to a single-channel grayscale image), and applies the Tanh (hyperbolic tangent) or Sigmoid activation function to constrain the output value within the normalized data range (e.g., ...). or The final output matrix is ​​geometrically similar to the real sample. Completely consistent, the spatial distribution of its internal pixel values ​​represents the predicted airfoil leading edge icing profile or thickness field.

[0072] In this embodiment of the invention, the discriminator network is constructed as a deep convolutional neural network classifier, whose core logic function is to distinguish between genuine and fake input two-dimensional icing morphology data. The discriminator network not only evaluates whether the input image conforms to the texture features of physical icing in terms of visual realism, but also verifies whether the icing morphology matches the current operating conditions, thereby providing the generator network with adversarial gradient signals with physical constraints.

[0073] The discriminator network's input is configured to receive joint feature data. To achieve conditional discrimination, the network performs feature fusion at the input layer or in the initial stage of feature extraction. Specifically, the input condition parameter vector... Through fully connected layer mapping or direct spatial dimension replication, its dimension is expanded from a one-dimensional vector to have the same spatial resolution as the input icing topography matrix. The feature mask for this working condition is then applied to a two-dimensional icing topography matrix. By concatenating the data along the channel dimension, a multi-channel composite input tensor is formed. Through this mechanism, the discriminator network can capture the joint probabilistic features between physical conditions such as ambient humidity and liquid water content and the geometric distribution of ice formation.

[0074] The feature extraction section consists of multiple convolutional modules stacked in series. Each convolutional module contains a convolutional layer, a normalization layer, and a non-linear activation layer. The operation process of each convolutional module is shown in the following formula: ; in, Indicates the first The input feature map of the layer; Represents the convolution kernel weight tensor; This represents the convolution operation; Represents the bias vector; This indicates a batch normalization operation, used to stabilize the feature distribution of intermediate layers and accelerate convergence. This represents the activation function of the discriminator. In a preferred embodiment, The Leaky ReLU function is used instead of the standard ReLU function to ensure that small gradient information is preserved even when the input value is negative, preventing neurons from entering a "dead" state during training.

[0075] In setting the stride of the convolutional layers, a stride greater than 1 (usually 2) is used instead of the traditional pooling layers for downsampling. This fully convolutional network structure allows the network to automatically learn the optimal spatial downsampling method, gradually increasing the number of feature channels while reducing the spatial resolution of the feature map, thereby abstracting low-level edge and texture features into high-level overall geometric features.

[0076] The discriminator network's output includes fully connected layers and a probability output layer. The deep feature map obtained after multiple convolutional layers is flattened into a one-dimensional feature vector, or converted into a vector through global average pooling. This vector is input to the final fully connected layer for linear weighted summation, yielding a scalar discriminant score (Logits). Finally, the sigmoid activation function maps this discriminant score to... The interval is used to obtain the final discrimination probability value. The calculation process is expressed as follows: ; in, This represents the input feature vector of the fully connected layer. and These are the weight vector and bias scalar of the output layer, respectively. Output value The confidence level that the input sample belongs to the real observation data is quantified. During training, the discriminator network updates its convolution kernel and weight parameters by maximizing the output probability of real samples and minimizing the output probability of generated samples, thereby enabling it to accurately delineate the boundary between the real icing morphology manifold and the generated morphology manifold in high-dimensional space.

[0077] In one embodiment of the present invention, a system of objective functions for driving model parameter optimization is defined. This system is constructed based on Nash equilibrium game theory and aims to quantify the performance of the generator and discriminator networks during training, transforming discrete prediction errors into differentiable gradient signals, which are then backpropagated to the various weight layers of the neural network. The core objective of training is to solve a minimax game problem, namely, to find a set of generator parameters that minimizes the Jensen-Shannon divergence between the generated ice morphology distribution and the real ice morphology distribution, and simultaneously to find a set of discriminator parameters that maximizes the accuracy of distinguishing between real and fake samples.

[0078] Discriminator Network The optimization objective is to maximize the probability of correctly classifying real icing samples and generating icing samples. To this end, a discriminator loss function is constructed. Its mathematical expression is as follows: ; in, This represents the mathematical expectation operator, used to calculate the average loss value of a batch of samples under a specific distribution. This represents the probability distribution of actual icing morphology data. This represents the prior probability distribution of the input operating condition parameters. The first term... Characterization discriminator for real samples The recognition error; when the input is a real sample, the discriminator outputs... It should approach 1, at which point The value approaches 0, resulting in the minimum loss. (Second term) Characterization discriminator on generated samples The recognition error; when the input is a generated sample, the discriminator outputs... It should approach 0, at which point The value approaches zero, minimizing the loss. During the training step, this is minimized using the gradient descent algorithm. This is equivalent to maximizing the discriminator's ability to distinguish between real and fake samples.

[0079] Generator Networks The optimization goal is to generate realistic icing morphology that can deceive the discriminator, i.e., to induce the discriminator to give a high-confidence "real" judgment to the generated sample. To this end, a generator loss function is constructed. Its mathematical expression is as follows: ; in, This is a normalized vector of chemical condition parameters sampled from the training set. This represents the predicted morphology generated by the generator based on this operating condition. The loss function is designed to maximize the probability that the discriminator will classify the generated sample as real. .when When it approaches 1, Approaching 0, loss function This design avoids the problem of generator gradient vanishing due to an overly strong discriminator in the early stages of training, providing the generator with stronger gradient guidance and enabling it to quickly learn the nonlinear mapping relationship between environmental parameters such as humidity and wind speed and the ice thickness distribution.

[0080] Combining the two loss functions mentioned above, the total value function of the entire generative adversarial network is... This constitutes a binary zero-sum game model. The final optimization objective of this model can be formally expressed as: ; in, Indicates the distribution of real data Real icing morphology samples obtained from sampling , Indicates the distribution of operating parameters The input vector sampled in the middle, during this optimization process, the working condition parameter vector As an explicit constraint, this condition is consistently applied throughout the entire generation and discrimination process. This means that the network not only optimizes the visual realism of the image, but also optimizes the morphology of icing and specific meteorological conditions (especially relative humidity). Physical consistency between them. Through alternating optimization and The system eventually converges to the Nash equilibrium point, at which point the ice morphology distribution generated by the generator is... Approximating the true distribution The discriminator cannot distinguish the source of the sample, that is... It has stabilized around 0.5.

[0081] See attached document Figure 3 In this embodiment of the invention, the training process of the model is designed as a dynamic iterative optimization process, aimed at solving the generator network parameters. Discriminator network parameters The optimal solution is found. To achieve stable and fast convergence on the non-convex loss function surface, a gradient-based optimization algorithm is used to update the network weights.

[0082] Specifically, an adaptive moment estimation optimizer is chosen as the core algorithm for parameter updates. The Adam optimizer combines the advantages of the AdaGrad and RMSProp algorithms, designing independent adaptive learning rates for different parameters by calculating the first and second moment estimates of the gradient. During the initialization phase, the learning rate hyperparameter is set. (e.g., 0.0002-0.0002), the exponential decay rate estimated by the first moment. (e.g., 0.50.5) and the exponential decay rate estimated by the second moment. (e.g., 0.9990.999).

[0083] The model training process is performed cyclically according to the following steps until the preset convergence condition is met: Step S401, mini-batch data sampling. Randomly sample data from the preprocessed training dataset. A mini-batch of samples. Each sample contains a normalized vector of operating condition parameters. and the corresponding real two-dimensional icing morphology matrix ,in For sample index, For the generator input, generate synchronously as needed. A noise vector that follows a standard distribution .

[0084] Step S402, Discriminator network parameters updated. In this step, the weight parameters of the generator network are frozen. Only for the discriminator network parameters To update, first, utilize the current generator network. Each sample contains a normalized vector of operating condition parameters. and noise vector Generate fake icing morphology samples : ; Subsequently, the real samples were paired and generated sample pairs Input to the discriminator network respectively Calculate the loss function value of the discriminator in the current batch. The loss function is calculated using the backpropagation algorithm. Relative to discriminator parameters gradient Finally, the Adam optimizer is applied to update the discriminator parameters based on this gradient: ;in, Indicates the assignment operator. The meaning is the direction of gradient descent. This refers to the Adam optimizer operator, which improves the discriminator's ability to distinguish between the actual icing morphology of wind turbine blades and the morphology predicted by the generative model.

[0085] Step S403, Generator network parameters updated. In this step, the weight parameters of the discriminator network are frozen. Only for generator network parameters Update the generator. The goal of the generator is to generate samples that will cause the discriminator to output a high probability value (i.e., judge it as true). Recalculate or use the sample generated in step S402. Input the discriminator and obtain the discrimination result. Calculate the loss function value of the generator in the current batch. The loss function is calculated using the backpropagation algorithm. Relative to generator parameters gradient The Adam optimizer is then applied to update the generator parameters based on this gradient: ; This step adjusts the weights of the generator network to make the generated 2D icing morphology more consistent with the geometric structure and texture details. This improves the physical authenticity of the sample, thereby reducing the probability that the discriminator will identify it as a fake sample.

[0086] Step S404, Iteration and Convergence Determination. Repeat steps S401 to S403. In each iteration (Epoch), the generator and discriminator compete and evolve together. As the number of training epochs increases, the generator gradually masters the nonlinear mapping law of environmental parameters such as humidity and temperature on the ice distribution, and the quality of the generated morphology gradually improves; at the same time, the discrimination difficulty of the discriminator gradually increases. The training termination condition is set to reach the preset maximum number of iterations, or the output probability of the discriminator stabilizes around 0.5, i.e. At this point, the model reaches Nash equilibrium, meaning the generated distribution... It has extremely approximated the actual data distribution. The discriminator can no longer effectively distinguish between generated and real samples, and model training is complete. Save the generator network parameters at this point. Used for subsequent online prediction tasks.

[0087] In this embodiment of the invention, the prediction implementation stage is based on a trained generative adversarial network model. This stage applies the complex nonlinear mapping capabilities learned offline to inference of actual or simulated working conditions. The prediction process supports both online prediction modes based on real-time monitoring data and offline simulation modes based on set working condition parameters. The prediction implementation steps are executed on a computing device configured with a deep learning inference engine, and specifically include the following steps: Step S501: Acquisition and analysis of target operating condition parameters. The input interface of the computing device receives the target operating condition data to be predicted. In the online prediction implementation, the target operating condition data comes from the SCADA (Supervisory Control and Data Acquisition) system of the wind turbine generator and real-time sensor feedback from the meteorological monitoring station. The sensors collect the current five-dimensional physical quantities in real time: ambient temperature. Incoming wind speed Ambient relative humidity Liquid water content and the median volume diameter of the water droplet In the offline forecasting implementation, the target operating condition data comes from user-preset simulation test sequences or historical meteorological records. The processor parses the acquired physical quantities into raw operating condition vectors. .

[0088] Step S502, Consistency normalization processing of operating condition parameters. To ensure that the distribution of input data is consistent with the feature space during model training, the statistical parameters stored during the training phase must be used to normalize the original operating condition vector. Perform normalization. Retrieve the global maximum vector of the training dataset from memory. and global minimum vector For the first element in the target working condition vector... Each component The normalized input value is calculated using the following formula. : ;in, and This must strictly correspond to the numerical boundaries determined during model training, rather than the maximum or minimum value of the current batch to be predicted. This step ensures that the physical parameters are correctly mapped to the sensitive response range of [0,1] before being input into the generator network, generating the normalized vector to be predicted. .

[0089] Step S503, forward inference computation of the generator network. The computing device loads the trained and fixed weight parameters of the generator network. The normalized vector obtained in step S502 The noise is fed as a conditional input to the generator network. In implementations that include noisy input, the processor synchronously generates a random noise vector that follows a standard normal distribution. and will and The concatenated input is then used. The generator network performs a single forward propagation operation, passing through feature mapping in the parameter embedding layer, dimensionality expansion in the fully connected layer, and upsampling and nonlinear activation in multiple transposed convolutional layers, directly outputting the prediction matrix. : ; During this process, the discriminator network no longer participates in the computation. The generator network reconstructs the corresponding two-dimensional ice morphology data within milliseconds based on the mapping weights between parameters such as environmental humidity and temperature and the geometric features of the ice shape stored internally.

[0090] Step S504: Output and caching of prediction results. The generator output... A normalized two-dimensional grayscale matrix or feature map, whose numerical distribution represents the target working condition. The probability distribution or relative thickness distribution of icing on the leading edge of the airfoil of a wind turbine blade. The computing device will then process this matrix. The data is temporarily stored in video memory or a memory buffer for subsequent denormalization processing and visualization rendering. For online monitoring applications, this prediction process is triggered periodically as the operating condition data is updated; for offline analysis applications, this process can be executed in parallel for a large number of operating condition sequences, thereby quickly establishing an icing database.

[0091] While the two-dimensional matrix output by the neural network contains characteristic information about icing morphology in its numerical distribution, it is essentially still an abstract feature map or normalized image data, and cannot be directly used for aerodynamic performance calculations or engineering evaluations. Therefore, the processor needs to perform a series of post-processing operations to transform the output tensor of the neural network into icing airfoil data with clear geometric coordinates and physical dimensions in physical space. The post-processing process specifically includes data denormalization, noise filtering and threshold truncation, and geometric coordinate reconstruction steps.

[0092] Step S505: Denormalization of the prediction results. The processor receives the prediction matrix output by the generator. The elements in this matrix are in the range [0,1] or [...]. Within the relative interval of [1,1]. To recover the true physical dimensions of the ice thickness, the maximum ice thickness parameter recorded during the training phase needs to be used. Perform the inverse transformation. For the matrix index ... elements The corresponding physical ice layer thickness The calculation formula is as follows: ;in, This is the inverse mapping function corresponding to the aforementioned preprocessing stage. If linear normalization is used in the preprocessing, the multiplication is performed directly here; if nonlinear operations such as logarithmic transformation are used in the preprocessing, the corresponding inverse exponential transformation is performed here. Through this step, the dimensionless grayscale values ​​are restored to the actual ice layer thickness values ​​in millimeters (mm), generating a physical thickness distribution matrix. .

[0093] Step S506: Noise Filtering and Threshold Truncation. Due to the generative mechanism of generative adversarial networks, the background region (i.e., the ice-free region) of the output matrix will have non-zero small numerical fluctuations, which manifest as background noise. To obtain clear ice-shaped boundaries, the processor applies a threshold truncation algorithm to the physical thickness distribution matrix. Perform cleaning. Set a thickness judgment threshold. (e.g., 0.1mm) for any element in the matrix The following decision logic is executed: ; After truncation, background noise interference is eliminated, and the topological boundaries between the icing region and the air region are clearly distinguished, resulting in a cleaning matrix that can accurately characterize the presence of ice. .

[0094] Step S507, Geometric Coordinate Reconstruction and Visualization Mapping. The cleaned matrix data is mapped back to the physical coordinate system of the airfoil leading edge. During the data preprocessing stage, the original airfoil leading edge was discretized and mapped as... The raster system here performs the inverse geometric transformation of this process. This is based on the preset raster physical size resolution. and the positioning coordinates of the airfoil leading edge stagnation point , index the matrix Convert to spatial coordinates in airfoil coordinate system .

[0095] ; The processor will clean the matrix The non-zero thickness values ​​are superimposed onto the airfoil reference profile normal direction, or the ice shape profile is directly reconstructed in the coordinate system in the form of a point cloud. Discrete outer edge points of the ice shape are connected using spline curve fitting algorithms (such as B-spline or cubic spline interpolation) to generate a continuous, closed icing airfoil geometry curve. This geometry curve directly reflects the ice layer coverage at the airfoil leading edge, the location of maximum ice thickness, and the specific shape of the ice angle under the target operating conditions.

[0096] Step S508, Quantitative assessment of icing severity. Based on the reconstructed geometry curves from step S507, the processor further calculates key metrics for engineering evaluation, including maximum icing thickness. and the cross-sectional area of ​​ice Cross-sectional area of ​​ice The following is obtained by numerically integrating and summing all non-zero elements in the cleaning matrix: ; in, The meaning is a double summation traversal loop. The meaning is the physical area of ​​a single grid cell. The computing device will generate an icing airfoil geometry coordinate file (such as .dat or .csv format) and , The output is displayed on a terminal for visualization or transmitted to the wind turbine's main control system via a data interface. Based on the predicted icing severity, the main control system automatically determines whether to trigger shutdown protection logic or initiate active anti-icing and de-icing operations such as electrothermal de-icing or gas-thermal de-icing, thereby achieving closed-loop control based on deep learning prediction results. For the curve fitting and numerical integration methods mentioned above, those skilled in the art can implement them using standard mathematical library functions; specific code implementation details will not be elaborated here.

Claims

1. A deep learning-based method for predicting icing on wind turbine blades, characterized in that, Includes the following steps: Step S1: Obtain two-dimensional icing morphology data and corresponding operating parameters of wind turbine blades under different operating conditions; the two-dimensional icing morphology data characterizes the icing profile or icing thickness distribution on the airfoil section of the blade. Step S2: Normalize the working condition parameters and map the two-dimensional icing morphology data into a two-dimensional matrix with a preset size to construct training sample pairs; Step S3: Construct a generative adversarial network model, which includes a generator network and a discriminator network that are in conflict with each other; Step S4: Input the processed real two-dimensional icing morphology data and the corresponding real working condition parameters into the generative adversarial network model; according to the preset loss function, alternately update the parameters of the generator network and the discriminator network until the generative adversarial network model reaches Nash equilibrium. Step S5: Receive the operating condition parameters to be predicted and input them into the trained generator network. The generator network outputs the two-dimensional icing morphology prediction results of the wind turbine blades under the corresponding operating conditions.

2. The deep learning-based method for predicting icing on wind turbine blades according to claim 1, characterized in that, In step S1, the operating parameters constitute a multi-dimensional vector, and the dimensions of the multi-dimensional vector include at least ambient temperature, incoming wind speed, ambient relative humidity, liquid water content, and median volume diameter of water droplets; in the multi-dimensional vector, the ambient relative humidity is used as an independent input variable to characterize the influence of water vapor content in the air on the icing process. In step S2, the normalization process uses the maximum and minimum values ​​obtained from the training dataset to map the operating condition parameters to a preset numerical range. The processing of the two-dimensional icing morphology data includes coordinate alignment, size unification, and rasterization operations, which convert the continuous airfoil leading edge icing profile into the two-dimensional matrix.

3. The deep learning-based method for predicting icing on wind turbine blades according to claim 1, characterized in that, In step S3, the generator network and the discriminator network are configured as a conditional generative adversarial network topology: The generator network is configured to receive the normalized operating parameters as conditional inputs and map the input operating parameters to predicted icing morphology samples. The discriminator network is configured to receive a two-dimensional icing morphology matrix as input, the input including real icing morphology samples and predicted icing morphology samples; The discriminator network simultaneously receives the normalized operating condition parameters as constraints and outputs a probability scalar representing that the input sample is real data.

4. The deep learning-based method for predicting icing on wind turbine blades according to claim 3, characterized in that, The generator network includes a parameter embedding module, a feature expansion module, and a shape generation module connected in sequence. The parameter embedding module uses a fully connected layer to extract features from the input operating condition parameters, mapping low-dimensional physical parameters into high-dimensional feature vectors. The feature expansion module uses a fully connected layer to expand the dimension of the high-dimensional feature vector and transforms it into a seed feature map with a three-dimensional tensor structure through a reshaping operation. The morphology generation module consists of multiple cascaded transposed convolutional layers. The seed feature map is upsampled through transposed convolution operations to gradually improve the spatial resolution of the feature map and reduce the number of channels. The predicted icing morphology sample is generated through the output layer.

5. The deep learning-based method for predicting icing on wind turbine blades according to claim 3, characterized in that, The discriminator network includes feature fusion operations and multiple cascaded convolutional modules: In the input layer or feature extraction stage, the discriminator network expands the spatial dimension of the input working condition parameters to form a feature mask with the same spatial resolution as the input two-dimensional icing morphology matrix, and then splices the feature mask and the two-dimensional icing morphology matrix in the channel dimension to form a composite input tensor. The convolution module includes convolutional layers, normalization layers, and nonlinear activation layers. It extracts the geometric texture and spatial distribution features of the input through convolution operations, and outputs the probability scalar through fully connected layers and activation functions.

6. The deep learning-based method for predicting icing on wind turbine blades according to claim 1, characterized in that, In step S4, the loss function includes a discriminator loss function and a generator loss function, and the alternating parameter update process solves the minimax game problem: The discriminator loss function is constructed based on the log-expected value of the probability that a real sample is judged as real and the log-expected value of the probability that a generated sample is judged as fake. The optimization objective is to maximize the accuracy of the discriminator network in distinguishing between real samples and generated samples; the generator loss function is constructed based on the log-expectation of the probability that a generated sample is judged as real. The optimization objective is to maximize the probability that the discriminator network classifies a generated sample as real.

7. The deep learning-based method for predicting icing on wind turbine blades according to claim 6, characterized in that, In step S4, the alternating update of the parameters of the generator network and the discriminator network includes cyclically performing the following steps: Freeze the weight parameters of the generator network, and use the generator network to generate fake icing morphology samples based on the operating parameters of the current batch; The discriminator network is input with real sample pairs and generated sample pairs, the discriminator loss function is calculated, and the parameters of the discriminator network are updated using an adaptive moment estimation optimizer. Freeze the weight parameters of the discriminator network, use the generator network to generate fake icing morphology samples and input them into the discriminator network; The generator loss function is calculated and the parameters of the generator network are updated using an adaptive moment estimation optimizer.

8. The deep learning-based method for predicting icing on wind turbine blades according to claim 1, characterized in that, In step S5, the prediction includes: Obtain the target working condition data to be predicted, and use the statistical parameters stored in the training phase to perform consistency normalization on the target working condition data to obtain a normalized vector. The normalized vector is fed as a conditional input to the trained generator network, which performs forward inference computation and directly outputs a normalized two-dimensional gray matrix as the prediction result.

9. The deep learning-based method for predicting icing on wind turbine blades according to claim 8, characterized in that, Following step S5, the method further includes a post-processing step for the prediction result, the post-processing including: Using the maximum ice thickness parameter recorded during the training phase, the two-dimensional gray matrix is ​​subjected to an inverse denormalization transformation to restore the dimensionless gray values ​​to the physical ice thickness values, thereby generating a physical thickness distribution matrix. The physical thickness distribution matrix is ​​truncated by applying a preset thickness determination threshold, and elements smaller than the thickness determination threshold are set to zero to eliminate background noise and distinguish between icing areas and air areas.

10. The deep learning-based method for predicting icing on wind turbine blades according to claim 9, characterized in that, The post-processing also includes: Based on the preset grid physical size resolution and the positioning coordinates of the airfoil leading edge stagnation point, the non-zero elements in the truncated physical thickness distribution matrix are mapped back to the airfoil coordinate system. By connecting discrete outer edge points of the ice shape using a curve fitting algorithm, a continuous and closed geometric curve of the icing airfoil can be reconstructed. Based on the reconstructed icing airfoil geometry, the maximum icing thickness and icing cross-sectional area are calculated to quantitatively assess the severity of icing.

Images