Wind turbine multi-condition flow field prediction method and device based on neural network

Abstract

The application relates to the technical fields of wind power generation, artificial intelligence and fluid mechanics, and discloses a wind turbine multi-working-condition flow field prediction method and device based on a neural network. The prediction method comprises the following steps: obtaining a sample data set, wherein the sample data set comprises multiple groups of multi-working-condition data; inputting the sample data set into a physical information field adversarial neural network model for training, so as to obtain the total loss of the sample data in forward propagation in the field adversarial neural network model. According to the total loss, it is determined whether the wake flow field prediction model is trained. In the case of training the wake flow field prediction model for a new round, the total loss is back propagated to optimize the neural network parameters. In this way, a double-branch loss collaborative optimization mechanism is formed. The physical information neural network and the field adversarial neural network are combined to fully exert the complementary advantages of the two. When the data is limited or the distribution difference is large, the wake flow field prediction with high precision and physical consistency can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of wind power generation, artificial intelligence and fluid mechanics, for example to a wind turbine multi-condition flow field prediction method and device based on a neural network. BACKGROUND

[0002] When a wind turbine is running, a complex wake flow field will be formed behind the impeller, and the flow field has highly nonlinear and strong space-time coupling characteristics. The shape and strength of the wake directly affect the wind farm layout, power generation efficiency and wind turbine life.

[0003] Currently, the prediction of the flow field can be achieved by the following four methods: a numerical simulation method based on control equations (such as CFD (Computational Fluid Dynamics)), a physical-based analytical model method (such as Jensen (a model for describing the wake effect of wind turbines in wind farms) and its variants), a data-driven neural network model method, and a field measurement method (such as PIV (Particle Image Velocimetry), LiDAR (Light Detection and Ranging), etc.). However, these methods have certain defects: the numerical simulation method has a large amount of simulation calculation and takes a long time, which is difficult to meet the real-time optimization and rapid iteration requirements in the actual engineering of wind farms. The physical-based analytical model method does not accurately describe the evolution of the vortex, and the precision of the flow details is low. The data-driven neural network prediction method is mainly limited to a single working condition, and the model generalization ability is insufficient, which is difficult to cope with the multiple influences of changes in incoming wind speed and yaw angle on the flow field structure. The field measurement method is limited by the resolution of the equipment, the spatial distribution and the time continuity, and it is difficult to obtain comprehensive, accurate, continuous and detailed three-dimensional flow field data.

[0004] The reason for the above-mentioned defects is that the existing model either ignores the complex physical laws, or only relies on data-driven, lacks multi-physical source information fusion and field adaptation ability, and is difficult to simultaneously consider high precision, physical consistency and multi-condition adaptability.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] In order to have a basic understanding of some aspects of the disclosed embodiments, the following is a simple summary. The summary is not a general review, nor is it intended to determine the key / important components or delineate the scope of protection of these embodiments, but as a prelude to the detailed description below.

[0007] The embodiment of the present disclosure provides a wind turbine multi-condition flow field prediction method and device based on a neural network, so as to realize high-precision prediction of a three-dimensional flow field.

[0008] In some embodiments, the wind turbine multi-condition flow field prediction method based on the neural network comprises: obtaining a sample data set, wherein the sample data set comprises multiple sets of multi-condition data; inputting the sample data set into a physical information field adversarial neural network model for training to obtain a total loss of forward propagation of sample data in the field adversarial neural network model; wherein the total loss comprises: data loss, physical equation loss and field classification loss; determining whether the wake flow field prediction model is trained according to the total loss; in the case of a new round of training of the wake flow field prediction model, the total loss is back propagated to optimize the neural network parameters.

[0009] In some embodiments, the wind turbine multi-condition flow field prediction device based on the neural network comprises: an obtaining module configured to obtain a sample data set, wherein the sample data set comprises multiple sets of multi-condition data; a training module configured to input the sample data set into a physical information field adversarial neural network model for training to obtain a total loss of forward propagation of sample data in the field adversarial neural network model; wherein the total loss comprises: data loss, physical equation loss and field classification loss; a determination module configured to determine whether the wake flow field prediction model is trained according to the total loss; and an optimization module configured to, in the case of a new round of training of the wake flow field prediction model, back propagate the total loss to optimize the neural network parameters.

[0010] In some embodiments, the wind turbine multi-condition flow field prediction device based on the neural network comprises: a processor and a memory storing program instructions, wherein the processor is configured to execute the wind turbine multi-condition flow field prediction method based on the neural network when running the program instructions.

[0011] In some embodiments, the electronic device comprises: an electronic device body; and the wind turbine multi-condition flow field prediction device based on the neural network is installed on the electronic device body.

[0012] The wind turbine multi-condition flow field prediction method and device based on the neural network provided by the embodiment of the present disclosure can achieve the following technical effects:

[0013] The sample data set input to the physical information field adversarial neural network model includes multiple groups of multi-working condition data. Compared with inputting a single working condition, the generalization ability of the wake flow field prediction model can be improved, and then the multiple influences of the change of the incoming flow wind speed and the yaw angle on the flow field structure can be fully dealt with. Moreover, the DANN effectively improves the generalization ability of the model under multiple working conditions by introducing field classification and adversarial training mechanism, so that it can adapt to the changes of different wind speeds, yaw angles and other complex working conditions. By calculating the total loss of the sample data forward propagation in the model, it is determined whether the wake flow field prediction model is trained. If a new round of training is performed, the total loss is back propagated, and the neural network parameters are optimized by using the total loss. In this way, a double-branch loss cooperative optimization mechanism is formed. In this way, the physical information neural network and the field adversarial neural network are combined to fully exert the complementary advantages of the two. In the case of limited data or large distribution difference, high-precision and physically consistent wake flow field prediction can still be realized, which provides a solid theoretical basis and data support for wind turbine arrangement optimization and control strategy.

[0014] The foregoing general description and the following description are only exemplary and explanatory, and are not intended to limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0015] One or more embodiments are exemplarily illustrated by corresponding drawings, which do not constitute limitations on the embodiments, elements with the same reference numerals in the drawings are shown as similar elements, the drawings do not constitute proportional limits, and wherein:

[0016] Figure 1 is a schematic diagram of a wake flow prediction model of a wind turbine based on a physical information field adversarial neural network provided by an embodiment of the present disclosure;

[0017] Figure 2 is a basic structure schematic diagram of a DANN provided by an embodiment of the present disclosure;

[0018] Figure 3 is a schematic diagram of a first neural network-based wind turbine multi-working condition flow field prediction method provided by an embodiment of the present disclosure;

[0019] Figure 4 is a schematic diagram of a second neural network-based wind turbine multi-working condition flow field prediction method provided by an embodiment of the present disclosure;

[0020] Figure 5 is a schematic diagram of a third neural network-based wind turbine multi-working condition flow field prediction method provided by an embodiment of the present disclosure;

[0021] Figure 6 is a schematic diagram of a fourth neural network-based wind turbine multi-working condition flow field prediction method provided by an embodiment of the present disclosure;

[0022] Figure 7 is a schematic diagram of a prediction result of the PI-DANN provided by an embodiment of the present disclosure under the working condition of

[0023] Figure 8 is a schematic diagram of a prediction result of the PI-DANN provided by an embodiment of the present disclosure under the working condition of

[0024] Figure 9 is a schematic diagram of a first neural network-based wind turbine multi-working condition flow field prediction device provided by an embodiment of the present disclosure;

[0025] Figure 10 is a schematic diagram of a second neural network-based wind turbine multi-working condition flow field prediction device provided by an embodiment of the present disclosure;

[0026] Figure 11 is a schematic diagram of a first neural network-based wind turbine multi-working condition flow field prediction device provided by an embodiment of the present disclosure installed in an electronic device;

[0027] Figure 12 is a schematic diagram of a second neural network-based wind turbine multi-working condition flow field prediction device provided by an embodiment of the present disclosure installed in an electronic device.

[0028] Reference signs:

[0029] 90, a first neural network-based wind turbine multi-working condition flow field prediction device; 91, an acquisition module; 92, a training module; 93, a determination module; 94, an optimization module;

[0030] 100, a second neural network-based wind turbine multi-working condition flow field prediction device; 101, a processor; 102, a memory; 103, a communication interface; 104, a bus;

[0031] 110, an electronic device. DETAILED DESCRIPTION

[0032] In order to enable persons skilled in the art to more fully understand the features and technical contents of the embodiments of the present disclosure, the implementation of the embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings, which are used only for reference and do not limit the embodiments of the present disclosure. In the following technical description, in order to facilitate explanation, a plurality of details are provided to provide a full understanding of the disclosed embodiments. However, one or more embodiments can still be implemented without these details. In other cases, in order to simplify the drawings, well-known structures and devices can be simplified.

[0033] ​​The terms "first", "second", and the like in the description and claims of the present disclosure and the above drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion.

[0034] Unless otherwise specified, the term "a plurality of" means two or more.

[0035] In the present disclosure, the character " / " represents a "or" relationship between the objects before and after it. For example, A / B means: A or B.

[0036] The term "and / or" is a description of the association between objects, which means that there can be three relationships. For example, A and / or B means: A or B, or, A and B, the three relationships.

[0037] The term "corresponding" can refer to an association or binding relationship, A corresponding to B means that there is an association or binding relationship between A and B.

[0038] In combination Figure 1 As shown, the present disclosure provides a wind turbine wake prediction model based on a physics-informed domain adversarial neural network (PI-DANN). The model includes a physics-informed neural network (PINN) and a domain adversarial neural network (DANN).

[0039] The domain adversarial neural network includes:

[0040] (1) Input layer:

[0041] The input features are multiple sets of working condition data, and each set of working condition data includes a five-dimensional vector composed of three-dimensional space coordinates , incoming flow speed , and yaw angle of the wind turbine .

[0042] (2) DANN basic structure, in combination Figure 2 As shown, it includes:

[0043] (2.1) Feature extractor (Feature Extractor, G f ):

[0044] The feature extractor G f is composed of 2-8 layers of long short-term memory units (LSTM), with 8-1024 neurons in each layer. Each layer is followed by a hyperbolic tangent activation function (Tanh), and batch normalization (Batch Norm) is used in some layers to improve training stability.

[0045] (2.2) Wake predictor (also referred to as Regression predictor, G r ):

[0046] The wake predictor G r is composed of 1-3 layers of fully connected neural networks (FCNN) with Tanh activation function. It outputs the target flow field physical quantities, including: the pressure of air and three-dimensional air velocity components.

[0047] (2.4) Working condition classifier (also referred to as Domain Classifier, G d ):

[0048] The feature extractor G f outputs are passed to the working condition classification branch through the gradient reversal layer (GRL). This branch is composed of 1-3 layers of FCNN, with 8-512 neurons in each layer and a Tanh activation function. The final output is the classification probability of each working condition.

[0049] During the training of DANN, the feature extractor G f first extracts deep features from the input data. These features are then processed through two parallel paths: on the one hand, they are used as input to the regression prediction network G r composed of fully connected neural networks, to predict the target physical quantities (velocity, pressure); on the other hand, they are passed to the working condition classifier G d through the gradient reversal layer (GRL) to identify the working condition of the input sample.

[0050] (3) Output layer: The final output of the network is the pressure and three-dimensional velocity components at a given spatial location under the predicted working condition.

[0051] The loss function in the physical information neural network includes:

[0052] (1) Regression prediction loss function , each term uses the mean squared error (MSE) form. It includes:

[0053] (1.1) Conventional data loss function .

[0054] (1.2) Navier-Stokes (NS) equation loss function for three-dimensional incompressible fluid .

[0055] (1.3) Vorticity equation loss function , Losses constitute physical equation (PDE) losses.

[0056] (2) Domain classification loss function , which is calculated by a cross-entropy loss function.

[0057] Based on the aforementioned wind turbine wake prediction model, combined with Figure 3 , the embodiment of the disclosure provides a wind turbine multi-condition flow field prediction method based on a neural network, which comprises the following steps:

[0058] S101, acquiring a sample data set, wherein the sample data set comprises multiple groups of multi-condition data.

[0059] S102, inputting the sample data set into a physical information domain adversarial neural network model for training to obtain a total loss of the sample data in the forward propagation of the domain adversarial neural network model, wherein the total loss comprises a data loss, a physical equation loss and a domain classification loss.

[0060] S103, determining whether the wake flow field prediction model is trained according to the total loss.

[0061] S104, in the case of training the wake flow field prediction model for a new round, performing back propagation on the total loss, and optimizing the neural network parameters by using the total loss.

[0062] Each combination of incoming wind speed and yaw angle is regarded as an independent Domain (set). Specifically, the incoming wind speed m / s, and the yaw angle , a total of 20 Domains. Each sample is labeled with a Domain label , which is used for adversarial training of the working condition classifier. In the process of model training and testing, the “leave-one-domain-out” method is used for evaluation: 19 of the 20 Domains are selected as Source Domains (training set) each time, and the remaining 1 is selected as Target Domain (test set), which is used to evaluate the adaptability and robustness of the model to different physical conditions.

[0063] The sample data set, that is, the training set, is acquired. The sample data set is input into the physical information domain adversarial neural network model for training, and the total loss in the process of forward propagation of the sample data set in the neural network model is calculated​ total loss including: data loss , physical equation loss and domain classification loss .

[0064] Whether the wake field prediction model is trained is determined according to the total loss. If the training is completed, the training is ended. If the training is not completed, the wake field prediction model is trained for a new round.

[0065] When the wake field prediction model is trained for a new round, the total loss is back propagated. In the process of back propagation, the total loss is used to optimize the total parameters of the neural network, wherein the total parameters of the neural network include: the parameters of the feature extractor , the parameters of the wake predictor and the parameters of the working condition classifier . The regression prediction loss is used to optimize the parameters of the wake predictor , and the domain classification loss is used to update the parameters of the working condition classifier . In this way, in each round of model training, the total parameters of the neural network are optimized and updated until the model training is completed. At the same time, the feature extractor (G f ) receives the back gradient through the gradient reversal layer (GRL) to maximize the domain classification error in an adversarial manner while ensuring the regression prediction accuracy, so as to force the network to extract working condition independent general flow field features, and finally improve the adaptability of the model to new working conditions on the premise of maintaining the prediction accuracy.

[0066] The method for predicting the flow field of a wind turbine under multiple working conditions based on a neural network provided by the embodiment of the present disclosure is used. The sample data set input to the physical information field adversarial neural network model includes multiple sets of multiple working condition data. Compared with inputting a single working condition, the generalization ability of the wake flow field prediction model can be improved, and the multiple influences of changes in the incoming flow speed and yaw angle on the flow field structure can be fully addressed. Moreover, the DANN effectively improves the generalization ability of the model under multiple working conditions by introducing a field classification and adversarial training mechanism, so that the model can adapt to changes in complex working conditions such as different wind speeds and yaw angles. The total loss of the sample data forward propagation in the model is calculated to determine whether the wake flow field prediction model is trained. If a new round of training is performed, the total loss is back propagated, and the neural network parameters are optimized using the total loss. In this way, a double-branch loss cooperative optimization mechanism is formed. In this way, the physical information neural network and the field adversarial neural network are combined to fully exert the complementary advantages of the two. In the case of limited data or large distribution difference, high-precision and physically consistent wake flow field prediction can still be achieved, providing a solid theoretical basis and data support for wind turbine arrangement optimization and control strategies.

[0067] Optionally, the total loss is determined by the following method, comprising:

[0068] The total loss includes a regression prediction loss and a field classification loss, and is specifically shown in equation (1). The field classification loss further includes a data loss, an NS equation loss and a vorticity equation loss, and is specifically shown in equation (2).

[0069] First, the mean square error of the three-dimensional air velocity vector prediction value and the true value of a single sample is calculated, and the sum of the mean square errors of all samples is determined as the data loss, and is specifically shown in equation (3). Then, the residual of the NS equation and the residual of the vorticity equation of each spatial coordinate vector are calculated, the sum of the residuals of the NS equation of each spatial coordinate vector is determined as the NS equation loss, and is specifically shown in equation (4). The sum of the residuals of the vorticity equation of each spatial coordinate vector is determined as the vorticity equation loss, and is specifically shown in equation (5). The sum of the NS equation loss and the vorticity equation loss is determined as the physical equation loss.

[0070] (1)

[0071] (2)

[0072] (3)

[0073] (4)

[0074] (5)

[0075] (6)

[0076] where, represents the loss function, the subscript respectively represent the total loss, data item, NS equation item, vorticity equation item and domain classification item loss function; is the weight coefficient of the domain classification item; is the weight of the NS equation loss item, is the weight of the vorticity equation loss item, The value of is dynamically adjusted during the training process; the superscript is the predicted value, is the true value; represents the spatial coordinate vector , is the three-dimensional air velocity vector , represents the selected data point; is the residual of the NS equation, is the residual of the vorticity equation; represents the working condition label, is the total number of all working conditions; is the pressure of the air, is the incoming flow speed, is the yaw angle.

[0077] The total loss function (equation 1) of PI-DANN is composed of regression prediction loss and domain classification loss. The regression loss adopts the mean square error (MSE) form, and the domain classification loss (equation 6) is calculated by the cross-entropy loss function. The regression prediction loss (equation 2) not only includes the conventional data loss (equation 3), but also includes the PDE loss composed of the Navier-Stokes equation loss (equation 4) and the vorticity equation loss (equation 5) of three-dimensional incompressible fluid. This composite loss function design not only maintains the flexibility of data-driven, but also enhances the expression ability of the model through physical constraints, and improves the cross-condition adaptability with the help of domain adversarial mechanism. The composite physical loss constraint method of parallel embedding of NS equation and vorticity equation realizes the deep integration of physical prior knowledge and domain adversarial learning, and can be used for physical consistency prediction of complex three-dimensional vortex structure flow field.

[0078] The NS equation set (equation 7) describes the basic evolution law of velocity and pressure in fluid, so that the prediction result satisfies the conservation of momentum and mass.

[0079] (7)

[0080] where, respectively represent the density and kinematic viscosity of air, is the total residual of the NS equation set, For the momentum conservation term, For the mass conservation term.

[0081] And because the wind wake field has significant local rotation, vortex structure stretching and merging characteristics, in the high Reynolds number flow field, only considering the NS equation is not enough to finely describe the actual vorticity distribution. Introducing the vorticity equation (equation 8) can make the model more sensitive to the evolution of vortex clusters and improve the accuracy of wind turbine complex flow field prediction.

[0082] (8)

[0083] wherein, is the residual of VE. Ω represents the vorticity, which is the vector of the rotational motion of the fluid element, defined as the curl of the velocity vector, and its calculation method is shown in equation 9.

[0084] (9)

[0085] The total loss function includes data fitting loss, physical equation loss and adversarial domain loss, and the weights (adversarial weight coefficient and physical loss weight w 1 、w 2)and learning rate are all dynamically adjusted, the purpose is to maintain the balance of each loss term and realize the efficient training of collaborative optimization. The adjustment method is as follows:

[0086] (1) Dynamic adjustment method of adversarial domain loss weight:

[0087] In DANN, the contribution proportion of the domain classification loss ( ) in the total loss ( ) function is controlled by the adversarial weight coefficient , which balances the regression prediction accuracy and the generalization ability of cross-condition characteristics. In order to ensure the stability of training, this embodiment adopts a gradual adjustment strategy: in the early stage of training ( is small), it focuses on main task feature learning, and gradually enhances cross-condition feature extraction ( increases) as the training progresses. This dynamic adjustment process follows formula (10):

[0088] (10)

[0089] wherein, represents the training progress (i.e. the ratio of the current training step number to the total step number, ), controls the rate of weight growth.

[0090] (2) Dynamic adjustment method of physical equation term loss weight:

[0091] Another component of the total loss is the regression prediction loss (Equation 2). To balance the data item, the NS equation item, and the VE equation item, and prevent one item from dominating the update of the model parameters during the training process, an adaptive dynamic weight adjustment strategy based on the gradient is adopted in this embodiment. Specifically, during the backpropagation process, the model parameters are updated by the gradient descent algorithm: the gradient of the data loss, the gradient of the NS equation loss, the gradient of the VE equation loss, and the gradient of the domain classification loss are calculated for all neural network parameters. The gradient of the NS equation loss is adjusted using the NS equation loss coefficient weight, the gradient of the VE equation loss is adjusted using the VE equation loss coefficient weight, and the gradient of the domain classification loss is adjusted using the domain classification loss coefficient weight. According to the neural network parameters corresponding to this iteration, the learning rate, the gradient of the data loss, the adjusted gradient of the NS equation loss, the adjusted gradient of the VE equation loss, and the adjusted gradient of the domain classification loss, the neural network parameters corresponding to the next iteration are determined. Specifically, as shown in Equation (11):

[0092] (11)

[0093] wherein, is the iteration round number. is the learning rate. represents the total parameters of the neural network, including the parameters of the feature extractor , the parameters of the wake prediction model , and the parameters of the working condition classifier .

[0094] In the first iteration, the estimated values of the loss coefficient weights of the physical equations , are dynamically estimated by the ratio of the data loss gradient to the corresponding physical equation loss gradient in the first iteration, as shown in Equation (12):

[0095] (12)

[0096] In Equation (12), represents the estimated weight value, represents the gradient of the loss calculated for all model parameters , and represents the average of the absolute values of the gradients, which is calculated by Equation (13):

[0097] (13)

[0098] In Equation (13), M is the number of model parameters.

[0099] To suppress the possible drastic fluctuations of the weights during the training process, a weighted smoothing strategy is introduced in the updating process, which improves the stability of the weights while maintaining their adaptive adjustment capability, as shown in equation (14):

[0100] (14)

[0101] In equation (14), is the updating coefficient, which controls the magnitude of the weight adjustment. Optionally, is set to 0.1.

[0102] On the basis of the dynamic weight strategy, setting a reasonable initial weight helps to accelerate the convergence of the model. Based on this, the initial weight is set as (the NS term), (the VE term).

[0103] The dynamic weight updating and adaptive learning rate updating strategies of the above loss terms can achieve stable convergence under the multi-source loss function.

[0104] Optionally, before the model training, a spatially structured down-sampling strategy is used to construct a training set for the subsequent PI-DANN model. First, based on the area range of the CFD simulated flow field, one-eighth of the diameter of the wind turbine impeller is taken as the interval to generate uniformly distributed grid reference points. Second, taking half of the maximum grid size in the CFD simulation as the error threshold, the fast KD tree (K-dimension tree, a data structure used to organize k-dimensional space data points in computer science) nearest neighbor search algorithm is used to select about 2.5% representative sample points from the original high-resolution data.

[0105] Optionally, to further optimize the utilization efficiency of computing resources, a dynamic batch loading mechanism is introduced. During the model training process, each batch loading batch size (batch size) group of sampling data is used to calculate the data loss, and a number of spatial points equal to half of the batch size are randomly generated for the calculation of the PDE loss. This mechanism not only avoids the risk of memory overflow, but also improves the training stability of the PI-DANN model.

[0106] Optionally, as shown in Figure 4 , the present disclosure provides another neural network-based method for predicting the flow field of a wind turbine under multiple operating conditions, which includes:

[0107] S101, obtaining a sample data set, wherein the sample data set includes multiple groups of multi-condition data.

[0108] S105, using the maximum and minimum value normalization method to linearly map the original physical quantities in the sample data set to the [0, 1] interval.

[0109] S106, input the sample data set into the physical information field adversarial neural network model for training, to obtain the total loss of the sample data in the field adversarial neural network model forward propagation, the total loss includes: data loss, physical equation loss and field classification loss; wherein, when calculating the physical equation residual, the predicted data is restored to the original physical dimension by inverse normalization.

[0110] S103, according to the total loss, determine whether the wake flow field prediction model is trained.

[0111] S104, in the case of a new round of training of the wake flow field prediction model, the total loss is back propagated, and the neural network parameters are optimized by using the total loss.

[0112] In order to accelerate the model convergence and adapt to multi-working condition tasks, an improved normalization method based on gradient compensation is proposed in this embodiment. Unlike the traditional dimensionless method (which depends on specific criterion numbers and is difficult to adapt to working condition changes), this method first uses Min-Max (minimum-maximum) normalization (formula 15) to linearly map the input parameters to the [0, 1] interval. When calculating the PDE residual (Loss PDE), the network prediction value is restored to the original physical dimension through inverse normalization (formula 16) to ensure the accuracy of the physical constraint.

[0113] (15)

[0114] (16)

[0115] wherein, is the normalized variable, is the original physical quantity, is the minimum value, is the maximum value.

[0116] After the above steps are completed, the PI-DANN model is used to train the multi-working condition data set end to end. The optimizer uses Adam, and the initial learning rate is 10 -4 , which is adjusted adaptively. Dynamic weight update and smooth normalization compensation are performed every iteration, and the batch size is 128-2048.

[0117] Optionally, in combination with Figure 5 , the embodiment of the disclosure provides another neural network-based prediction method for multi-working condition flow field of a wind turbine, comprising:

[0118] S101, obtaining a sample data set, wherein the sample data set includes multiple groups of multi-working condition data.

[0119] S105, the original physical quantity in the sample data set is linearly mapped to the interval [0, 1] by using the maximum minimum value normalization method.

[0120] S106, the sample data set is input into the physical information field adversarial neural network model for training, to obtain the total loss of the sample data in the field adversarial neural network model for forward propagation, the total loss includes: data loss, physical equation loss and field classification loss; wherein, when calculating the physical equation residual, the predicted data is restored to the original physical dimension by reverse normalization.

[0121] S103, according to the total loss, whether the wake flow field prediction model is trained is determined.

[0122] S107, in the case of a new round of training of the wake flow field prediction model, the total loss is back propagated, the neural network parameters are optimized by using the total loss, and the gradient is compensated by using the dimension coefficient.

[0123] In the back propagation process, the gradient needs to be compensated by using the dimension coefficient. Specifically, when calculating the residual term by using automatic differentiation, the scaling factor brought by normalization is introduced into the gradient calculation according to the chain rule (formula 17), so as to ensure the correctness of the optimization direction.

[0124] (17)

[0125] In the formula, is the gradient compensation, is the loss function of the PDE term, is the scaling factor.

[0126] In this way, not only the problem of frequently adjusting the criterion number under multiple working conditions is avoided, but also the consistency of the physical law is maintained by using the gradient compensation mechanism, and the training stability and convergence efficiency under multiple working conditions are significantly improved.

[0127] As can be seen from the above, in order to maximize the training efficiency and accuracy, the original data is subjected to targeted preprocessing. First, a spatially structured down-sampling strategy is adopted, and only 2.5% of the original data is selected as the training set on the premise of retaining the key features of the flow field. Secondly, the Min-Max normalization method is used to eliminate the difference between the physical quantities, and a gradient compensation mechanism is introduced in the back propagation stage in view of the particularity of the PINN physical equation calculation. In this way, the training efficiency of the PI-DANN model is effectively improved, and the simulation accuracy and the generalization ability of the model are guaranteed.

[0128] Optionally, as shown in Figure 6 The embodiment of the disclosure provides another neural network-based prediction method for a multi-condition flow field of a wind turbine, which comprises the following steps:

[0129] S101, Obtain the sample dataset, which includes multiple sets of multi-condition data.

[0130] S105 uses the maximum and minimum value normalization method to linearly map the original physical quantities in the sample dataset to the [0,1] interval.

[0131] S106, input the sample dataset into the physical information domain adversarial neural network model for training, and obtain the total loss of the sample data in the domain adversarial neural network model for forward propagation. The total loss includes: data loss, physical equation loss and domain classification loss; wherein, when calculating the physical equation residual, the predicted data is restored to the original physical dimensions through inverse normalization.

[0132] S113, if the total loss is less than the loss threshold, the wake field prediction model training is considered complete.

[0133] S123, if the total loss is greater than or equal to the loss threshold, determine the wake field prediction model for a new round of training.

[0134] S107, in the case of a new round of training of the wake field prediction model, backpropagation of the total loss is performed, the total loss is used to optimize the neural network parameters, and the gradient is compensated by the dimensional coefficient.

[0135] Pre-store the loss threshold in memory This is used to determine whether the total loss is small enough. The total loss... With loss threshold Compare. If If so, then the wake field prediction model training is complete. If this is the case, then the wake field prediction model needs to undergo a new round of training.

[0136] Once the wake field prediction model is trained, it can quickly and accurately predict the three-dimensional flow field distribution of the wind turbine wake field for any input operating condition, achieving physically consistent flow field prediction under complex multi-operating conditions.

[0137] By analyzing the two-dimensional flow field profile, the performance of the PI-DANN model in predicting the three-dimensional wake field distribution of wind turbines is visually demonstrated.

[0138] like Figure 7 As shown, it demonstrates in The prediction results of PI-DANN under the operating conditions.

[0139] like Figure 8 As shown, it demonstrates in The prediction results of PI-DANN under the operating conditions.

[0140] Combination Figure 9As shown, the embodiment of the present disclosure provides a first neural network-based wind turbine multi-condition flow field prediction device 90, which comprises an acquisition module 91, a training module 92, a determination module 93 and an optimization module 94. Wherein, the acquisition module 91 is configured to acquire a sample data set, wherein the sample data set comprises multiple sets of multi-condition data; the training module 92 is configured to input the sample data set into a physical information field adversarial neural network model for training to obtain a total loss of the sample data propagated forward in the field adversarial neural network model; wherein the total loss comprises: data loss, physical equation loss and field classification loss; the determination module 93 is configured to determine whether the wake flow field prediction model is trained according to the total loss; the optimization module 94 is configured to, in the case of a new round of training of the wake flow field prediction model, propagate the total loss backward to optimize the neural network parameters.

[0141] By using the first neural network-based wind turbine multi-condition flow field prediction device 90 provided by the embodiment of the present disclosure, the neural network-based wind turbine multi-condition flow field prediction method, the sample data set input into the physical information field adversarial neural network model comprises multiple sets of multi-condition data, compared with inputting a single condition, the generalization ability of the wake flow field prediction model can be improved, and then the multiple influences of the changes of the incoming flow wind speed and the yaw angle on the flow field structure can be fully coped with. Moreover, DANN effectively improves the generalization ability of the model under multiple conditions by introducing the field classification and adversarial training mechanism, so that it can adapt to the changes of complex conditions such as different wind speeds and yaw angles. By calculating the total loss of the sample data propagated forward in the model, it is determined whether the wake flow field prediction model is trained. If a new round of training is performed, the total loss is propagated backward, and the neural network parameters are optimized by using the total loss. In this way, a double-branch loss cooperative optimization mechanism is formed. In this way, the physical information neural network and the field adversarial neural network are combined to fully exert the complementary advantages of the two. In the case of limited data or large distribution difference, high-precision and physically consistent wake flow field prediction can still be realized, which provides a solid theoretical basis and data support for wind turbine arrangement optimization and control strategy.

[0142] In combination Figure 10 As shown, the embodiment of the present disclosure provides a second neural network-based wind turbine multi-condition flow field prediction device 100, which comprises a processor 101 and a memory 102. Optionally, the second neural network-based wind turbine multi-condition flow field prediction device 100 can further comprise a communication interface 103 and a bus 104. Wherein, the processor 101, the communication interface 103 and the memory 102 can complete the communication among each other through the bus 104. The communication interface 103 can be used for information transmission. The processor 101 can call the logical instructions in the memory 102 to execute the neural network-based wind turbine multi-condition flow field prediction method of the above-mentioned embodiment.

[0143] In addition, the logic instructions in the memory 102 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium.

[0144] The memory 102 as a computer readable storage medium can be used to store software programs, computer executable programs, such as program instructions / modules corresponding to the method in the embodiments of the present disclosure. The processor 101 executes the program instructions / modules stored in the memory 102, thereby performing functional applications and data processing, that is, implementing the above-mentioned method for predicting the flow field of the wind turbine in multiple working conditions based on the neural network.

[0145] The memory 102 can include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required by at least one function; the data storage area can store data created according to the use of the terminal device, etc. In addition, the memory 102 can include a high-speed random access memory, and can also include a non-volatile memory.

[0146] In combination with Figure 11 and Figure 12 As shown in the above-mentioned embodiments, the present disclosure provides an electronic device 110, which includes an electronic device body and the first prediction device 90 or the second prediction device 100 for predicting the flow field of the wind turbine in multiple working conditions based on the neural network. The first prediction device 90 or the second prediction device 100 for predicting the flow field of the wind turbine in multiple working conditions based on the neural network is installed on the electronic device body. The installation relationship described herein is not limited to placing in the inside of the product body, but also includes the installation connection with other components of the electronic device 110, including but not limited to physical connection, electrical connection or signal transmission connection, etc. Those skilled in the art can understand that the first prediction device 90 or the second prediction device 100 for predicting the flow field of the wind turbine in multiple working conditions based on the neural network can be adapted to the feasible product body, and thus other feasible embodiments can be realized.

[0147] The present disclosure provides a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are set to execute the above-mentioned method for predicting the flow field of the wind turbine in multiple working conditions based on the neural network.

[0148] The technical solutions of the embodiments of the present disclosure can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes one or more instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present disclosure. The aforementioned storage medium can be a non-transitory storage medium, such as a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc. various media that can store program codes.

[0149] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0150] Although the specific embodiments of the present application are described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A method for predicting the flow field of a wind turbine under multiple operating conditions based on a neural network, characterized in that, include: Step 1, Sample Data Acquisition: Acquire a sample dataset, wherein the sample dataset includes multiple sets of multi-condition data; Step 2, Model Training: Input the sample dataset into the physical information domain adversarial neural network model for training, and obtain the total loss of the sample data in the forward propagation of the domain adversarial neural network model; wherein, the total loss includes: data loss, physical equation loss and domain classification loss; the total loss is determined in the following way: The mean square error between the predicted and actual values ​​of the three-dimensional air velocity vector for a single sample is calculated, and the sum of the mean square errors of all samples is determined as the data loss. Calculate the residuals of the NS equations and the residuals of the vortex equations for each spatial coordinate vector. The sum of the residuals of the NS equations for each spatial coordinate vector is determined as the NS equation loss, and the sum of the residuals of the vortex equations for each spatial coordinate vector is determined as the vortex equation loss. The sum of the NS equation loss and the vortex equation loss is determined as the physical equation loss. The classification probability error of the model to the domain of the sample data is calculated. The cross-entropy of the three-dimensional prediction velocity vector in the domain classification task is used to determine the single sample loss. The sum of the domain classification cross-entropies of all samples is determined as the domain classification loss. Step 3, Model Training Completion Confirmation: Based on the total loss, determine whether the wake field prediction model has been trained successfully; Step 4, Neural Network Parameter Optimization: After a new round of training on the wake field prediction model, the total loss is backpropagated to optimize the neural network parameters.

2. The method for predicting the multi-condition flow field of a wind turbine based on a neural network according to claim 1, characterized in that, The optimized neural network parameters include: Calculate the gradient of the data loss, the gradient of the NS equation loss, the gradient of the vortex equation loss, and the gradient of the domain classification loss for all neural network parameters respectively. The gradient of the loss of the NS equation is adjusted using the weights of the loss coefficients of the NS equation, the gradient of the loss of the vortex equation is adjusted using the weights of the loss coefficients of the vortex equation, and the gradient of the loss of the domain classification is adjusted using the weights of the loss coefficients of the domain classification. Based on the neural network parameters, learning rate, gradient of the data loss, gradient of the adjusted NS equation loss, gradient of the adjusted vortex equation loss, and gradient of the adjusted domain classification loss corresponding to this iteration, determine the neural network parameters corresponding to the next iteration.

3. The method for predicting the multi-condition flow field of a wind turbine based on a neural network according to claim 1, characterized in that, The method further includes: Before the sample dataset is input, a maximum-minimum normalization method is used to linearly map the original physical quantities in the sample dataset to the interval [0, 1]. When calculating the residuals of the physical equations, the predicted data are restored to the original physical dimensions through inverse normalization.

4. The method for predicting the multi-condition flow field of a wind turbine based on a neural network according to claim 3, characterized in that, When backpropagating the total loss, the method further includes: Calculate the normalized variable based on the original physical quantity, the minimum value of the original physical quantity, and the maximum value of the original physical quantity; Calculate the scaling factor based on the minimum value of the original physical quantity and the maximum value of the original data physical quantity; Gradient compensation is calculated based on the loss function of the physical equation terms, the normalization variable, the scaling factor, and the neural network parameters corresponding to the current iteration, so as to compensate for the gradient through dimensional coefficients.

5. The method for predicting the multi-condition flow field of a wind turbine based on a neural network according to claim 1, characterized in that, The method further includes: During model training, the adversarial weight coefficients are determined based on the training progress: in, This represents the adversarial weight coefficient. Indicates training progress. , Control the rate at which the weights increase; The contribution ratio of the domain classification loss to the adversarial weight coefficient is adjusted according to the adversarial weight coefficient.

6. The method for predicting the multi-condition flow field of a wind turbine based on a neural network according to claim 2, characterized in that, The weights of the loss coefficients in the physical equations are dynamically adjusted in the following ways: In the In the first iteration, the estimated values ​​of the weights of the loss coefficients of each physical equation are obtained through the second iteration. The ratio of the data loss gradient of the round to the loss gradient of the corresponding physical equation is dynamically estimated; Based on the estimated weights of the loss coefficients of each physical equation, a weighted smoothing strategy is introduced to obtain... The weight of the loss coefficient of the physical equation corresponding to the round of iteration.

7. A device for predicting the multi-condition flow field of a wind turbine based on a neural network, characterized in that, include: The acquisition module is configured to acquire a sample dataset, wherein the sample dataset includes multiple sets of multi-condition data; The training module is configured to input the sample dataset into a physical information domain adversarial neural network model for training, and obtain the total loss of the sample data propagating forward in the domain adversarial neural network model; wherein, the total loss includes: data loss, physical equation loss, and domain classification loss; the total loss is determined in the following manner: The mean square error between the predicted and actual values ​​of the three-dimensional air velocity vector for a single sample is calculated, and the sum of the mean square errors of all samples is determined as the data loss. Calculate the residuals of the NS equations and the residuals of the vortex equations for each spatial coordinate vector. The sum of the residuals of the NS equations for each spatial coordinate vector is determined as the NS equation loss, and the sum of the residuals of the vortex equations for each spatial coordinate vector is determined as the vortex equation loss. The sum of the NS equation loss and the vortex equation loss is determined as the physical equation loss. The classification probability error of the model to the domain of the sample data is calculated. The cross-entropy of the three-dimensional prediction velocity vector in the domain classification task is used to determine the single sample loss. The sum of the domain classification cross-entropies of all samples is determined as the domain classification loss. The determination module is configured to determine whether the wake field prediction model has been trained successfully based on the total loss. The optimization module is configured to backpropagate the total loss and optimize the neural network parameters during a new round of training of the wake field prediction model.

8. A prediction device for multi-condition flow field of a wind turbine based on a neural network, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to, when running the program instructions, execute the neural network-based method for predicting the multi-condition flow field of a wind turbine as described in any one of claims 1 to 6.

Images