Wind farm wake control optimization method, device, equipment and storage medium

Abstract

The application discloses a wind farm wake control optimization method, device, equipment and storage medium. The method comprises the following steps: simulating wind farm data by using a preset simulation engine to obtain a simulation data set; training a preset convolutional neural network based on the simulation data set until the preset convolutional neural network reaches a preset convergence condition to obtain a single-wind-turbine wake distribution model, thereby establishing a digital twin system, reducing simulation cost and improving simulation efficiency; utilizing the wake distribution model to make control decisions for a wind farm group according to real-time wind farm data, generating an optimized control strategy, and performing wake control on the wind farm group based on the optimized control strategy, so that the digital twin system adjusts wind turbine parameters according to the wake distribution, and the power generation efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power generation, and in particular to a wind farm wake control optimization method, device, equipment and storage medium. BACKGROUND

[0002] A large-scale wind farm is often composed of dozens or even hundreds of wind turbines in parallel. Due to the wind energy absorbed by the wind turbine at the upwind, a wake area with reduced wind speed is generated behind the wind turbine. The wake area develops to the downstream wind turbine, which reduces the incoming wind speed of the downstream wind turbine, thereby affecting the power generation efficiency of the downstream wind turbine. Therefore, it is necessary to control the wake effect of the wind farm to improve the overall wind receiving condition of the upstream and downstream wind turbines.

[0003] At present, the wake simulation analysis based on the traditional wake engineering model has a large number of approximate calculations and false calculations, resulting in a large error in the analysis result. The wake simulation analysis based on fluid mechanics can obtain a relatively accurate simulation analysis result, but high-precision simulation divides the wind farm into a large number of grid spaces for numerical calculation. If the simulation analysis result is directly applied to the traditional optimization control algorithm, it will result in high simulation cost and low simulation efficiency. SUMMARY

[0004] The present application provides a wind farm wake control optimization method, device, equipment and storage medium to solve the technical problems of high cost and low efficiency of wind farm wake simulation analysis.

[0005] To solve the above technical problems, in a first aspect, the present application provides a wind farm wake control optimization method, comprising:

[0006] using a preset simulation engine to simulate wind farm data to obtain a simulation data set;

[0007] training a preset convolutional neural network based on the simulation data set until the preset convolutional neural network reaches a preset convergence condition to obtain a single wind turbine wake distribution model;

[0008] using the wake distribution model to make control decisions for a wind farm group according to real-time wind farm data to generate an optimization control strategy;

[0009] controlling the wind farm group based on the optimization control strategy.

[0010] In some implementations, the simulation of wind farm data using a preset simulation engine to obtain a simulation data set comprises:

[0011] using the preset simulation engine to perform atmospheric boundary layer calculation and wind turbine wake calculation according to the wind farm data to output the simulation data set.

[0012] In some implementations, the utilizing the preset simulation engine, performing atmospheric boundary layer calculation and wind turbine wake calculation according to the wind farm data, and outputting the simulation data set, comprises:

[0013] Utilizing the preset simulation engine, simulating the initial non-physical state of the wind farm into a physical state with fully developed turbulence according to the wind farm data to obtain wind farm simulation data;

[0014] Utilizing the actuator line model or the disc model, iterating the wind farm simulation data to output the simulation data set.

[0015] In some implementations, before the utilizing the preset simulation engine, simulating the wind farm data to obtain the simulation data set, further comprising:

[0016] Based on Dakota, a preset simulation driver is built;

[0017] Based on a preset driver optimization algorithm, the preset simulation driver is optimized according to a preset initial training set to obtain the preset simulation engine, and the preset driver optimization algorithm comprises at least one of a gradient optimization algorithm, a non-gradient optimization algorithm, and a non-gradient full set optimization algorithm.

[0018] In some implementations, before the training the preset convolutional neural network based on the simulation data set until the preset convolutional neural network reaches a preset convergence condition to obtain the wake distribution model of a single wind turbine, comprising:

[0019] Based on the encoder and the decoder trained in the preset convolutional neural network, a characteristic function of the simulation data set is determined;

[0020] The characteristic function is mapped to a hidden vector space to generate linear evolution characteristics under the influence of a preset control parameter;

[0021] The linear evolution characteristics are combined with the wind farm data in the simulation data set to train a feature embedding;

[0022] The feature embedding is mapped to a linear vector space required by the encoder to generate the wake distribution model.

[0023] In some implementations, before the training the preset convolutional neural network based on the simulation data set until the preset convolutional neural network reaches a preset convergence condition to obtain the wake distribution model of a single wind turbine, further comprising:

[0024] Based on a secure shell protocol, a communication channel between the simulation data set and the preset simulation engine is established;

[0025] packaging and encapsulating the simulation data set;

[0026] transmitting the encapsulated simulation data set to the preset convolutional neural network based on the communication channel.

[0027] In some implementations, the tail flow distribution model is used to make control decisions for a wind farm cluster based on real-time wind farm data, and an optimized control strategy is generated.

[0028] obtaining real-time wind farm data;

[0029] inputting the real-time wind farm data into the tail flow distribution model, and outputting the optimized control strategy.

[0030] In a second aspect, the present application provides a wind farm tail flow control optimization method, comprising:

[0031] a simulation module configured to simulate wind farm data by using a preset simulation engine to obtain a simulation data set;

[0032] a training module configured to train a preset convolutional neural network based on the simulation data set until the preset convolutional neural network reaches a preset convergence condition, and obtain a tail flow distribution model of a single wind turbine;

[0033] a generation module configured to use the tail flow distribution model to make control decisions for a wind farm cluster based on real-time wind farm data, and generate an optimized control strategy;

[0034] a control module configured to control the tail flow of the wind farm cluster based on the optimized control strategy.

[0035] In a third aspect, the present application further provides a computer device comprising a processor and a memory, wherein the memory is configured to store a computer program, and the computer program is configured to be executed by the processor to implement the wind farm tail flow control optimization method according to the first aspect.

[0036] In a fourth aspect, the present application provides a computer readable storage medium storing a computer program, and the computer program is configured to be executed by a processor to implement the wind farm tail flow control optimization method according to the first aspect.

[0037] Compared with the prior art, the present application has at least the following beneficial effects:

[0038] By utilizing a preset simulation engine, wind farm data is simulated to obtain a simulation dataset, and based on the simulation dataset, a preset convolutional neural network is trained until the preset convolutional neural network reaches a preset convergence condition, thereby obtaining a wake flow distribution model of a single wind turbine, so as to establish a digital twin system, thereby reducing simulation cost and improving simulation efficiency; then, by utilizing the wake flow distribution model, real-time wind farm data is used to make control decisions for a wind farm cluster, thereby generating an optimized control strategy, and based on the optimized control strategy, the wind farm cluster is subjected to wake flow control, so as to use the digital twin system to adjust wind turbine parameters according to the wake flow distribution, thereby improving power generation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A flowchart of a wind farm wake flow control optimization method according to an embodiment of the present application is shown.

[0040] Figure 2 A running mechanism diagram of Dakota according to an embodiment of the present application is shown.

[0041] Figure 3 A wake flow characteristic diagram according to an embodiment of the present application is shown.

[0042] Figure 4 A structural diagram of a wind farm wake flow control optimization apparatus according to an embodiment of the present application is shown.

[0043] Figure 5 A structural diagram of a computer device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0045] Please refer to Figure 1 , Figure 1 A flowchart of a wind farm wake flow control optimization method according to an embodiment of the present application is shown. The wind farm wake flow control optimization method according to the embodiment of the present application can be applied to a computer device, which includes but is not limited to a smart phone, a notebook computer, a tablet computer, a desktop computer, a physical server, a cloud server, and the like. As shown in Figure 1 , the wind farm wake flow control optimization method according to the embodiment includes steps S101 to S104, which are described in detail as follows.

[0046] In step S101, a preset simulation engine is used to simulate wind farm data to obtain a simulation data set.

[0047] In this step, the wind farm data includes, but is not limited to, the incoming flow speed and deflection angle of the wind turbine, etc. The preset simulation engine can be SOWFA.

[0048] In some embodiments, the step S101 comprises:

[0049] The preset simulation engine is used to perform atmospheric boundary layer calculation and wind turbine wake calculation according to the wind farm data, and output the simulation data set.

[0050] In this embodiment, the atmospheric boundary calculation and the wind turbine wake calculation are all preprocessed, simulated and post-processed to output the simulation data set. This embodiment is based on SOWFA to perform wind farm numerical simulation, and uses a numerical simulation method to obtain the wind field evolution of the wind farm, which helps to further understand the wind farm wake action mechanism and improve the calculation accuracy of the wake optimization problem. SOWFA is a wind farm numerical simulation engine developed based on the OpenFoam computational fluid dynamics library. Compared with various engineering models and simulation models commonly used in wake analysis, SOWFA performs more detailed physical modeling of the wind farm. SOWFA inputs specific environmental information and wind turbine information of the wind farm, performs a large number of grid calculations through physical methods, and thus more accurately obtains the specific wind distribution of the entire wind field over time.

[0051] Optionally, the simulation engine is used to perform atmospheric boundary layer calculation and wind turbine wake calculation according to the wind farm data, and output the simulation data set, comprising: using the preset simulation engine to simulate the initial non-physical state of the wind farm to a physical state with fully developed turbulence according to the wind farm data, to obtain wind farm simulation data; using an actuator line model or a disc model to iterate the wind farm simulation data, and output the simulation data set.

[0052] In this optional embodiment, the simulation calculation of SOWFA includes ABL atmospheric boundary layer calculation and wind turbine wake calculation, wherein the wind turbine wake model calculation can select two models of ALM (actuator line model) and ADM (disc model), wherein ALM is more detailed in modeling the wind turbine blades, and thus has a higher requirement for the simulation time step.

[0053] The SOWFA program needs to be compiled and built from source code, and in this embodiment, the SOWFA program is compiled and stored in a Docker container to facilitate management, upgrading and migration. In order to facilitate access and automation, an ssh secure shell protocol service is simultaneously running in the container, and port forwarding is performed through the server to enable external access. Since a large amount of disk space is required during the calculation of the example, the example directory in the container needs to be mapped to point to the example directory on the server data disk.

[0054] SOWFA performs numerical calculation on a per-example basis. An example folder contains multiple simulation description files and script files that specify specific information such as simulation parameters, simulation domain parameters, wind field environment parameters, and wind turbine parameters. The complete execution of an example is divided into multiple steps, and the corresponding command line scripts need to be executed in sequence. The execution of an example is divided into ABL atmospheric boundary layer calculation and ALM or ADM wind turbine wake calculation, and each part is further divided into preprocessing, simulation execution and post-processing. The purpose of the ABL atmospheric boundary layer example is to simulate the wind field from the initial non-physical state to the physical state with fully developed turbulence, and the influence of the wind turbine is ignored during the ABL calculation. The ALM or ADM wind turbine wake calculation is based on the simulation results of the ABL wind field, and uses the actuator line model or disc model to model the wake and iteratively affect it over time. After the incoming flow wind flows through the entire wind field for a sufficient time, the entire wind field can see a significant wake effect, and the power of each wind turbine will also change in response.

[0055] The simulation is performed in parallel, and during the simulation, wind field time section data in OPENFOAM format will be saved according to the CPU cores. This data is the simulation result output data, which includes the average wind speed distribution, instantaneous wind speed distribution, temperature distribution and other field information of the entire simulation region. It can be read by Paraview software and further analyzed.

[0056] Taking a single wind turbine as an example, the SOWFA simulation results are used to study the wake distribution under a specific incoming flow direction. The wind turbine height wind field average speed horizontal section is obtained through Paraview. According to the simulation results, as the wind turbine yaw angle deflects, the wind turbine rotor plane is at a certain angle to the incoming flow direction, which causes the wind turbine to mainly absorb the component of the incoming flow along the normal direction of the rotor plane, resulting in changes in the size and direction of the wake.

[0057] In some embodiments, before the simulation data set is obtained by simulating the wind farm data using the preset simulation engine, the method further includes:

[0058] A preset simulation driver is built based on Dakota.

[0059] The preset simulation driver is optimized according to a preset initial training set based on a preset driving optimization algorithm, and the preset simulation driver is obtained, wherein the preset driving optimization algorithm comprises at least one of a gradient optimization algorithm, a non-gradient optimization algorithm and a non-gradient full-set optimization algorithm.

[0060] In the embodiment, after the initial training data is obtained, the preset simulation driver is optimized by one or more combinations of the gradient optimization algorithm, the non-gradient optimization algorithm and the non-gradient full-set optimization algorithm; in the optimization process, the initial training data is processed by the least square method, and the optimized simulation driver is output.

[0061] The simulation driver optimization method aims to minimize the cost or maximize the system performance. The simulation driver optimization algorithm comprises a gradient-based optimization algorithm, a non-gradient information optimization algorithm and a global optimization algorithm. Due to the differences in the characteristics of different target problems and the analysis capabilities of the established simulation model (for example, the accuracy and uncertainty of the simulation results, and whether further information such as gradient can be provided), the most suitable optimization method needs to be selected according to the actual situation. In addition, sometimes multiple optimization analysis algorithms are combined to improve the optimization solving effect.

[0062] In the embodiment, the gradient optimization algorithm is most suitable for effectively navigating to the local minimum value near the initial point based on the gradient-based optimizer. Common algorithms include the conjugate gradient method and the Newton method. These methods are not suitable for finding the global optimum in a non-convex design space. The gradient-based optimization method is an efficient method with the best convergence speed among all local optimization methods, and is the method of choice for problems that are smooth, unimodal and well-behaved. However, when the problem exhibits non-smooth, discontinuous or multi-modal behavior, these methods can be the least robust. Gradient accuracy is a key factor for gradient-based optimizers, because inaccurate derivatives often lead to search failures or premature termination of the method. In an ideal case, analytical gradient information can be obtained, thus achieving the fastest convergence speed and relatively high accuracy. If an analytical gradient and Hessian information can be provided by a simulation program, using the full Newton method will achieve a quadratic convergence speed close to the solution. However, in actual engineering, it is often difficult to obtain an analytical gradient, at which time a finite difference method can be used to approximate the gradient, so as to solve by using the gradient optimization method. However, in order to ensure that reliable gradient information can be obtained, the smoothness of the influence of the design parameters on the simulation results of the optimization target should be understood by the method of parameter analysis first. After determining the effectiveness of the difference approximation method, a suitable approximation step size needs to be selected.

[0063] Gradient-free optimization algorithms can be applied in situations where gradient computation is too expensive or unreliable compared to gradient-based methods. In addition, some derivative-free methods can be used for global optimization. For these reasons, derivative-free methods are often the method of choice when the problem can be non-smooth, multi-modal, or otherwise poorly behaved. However, they are much slower to converge when seeking an optimal solution, and thus tend to require more computation than gradient-based methods. Depending on the number of variables, they typically require several hundred to several thousand or more function evaluations for local methods, and tens of thousands to hundreds of thousands or more for global methods. Given the computational cost, it is often prudent to use a derivative-free method to identify a region of interest, and then use a gradient-based method to locate the solution. Commonly used gradient-free optimization algorithms include pattern search, simplex method, heuristic greedy search algorithms.

[0064] Gradient-free global optimization algorithms can be used to find a global optimal solution. Commonly used algorithms include evolutionary algorithms, particle swarm algorithms, and efficient global optimization (EGO) methods.

[0065] This embodiment is based on the simulation driver built by Dakota, which is an open source, cross-platform project that provides relatively complete support for various methods of simulation driver optimization. In subsequent versions, it provides other simulation-based analysis functions in addition to optimization, including global sensitivity and variance analysis, parameter estimation, uncertainty quantification and verification methods, and more advanced optimization methods such as surrogate optimization, hybrid optimization, and optimization under uncertain conditions.

[0066] To obtain the Dakota application, you can either build it from source code or directly download the compiled file. The official Dakota release versions for Windows and Linux platforms differ slightly in functionality due to different parameter settings when executing Cmake (a cross-platform installation (compilation) tool that describes the installation (compilation process) on all platforms). For example, the Windows release version is missing support for the Python surrogate model toolkit.

[0067] The core functionality of Dakota is contained in the command-line executable program. In addition, Dakota provides an IDE for visual operations, which is based on eclipse and relies on the command-line program. In addition, Dakota provides command-line tools and Python tool libraries to wrap user simulation programs so that they can be used in Dakota. For surrogate models, Dakota also provides a Python library that can be used independently.

[0068] A key advantage of Dakota is the wide range of iterative capabilities available through a relatively simple interface with the simulator. Trying different iterative methods often requires only a few changes to the Dakota text input file, and then starting a new analysis, without the need to understand the underlying software packages integrated in Dakota. In addition, Dakota will manage the concurrent execution of the computational model in parallel.

[0069] Figure 2 An illustration of the Dakota execution mechanism is shown. The Dakota iterative algorithm and the user's simulation code are loosely coupled, a coupling often referred to as "black box" because Dakota does not know (or knows very little about) the internal details of the computational model, and therefore does not require any access to its source code. Dakota and the simulation code exchange data by reading and writing short data files. Dakota is executed through commands provided by the user in a text input file, which specifies the type of analysis to be performed (e.g., parameter study, optimization, uncertainty quantification), as well as the file names associated with the user's simulation code. During operation, Dakota automatically executes the user's simulation code by creating a separate process outside of Dakota. When Dakota runs, it gives a parameter file containing the current variable values. Dakota then launches the user's simulation code (or a script that encapsulates it), and when the simulation is complete, reads the response data from the results file. This process is repeated until all simulations required by the iterative study are complete, and Dakota outputs the results of the analysis.

[0070] The behavior of a Dakota project is defined by the Dakota input file, which is typically composed of six sections: variables, interfaces, responses, models, methods, and environment. While these keyword blocks can appear in any order in the Dakota input file, there is an inherent relationship between them, so that at each iteration of the algorithm, a method block requests a mapping of variables to responses, which is implemented by a model through an interface that will connect Dakota and the user's simulation program. For more complex models, these sections can appear multiple times in the same Dakota input file to define nested or associated model relationships (e.g., building a surrogate model on top of a simulation model, and performing optimization using the surrogate model). The interaction file between Dakota and the user's simulation program is governed by specific format conventions, which can be met by quickly wrapping the simulation program with the toolkits or Python libraries provided by Dakota.

[0071] In step S102, the preset convolutional neural network is trained based on the simulation dataset until the preset convolutional neural network reaches a preset convergence condition, and a wake flow distribution model of a single fan is obtained.

[0072] In this step, the preset convergence condition can be that the number of training times reaches a preset number.

[0073] In some embodiments, the step S102 comprises:

[0074] Based on the trained encoder and decoder in the preset convolutional neural network, a characteristic function of the simulation data set is determined;

[0075] The characteristic function is mapped to a hidden vector space, and linear evolution characteristics are generated under the influence of a preset control parameter;

[0076] The linear evolution characteristics are combined with wind farm data in the simulation data set to train a feature embedder;

[0077] The feature embedder is mapped to a linear vector space required by the encoder to generate the wake flow distribution model.

[0078] In this embodiment, the wake flow distribution model is a proxy model, which is a high-precision and high-performance approximation of a high-fidelity and low-performance simulation model (or "real" model), aiming to capture the significant features of the expensive high-fidelity model. They can be used to explore the changes in the parameter space region response quantity, usually with lower computational cost. In the context of optimization problems, proxy models can speed up convergence by reducing function evaluation costs or smoothing noisy response functions.

[0079] Optionally, the proxy model includes but is not limited to data fitting based proxy model, multi-fidelity proxy model and reduced order model proxy. Among them, the data fitting method includes using the data (target value, gradient, etc.) generated by the original real model to construct an approximate model. Common numerical fitting methods include Taylor series expansion, polynomial regression, Gaussian process regression or Kriging interpolation, artificial neural network, multivariate adaptive regression spline, radial basis method. The multi-fidelity proxy model uses a low-fidelity physical model as a substitute for the original high-fidelity model. The low-fidelity model usually includes a coarser grid, a looser convergence tolerance, and a lower unit order. It is an independent model itself and does not need to obtain data from the high-fidelity model for construction, while the high-fidelity model is used to check the parameters of the low-fidelity model. The reduced order model is derived mathematically from the high-fidelity model by Galerkin projection technology. By calculating a set of basis functions to capture the main dynamics of the system, the original high-order system can be projected to a smaller system.

[0080] It should be noted that the CFD simulation data based on SOWFA has high computational cost for direct application to optimization, however, a large number of simulation results under different conditions can be generated by pre-computation for model training, realizing the construction of a high-precision proxy model based on data driving.

[0081] In some embodiments, the step S102 further comprises, before the step S102:

[0082] establishing a communication channel between the simulation dataset and the preset simulation engine based on a secure shell protocol;

[0083] packing and encapsulating the simulation dataset;

[0084] transmitting the encapsulated simulation dataset to the preset convolutional neural network based on the communication channel.

[0085] In the embodiment, the complete execution process of the SOWFA example is relatively cumbersome, which needs to copy the example file, modify the example parameters, execute the multi-step calculation instructions in sequence and wait for the execution to be completed. After the execution is completed, the parallel stored result file needs to be packaged and downloaded to the local for viewing and analyzing. The execution of the optimization process needs to perform batch execution of the example to obtain sufficient sample data, and therefore the above links need to be automatically encapsulated, and the interface connected with the optimization analysis engine is developed. The embodiment realizes the automatic data interface program for automatic example generation, execution and post-processing based on Python.

[0086] The embodiment sets a basic communication module based on the paramiko library, so as to communicate with the SOWFA software running in the Docker container in the server through the secure shell protocol (SSH). Based on the basic communication module, each step of executing the SOWFA example is automatically encapsulated, including copying and parameter setting of the example, preprocessing, numerical calculation execution, post-processing and packaging of output data. After the encapsulation, two types of ABLConnecter and ALM / ADMConnecter are formed. Since the execution of the wake flow calculation needs to depend on the result of the atmospheric boundary layer simulation, the instantiation of the latter needs to depend on the instantiation of the former. In the case of a specific incoming flow speed, the yaw angle can be changed to obtain different examples, and therefore one ABLConnecter can be applied to multiple ADMConnecters.

[0087] Optionally, two automatic functions of makeABL and makeALM are set based on the basic interface, so as to maintain the ABL sample library and realize one-key sample generation.

[0088] Optionally, in order to realize the visualized and convenient management of large samples, an example management program based on Java is set, so as to support reading the list of executed examples from the server example directory, and updating the local example library or submitting a new example based on the list. In the embodiment, after the simulation dataset is packaged and pulled to the local, the automatic loading of the data is supported, and the cross-section data tensor file required by the research is generated.

[0089] Step S103, using the wake distribution model, making control decisions for the wind farm group according to real-time wind farm data, and generating an optimized control strategy.

[0090] In this step, the wake distribution model can quickly calculate the wake distribution under different control parameters, so the control parameters can be determined based on the expected wake distribution, thereby realizing control strategy optimization, improving simulation efficiency, and reducing cost.

[0091] In some embodiments, the step S103 comprises:

[0092] Obtaining real-time wind farm data;

[0093] Inputting the real-time wind farm data into the wake distribution model, and outputting the optimized control strategy.

[0094] In this embodiment, the wake characteristics of a single wind turbine under different incoming flow wind speeds and yaw conditions are learned based on a deep neural network of AutoEncoder, and the structure is as shown in Figure 3 The network finds a set of characteristic functions of the original wind farm data (ALM atmospheric boundary layer simulation and ADM wind turbine wake simulation data) by training a pair of Encoder / Decoder, and maps them to a set of linear vector spaces, which have linear evolution characteristics under the influence of control parameters. At the same time, the network trains an embedder for the yaw angle and incoming flow wind speed data as control parameters, which maps the influence to the linear vector space sought by the Encoder, so as to superimpose the influence of the control parameters on the basis of the ABL atmospheric boundary layer conditions, and obtain the target wake distribution. After training, a proxy model is obtained with incoming flow wind speed and yaw angle as control variables and wind turbine near-field wake distribution as output, and the wind farm optimization based on the proxy model is completed.

[0095] Step S104, based on the optimized control strategy, performing wake control on the wind farm group.

[0096] In this step, according to the optimized control strategy, the yaw angle and pitch angle of the wind turbine are adjusted, so as to change the size and direction of the wake, and improve the overall wind receiving condition of each wind turbine in the upstream and downstream.

[0097] It should be noted that this embodiment outputs a simulation data set by simulating the wind farm numerically, completes the training of the convolutional neural network, outputs a proxy model, establishes a digital twin system, reduces the simulation cost, and improves the simulation efficiency. And through the digital twin system, the wind turbine parameters can be adjusted according to the wake distribution, so as to improve the power generation efficiency.

[0098] To perform the wind farm wake control optimization method corresponding to the above-mentioned method embodiment, to realize the corresponding functions and technical effects. Referring to Figure 4 , Figure 4 A structural block diagram of a wind farm wake control optimization device provided by an embodiment of the application is shown. For ease of illustration, only the parts related to the present embodiment are shown. The wind farm wake control optimization device provided by the embodiment of the application comprises:

[0099] The simulation module 401 is configured to simulate the wind farm data by using a preset simulation engine to obtain a simulation data set.

[0100] The training module 402 is configured to train a preset convolutional neural network based on the simulation data set until the preset convolutional neural network reaches a preset convergence condition to obtain a single-wind-turbine wake distribution model.

[0101] The generation module 403 is configured to use the wake distribution model to make control decisions for a wind farm group according to real-time wind farm data to generate an optimized control strategy.

[0102] The control module 404 is configured to control the wind farm group based on the optimized control strategy.

[0103] In some embodiments, the simulation module 401 is configured to:

[0104] The simulation module 401 is configured to use the preset simulation engine to perform atmospheric boundary layer calculation and wind turbine wake calculation according to the wind farm data to output the simulation data set.

[0105] In some embodiments, the simulation module 401 is specifically configured to:

[0106] The simulation module 401 is configured to use the preset simulation engine to simulate the initial non-physical state of the wind farm into a physical state with fully developed turbulence according to the wind farm data to obtain wind farm simulation data.

[0107] The simulation module 401 is configured to use an actuator line model or a disc model to iterate the wind farm simulation data to output the simulation data set.

[0108] In some embodiments, the device further comprises:

[0109] The building module is configured to build a preset simulation driver based on Dakota.

[0110] The optimization module is configured to optimize the preset simulation driver based on a preset driver optimization algorithm according to a preset initial training set to obtain the preset simulation engine. The preset driver optimization algorithm comprises at least one of a gradient optimization algorithm, a gradient-free optimization algorithm, and a gradient-free full-set optimization algorithm.

[0111] In some embodiments, the training module 402 is configured to:

[0112] Based on the encoder and decoder trained in the preset convolutional neural network, the feature function of the simulation dataset is determined;

[0113] The feature function is mapped to the hidden vector space, and linear evolution features are generated under the influence of preset control parameters;

[0114] The linear evolution features are combined with wind farm data from the simulation dataset to train a feature embedder;

[0115] The feature embedder is mapped into the linear vector space required by the encoder to generate the wake distribution model.

[0116] In some embodiments, the apparatus further includes:

[0117] A communication channel is established between the simulation dataset and the preset simulation engine based on the secure enclosure protocol.

[0118] The encapsulation module is used to package and encapsulate the simulation dataset;

[0119] The transmission module is used to transmit the encapsulated simulation dataset to the preset convolutional neural network based on the communication channel.

[0120] In some embodiments, the generation module 403 is configured to:

[0121] Obtain real-time wind farm data;

[0122] The real-time wind farm data is input into the wake distribution model, and the optimized control strategy is output.

[0123] The wind farm wake control optimization device described above can implement the wind farm wake control optimization method of the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0124] Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 5 As shown, the computer device 5 of this embodiment includes: at least one processor 50 ( Figure 5 (Only one is shown) a processor, a memory 51, and a computer program 52 stored in the memory 51 and executable on the at least one processor 50, wherein the processor 50 executes the computer program 52 to implement the steps in any of the above method embodiments.

[0125] The computer device 5 can be a smart phone, a tablet computer, a desktop computer, a cloud server, and the like. The computer device can include, but is not limited to, a processor 50, a memory 51. Those skilled in the art can understand that Figure 5 The computer device 5 is only an example and does not constitute a limitation on the computer device 5, and can include more or fewer components than shown, or combine certain components, or include different components, for example, can also include an input / output device, a network access device, and the like.

[0126] The processor 50 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or can also be any conventional processor.

[0127] The memory 51 can be an internal storage unit of the computer device 5 in some embodiments, for example, a hard disk or a memory of the computer device 5. The memory 51 can also be an external storage device of the computer device 5 in other embodiments, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like. Further, the memory 51 can include both the internal storage unit and the external storage device of the computer device 5. The memory 51 is used to store an operating system, an application program, a boot loader, data, and other programs, for example, program codes of the computer program, and the like. The memory 51 can also be used to temporarily store data that has been output or will be output.

[0128] In addition, the embodiment of the present application further provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps in any method embodiment described above.

[0129] The embodiment of the present application provides a computer program product, when the computer program product is run on a computer device, so that the computer device executes the steps in each method embodiment described above.

[0130] In several embodiments provided in the present application, it can be understood that each block in the flowchart or block diagram can represent a module, a segment or a portion of code which includes one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figure. For example, two blocks noted in succession can in fact be executed substantially concurrently or in the reverse order, depending on the functionality involved.

[0131] If the functions are realized in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the parts of the technical solutions which essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device to perform all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various other media capable of storing program codes.

[0132] The specific embodiments described above further illustrate the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above description is only for specific embodiments of the present application and is not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A wind farm wake control optimization method, characterized in that, include: Using a pre-set simulation engine, wind farm data is simulated to obtain a simulation dataset; Based on the simulation dataset, a preset convolutional neural network is trained until it reaches a preset convergence condition, resulting in a wake distribution model for a single wind turbine. Specifically, based on the encoder and decoder trained in the preset convolutional neural network, feature functions of the simulation dataset are determined. These feature functions are mapped to a hidden vector space, and under the influence of preset control parameters, linear evolution features are generated. These linear evolution features are combined with wind farm data from the simulation dataset to train a feature embedder. The feature embedder is then mapped to the linear vector space required by the encoder to generate the wake distribution model. Using the aforementioned wake distribution model, control decisions are made for the wind farm cluster based on real-time wind farm data, and an optimized control strategy is generated. Based on the optimized control strategy, wake control is performed on the wind farm group.

2. The wind farm wake control optimization method as described in claim 1, characterized in that, The process involves using a pre-set simulation engine to simulate wind farm data, resulting in a simulation dataset, including: Using the preset simulation engine, atmospheric boundary layer calculation and wind turbine wake calculation are performed based on the wind farm data, and the simulation dataset is output.

3. The wind farm wake control optimization method as described in claim 2, characterized in that, The process involves using the preset simulation engine to perform atmospheric boundary layer calculations and wind turbine wake calculations based on the wind farm data, and outputting the simulation dataset, including: Using the preset simulation engine, based on the wind farm data, the initial non-physical state of the wind farm is simulated as a physical state with fully developed turbulence, thus obtaining wind farm simulation data; The simulation data of the wind farm is iterated using an actuation line model or a disk model to output the simulation dataset.

4. The wind farm wake control optimization method as described in claim 1, characterized in that, Before simulating wind farm data using a preset simulation engine to obtain the simulation dataset, the process also includes: Based on Dakota, a pre-defined simulation driver was built; Based on a preset driving optimization algorithm, the preset simulation driver is optimized according to a preset initial training set to obtain the preset simulation engine. The preset driving optimization algorithm includes at least one of gradient optimization algorithm, gradient-free optimization algorithm, and gradient-free full set optimization algorithm.

5. The wind farm wake control optimization method as described in claim 1, characterized in that, Before training the preset convolutional neural network based on the simulation dataset until the preset convolutional neural network reaches the preset convergence condition and obtains the wake distribution model of a single wind turbine, the process further includes: A communication channel is established between the simulation dataset and the preset simulation engine based on the secure enclosure protocol. The simulation dataset is packaged and encapsulated; Based on the communication channel, the encapsulated simulation dataset is transmitted to the preset convolutional neural network.

6. The wind farm wake control optimization method as described in claim 1, characterized in that, The process of using the wake distribution model to make control decisions on the wind farm cluster based on real-time wind farm data and generating optimized control strategies includes: Obtain real-time wind farm data; The real-time wind farm data is input into the wake distribution model, and the optimized control strategy is output.

7. A wind farm wake control optimization method, characterized in that, include: The simulation module is used to simulate wind farm data using a preset simulation engine to obtain a simulation dataset; A training module is used to train a preset convolutional neural network based on the simulation dataset until the preset convolutional neural network reaches a preset convergence condition to obtain a wake distribution model of a single wind turbine. Specifically, based on the encoder and decoder trained in the preset convolutional neural network, the feature functions of the simulation dataset are determined; the feature functions are mapped to a hidden vector space, and linear evolution features are generated under the influence of preset control parameters; the linear evolution features are combined with wind farm data in the simulation dataset to train a feature embedder; and the feature embedder is mapped to the linear vector space required by the encoder to generate the wake distribution model. The generation module is used to make control decisions on the wind farm group based on the wake distribution model and real-time wind farm data, and generate an optimized control strategy. The control module is used to perform wake control on the wind farm group based on the optimized control strategy.

8. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer program, which, when executed by the processor, implements the wind farm wake control optimization method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the wind farm wake control optimization method as described in any one of claims 1 to 6.

Images