A wind turbine wake model and wind farm layout optimization method based on machine learning

Abstract

The present invention belongs to the field of wind power generation, and discloses a method for constructing a wind power wake model based on machine learning and optimizing the layout of a wind farm. The method uses a numerical simulation data set to train a machine learning wake model for each sub-model, and constructs a wake model based on machine learning according to the spatial position. Each time, only the sub-model adjacent to the prediction point needs to be called for prediction. Each sub-model is a simple single-layer neural network architecture, thus ensuring the high efficiency of the model. Since the wake model used has an accuracy similar to that of numerical simulation, the layout optimization method using the wake model can have a high accuracy that could not be achieved by the analytical wake model in the past. On the basis of accurately predicting the influence of the wake, the output power of the entire field can be accurately predicted in the optimization of the wind farm layout. Multiple optimizations are performed using multiple random starting points, and then the best one is selected from the multiple local optima obtained, so that the result is as close to the global optimal solution as possible.

Description

Technical Field

[0001] The present invention belongs to the field of wind power generation, and specifically relates to a method for constructing a wind power wake model based on machine learning and a wind farm layout optimization method applying the wake model. Background Art

[0002] The layout design of wind turbines within wind farms primarily utilizes an empirically based array arrangement. Wind turbines are often arranged in fixed rows and columns, which can easily produce severe wake effects when the incoming wind direction aligns with the row and column directions. The wake effect significantly reduces the power generation efficiency of downstream wind turbines, leading to significant cost losses. To this end, some wind farms use longer turbine spacing to minimize the losses caused by the wake effect; however, this further results in lower installed capacity per unit area. In this case, accurate calculation of the wake effect is necessary to design an appropriate wind turbine layout.

[0003] At present, the calculation software that considers the wake effect of wind turbines mainly adopts analytical wake models and numerical simulation methods. Actual large-scale wind farms often contain dozens or hundreds of wind turbines. If the numerical simulation method based on fluid dynamics is used, the accuracy is good, but the calculation cost is high and it cannot meet the efficiency requirements of optimization problems. The analytical wake models based on analytical formulas, such as the Jensen model (also known as the Park model) and the Gaussian model, have high calculation efficiency because they contain many assumptions that simplify the problem and can meet the efficiency requirements of optimization problems, but they lack accuracy and are not as precise as numerical simulation methods. For example, invention patent application CN109992889A uses the Jensen analytical wake model as a basis to quickly predict the wake velocity at a fixed point, but its numerical simulation accuracy is low and does not include turbulence field prediction, so it is only applicable to specific working conditions.

[0004] Currently, existing wake prediction technologies lack a method that can achieve both numerical simulation accuracy and analytical model efficiency. In particular, in the problem of wind farm layout optimization, a more precise method is urgently needed to improve the accuracy of wake prediction. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a machine learning based wake model (Machine Learning Wake Model) and a layout optimization (Wind Farm Layout Optimization) method using the wake model. The wake model is an artificial neural network (Artificial Neural Network) model constructed using an artificial intelligence (Artificial Intelligence) method. A numerical simulation data set (CFD simulation database) is used as a training data set to train a machine learning wake model for each sub-model (Sub-model), which is numbered and summarized in order of spatial position to construct a wake model based on machine learning, which can achieve high-precision and high-efficiency wake prediction at the same time; the present invention also provides a layout optimization method, in which a wake model based on machine learning is used to solve the wake effect, which can have high precision that could not be achieved by the analytical wake model in the past, and multiple random starting points (Multi-start Method) are used for multiple optimizations to find the best from multiple local optimal solutions, which can be as close to the global optimal solution as possible, and is effectively applicable to complex optimization problems with high nonlinearity and high variable dimensions such as layout optimization.

[0006] The method for constructing a wind turbine wake model based on machine learning described in the present invention comprises the following steps:

[0007] Step 1: Conduct fluid dynamics numerical simulation modeling for the wind turbine;

[0008] Step 2: Based on the atmospheric boundary layer theory and the inflow conditions of wind turbine operation, determine the inflow conditions included in the single wind turbine wake database;

[0009] Step 3: Perform numerical calculations on each selected inflow condition and save the numerical simulation results of the computational domain as raw data files in a single wind turbine wake database;

[0010] Step 4: interpolate and extract the original data files in the single wind turbine wake database so that the size of each data file is fixed, thereby making the size of the obtained original output data set controllable;

[0011] Step 5: Divide all output nodes into several sub-models according to the output node positions in the width and height planes of the computational domain. If there are n output nodes in the width direction and m output nodes in the height direction, then there are n×m sub-models in total. Divide the original output data into the corresponding sub-model output data according to the division method.

[0012] Step 6. Build an artificial neural network model, consisting of a two-variable input layer, a hidden layer, and an output layer with the same number of output nodes as the length direction. The two-variable input layer simplifies the input dimension as much as possible, so the number of working conditions included in the database is not excessive, and the time consumption for generating the dataset is within an acceptable range. The activation function of the hidden layer is tanh. Its single-layer structure makes the model as lightweight as possible, thus ensuring high efficiency when the model is subsequently called for prediction. The activation function of the output layer is relu.

[0013] Step 7: Using different sub-models to predict different spatial positions ensures the high accuracy of the entire model. Therefore, each sub-model is trained with an artificial neural network model to obtain a machine learning wake model for each sub-model.

[0014] Step 8: The machine learning wake models of all sub-models are numbered and aggregated in order of spatial position to form a complete machine learning wake model of the wind turbine.

[0015] Furthermore, in step 1, the k-epsilon turbulence model is combined with the ADM-R model to perform fluid dynamics numerical simulation modeling of the wind turbine. The length × width × height dimensions of the calculation domain are greater than or equal to 28D × 12D × 4D, where D is the rotor diameter of the wind turbine.

[0016] Furthermore, in step 2, the inflow conditions are uniformly selected within the entire operating inflow condition range, and the wind speed and turbulence at the hub height are determined for each inflow condition.

[0017] Furthermore, in step 4, a fixed interval is determined in the calculation domain, and the output node is determined based on the interval; each output node is interpolated to obtain the data of the output node, and the data under different working conditions are concatenated in a file as the original output data, and the wind speed and turbulence at the hub height of the working condition are recorded in sequence in a separate file as input data.

[0018] Furthermore, the output data of each sub-model and the input data of the working condition sequence are read in, the training number threshold is set, the artificial neural network model is trained, and the model with the highest accuracy is recorded as the final machine learning wake model of each sub-model.

[0019] A wind farm layout optimization method uses the wake model constructed by the above method and combines it with a multi-starting point optimization method to solve the layout optimization problem and approach the global optimal solution. The steps are as follows:

[0020] Step 9: Determine the inflow conditions that need to be calculated based on the inflow conditions of the wind farm, including wind speed, turbulence, and direction;

[0021] Step 10: Provide a random layout plan, determine the upstream and downstream relationship of the wind turbines according to each inflow condition, and use the machine learning wake model to solve the wake effect of the wind turbine from the upstreammost wind turbine in sequence;

[0022] Step 11: For the downstream wind turbines, the wind speed and turbulence are superimposed using the corresponding wake superposition model; the wind speed and turbulence under the combined influence of the wake effects of multiple wind turbines are obtained as the inflow conditions for the downstream wind turbines;

[0023] Step 12: Determine the power generation capacity of the wind turbine generator set based on its power-wind speed curve and the inflow conditions of the wind turbine generator set;

[0024] Step 13: Add the generated power of all wind turbines to obtain the initial power output of the wind farm;

[0025] Step 14: Calculate the next possible optimized layout scheme using the optimization algorithm, and calculate the total power output of the wind farm under the layout scheme; if the total power output is higher than the initial total power output, record the layout scheme as the optimized solution;

[0026] Step 15: Repeat step 14 until the convergence condition of the optimization problem is met, then stop the optimization and take the final optimization solution as the local optimal solution of the optimization problem;

[0027] Step 16: Randomize an initial distribution again and repeat steps 10-15 until the set total number of randomizations is reached to obtain multiple local optimal solutions;

[0028] Step 17: The best solution among multiple local optimal solutions is used as the optimal layout.

[0029] The beneficial effects described in the present invention are as follows: the present invention uses a numerical simulation data set as a training data set to train a machine learning wake model for each sub-model, and then numbers and summarizes them in order of spatial position to construct a wake model based on machine learning. It is only necessary to record the spatial position, and each time only four adjacent sub-models around the prediction point need to be called for prediction. The neural network architecture of each sub-model is a simple single-layer architecture, thus ensuring the high efficiency of the model. Since the wake model used has an accuracy similar to that of numerical simulation, the layout optimization method using the machine learning wake model can achieve a high accuracy that could not be achieved by the analytical wake model in the past; furthermore, the multi-starting point optimization method used in the layout optimization method can be effectively applied to highly nonlinear optimization problems such as wind farm layout optimization, and the optimal solution can be selected from multiple local optimal solutions obtained by multiple optimizations, which can be as close to the global optimal solution as possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of sub-model division;

[0031] Figure 2 It is a schematic diagram of the artificial neural network structure of the sub-model;

[0032] Figure 3 is a flow chart of a layout optimization method using a wake model based on machine learning;

[0033] Figure 4 This is a flow chart of the wake model construction method based on machine learning;

[0034] Figure 5 This is a display of the training history of a single sub-model in the Horns Rev wind farm case;

[0035] Figure 6 This is a comparison of the error between the trained machine learning wake model and the numerical simulation in the Horns Rev wind farm case.

[0036] Figure 7 In the Horns Rev wind farm case, the layout optimization method was used and the machine learning-based wake model was applied to produce layout and wind speed contour maps before and after optimization.

[0037] Figure 8 This is a comparison of the optimization results using different wake models in the layout optimization method for the Horns Rev wind farm case. DETAILED DESCRIPTION

[0038] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0039] like Figure 4 As shown, the method for constructing a wind turbine wake model based on machine learning described in the present invention takes a typical wind farm Horns Rev as an example, and the steps are as follows:

[0040] Step 1: Use the k-epsilon turbulence model combined with the ADM-R model to perform fluid dynamics numerical simulation modeling of the wind turbine. The computational domain size is greater than or equal to 28D × 12D × 4D (length, width, and height), where D is the rotor diameter of the wind turbine. At the HornsRev wind farm, the wind turbine model is the Vestas V80 2MW, with a rotor diameter of 80m.

[0041] Step 2: Based on the atmospheric boundary layer theory and the inflow conditions of the wind turbine operation, determine the inflow conditions included in the single wind turbine wake database; the inflow conditions are uniformly selected within the entire operating inflow condition range, and the wind speed and turbulence at the hub height are determined for each inflow condition;

[0042] The main inflow operating conditions of the wind turbine model used in the Horns Rev wind farm are 5m / s-20m / s at hub height (70m), with a spacing of 0.5m / s, a turbulence range of 2%-30%, and a spacing of 2%, for a total of 465 operating conditions;

[0043] Step 3: Perform numerical calculations for each selected inflow condition and save the numerical simulation results of the computational domain as raw data files in a single wind turbine wake database, totaling 465 raw output data files;

[0044] Step 4: Use interpolation methods such as three-dimensional linear interpolation of adjacent points to interpolate and extract the raw data files in the single wind turbine wake database; determine a fixed interval within the calculation domain, such as 10 meters, and determine the output node based on this interval; interpolate each output node to obtain the data of the output node, and concatenate the data under different working conditions into a file as the raw output data, and record the wind speed and turbulence at the hub height of the working condition in sequence in a separate file as the input data;

[0045] Step 5: Divide all output nodes into several sub-models according to the output node positions in the width and height planes of the computational domain. If there are n output nodes in the width direction and m output nodes in the height direction, then there are n×m sub-models in total. Divide the original output data into the corresponding sub-model output data according to the division method. Figure 1 As shown in FIG, it is a schematic diagram of sub-model division. When generating a data set, the calculation results of the original computational fluid dynamics numerical simulation are preprocessed according to the sub-model division method;

[0046] In the Horns Rev wind farm case, there are a total of 12×80 / 10+1=97 output nodes in the width direction and 4×80 / 10=32 output nodes in the height direction (not including the ground position). Therefore, a total of 97×32=3104 sub-models need to be trained to predict the results of the entire field. Since the wake characteristics under low wind speed conditions and high wind speed conditions (separated by 10m / s) are quite different, they need to be trained separately, and the velocity field and turbulence field require independent models for prediction. Therefore, the model is divided into four cases: low wind speed segment velocity model, high wind speed segment velocity model, low wind speed segment turbulence model, and high wind speed segment turbulence model. Each case requires training 3104 sub-models, so a total of 12416 sub-models;

[0047] Step 6: Build an artificial neural network model, which includes a two-variable input layer, a hidden layer, and an output layer with the same number of output nodes as the length direction; the activation function of the hidden layer is tanh, and the activation function of the output layer is relu; the activation functions of the hidden layer and the output layer can also be replaced with similar activation functions to achieve similar effects; for example Figure 2 Figure 1 shows the structure of the sub-model artificial neural network. Each sub-model is built using the same architecture as shown in the figure. Since each sub-model represents a different position, each trained sub-model is different. In the Horns Rev wind farm case, the artificial neural network was trained in a Python environment. Each sub-model consists of: an input layer of size 2, a hidden layer of size 10 with a tanh activation function, and an output layer of size 225 with a relu activation function.

[0048] Step 7: Perform artificial neural network model training on each sub-model in turn to obtain a machine learning wake model for each sub-model, totaling 12416;

[0049] like Figure 5 As shown in the figure, the training history of a single sub-model in the Horns Rev wind farm case shows that when the sub-model is trained for more than 30,000 times, its loss function and model accuracy have reached a good level and tend to be stable. Therefore, for this case, the sub-model only needs to be trained for more than 30,000 times.

[0050] Step 8: The machine learning wake models of all sub-models are numbered and aggregated in order of spatial position to form a complete machine learning wake model of the wind turbine.

[0051] like Figure 6 As shown in the figure, in the Horns Rev wind farm case, the error comparison between the trained machine learning wake model and the numerical simulation is shown. The horizontal axis is the numerical simulation calculation result, and the vertical axis is the prediction result of the machine learning wake model. The scatter plot reveals the correlation between the two. By calculating the determination coefficient R 2 , it can be seen that in all cases, the correlation is close to 1.0, which also indicates that the prediction results of the machine learning wake model are very accurate.

[0052] like Figure 3 As shown in FIG, a flowchart of the application of a machine learning-based wind turbine wake model in wind farm layout optimization is shown. The steps for optimizing wind farm layout using the wake model are as follows:

[0053] Step 9: Determine the inflow conditions that need to be calculated based on the inflow conditions of the wind farm, including wind speed, turbulence, and direction;

[0054] Step 10: Normalize the two orthogonal directions x and y of the entire designed wind farm to the interval [0,1]. If the number of designed wind turbines is N, a random array coor = [x1, y1, ...x N ,y N ],make This 2N-length array can represent the coordinate positions of all wind turbines in the wind farm, and also represents a random layout scheme. According to each inflow condition, the upstream and downstream relationships of the wind turbines are determined, and the wake effect is solved using the machine learning wake model in sequence, starting from the upstream wind turbine.

[0055] Step 11: For the downstream wind turbines, the wind speed and turbulence are superimposed using the corresponding wake superposition model; the speed superposition model uses the square sum loss ratio model: Turbulence intensity superposition adopts the turbulent kinetic energy superposition model: Subscript i represents the downstream wind turbine to be calculated, subscript j represents the wind turbine that affects wind turbine i, and subscript inflow represents the background wind field data. The wind speed and turbulence under the combined influence of the wake effect of multiple wind turbines are obtained as the inflow conditions for the downstream wind turbines.

[0056] Step 12: Determine the power generation capacity of the wind turbine generator set based on its power-wind speed curve and the inflow conditions of the wind turbine generator set;

[0057] Step 13: Add the generated power of all wind turbines to obtain the initial power output of the wind farm. In the HornsRev wind farm example, the initial power output is about 35.83 MW at an inflow of 8 m / s at 270 degrees and a turbulence of 7.7%.

[0058] Step 14: Calculate the next possible optimal layout scheme using an optimization algorithm, such as a gradient-based algorithm, such as sequential quadratic programming, interior point method, gradient descent method, active set method, or a heuristic algorithm, such as a genetic algorithm, particle swarm algorithm, annealing algorithm, or ant colony algorithm, and calculate the total power output of the wind farm under the layout scheme; if the total power output is higher than the initial total power output, record the layout scheme as the optimized solution;

[0059] Step 15: Repeat step 14 until the convergence condition of the optimization problem is met, then stop the optimization and take the final optimization solution as the local optimal solution of the optimization problem;

[0060] Step 16: Randomize an initial distribution again and repeat steps 10-15 for about 200 times or more, or optimize the random initial layout as many times as possible within an acceptable computing time, such as 7 days (168 hours), to obtain multiple local optimal solutions;

[0061] Step 17: The best solution among multiple local optimal solutions is used as the optimal layout.

[0062] like Figure 7 As shown in the Horns Rev wind farm case, the layout and wind speed field contour maps before and after optimization demonstrate that by adopting a high-precision machine learning wake model and layout optimization method, the optimized layout enables the downstream wind turbines to avoid the strong wake influence of the upstream wind turbines as much as possible, thereby significantly improving the total power.

[0063] In the Horns Rev wind farm case, based on the same optimization framework, the method described in the present invention can obtain wake velocity and turbulence field prediction results with an efficiency similar to that of the analytical model, with an accuracy error of less than 2% compared with the numerical simulation; its accuracy is about 7% to 15% higher than that of the analytical model; under uniform flow conditions, the optimal solution found by the machine learning wake model described in the present invention (such as Figure 7 The total power output (in the lower middle figure) is 46.14 MW, an increase of 28.77%. The optimal solution found by using Ishihara's analytical model (adopted in FLORIS software) that takes into account the influence of turbulence is 44.72 MW, an increase of 24.81%. Compared with the result of using the machine learning wake model published by this invention, the improvement rate is about 4% lower. Figure 8 .

[0064] The above description is only a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made using the contents of the present invention description and drawings are within the scope of protection of the present invention.

Claims

1. A method for constructing a wind turbine wake model based on machine learning, characterized in that: The steps of the wake model establishment method are: Step 1: Conduct fluid dynamics numerical simulation modeling for the wind turbine; Step 2: Based on the atmospheric boundary layer theory and the inflow conditions of the wind turbine operation, determine the inflow conditions included in the single wind turbine wake database; Step 3: Perform numerical calculations on each selected inflow condition and save the numerical simulation results of the computational domain as raw data files in a single wind turbine wake database; Step 4: interpolate and extract the original data file in the single wind turbine wake database to obtain the original output data; Step 5: Divide all output nodes into several sub-models according to the output node positions in the width and height planes of the computational domain. If there are n output nodes in the width direction and m output nodes in the height direction, then there are n×m sub-models in total. Divide the original output data into the corresponding sub-model output data according to the division method. Step 6: Build an artificial neural network model, which includes a two-variable input layer, a hidden layer, and an output layer with the same number of output nodes as the length direction; the activation function of the hidden layer is tanh, and the activation function of the output layer is relu; Step 7: Perform artificial neural network model training on each sub-model in turn to obtain a machine learning wake model for each sub-model; Step 8: The machine learning wake models of all sub-models are numbered and aggregated in order of spatial position to form a complete machine learning wake model of the wind turbine.

2. The method for constructing a wind turbine wake model based on machine learning according to claim 1, characterized in that: In step 1, the k-epsilon turbulence model combined with the ADM-R model is used to perform fluid dynamics numerical simulation modeling of the wind turbine. The length × width × height of the calculation domain is greater than or equal to 28D × 12D × 4D, where D is the rotor diameter of the wind turbine.

3. The method for constructing a wind turbine wake model based on machine learning according to claim 1, characterized in that: In step 2, the inflow conditions are uniformly selected within the entire operating inflow condition range, and the wind speed and turbulence at the hub height are determined for each inflow condition.

4. The method for constructing a wind turbine wake model based on machine learning according to claim 1, wherein: In step 4, a fixed interval is determined in the calculation domain, and the output node is determined based on the interval; each output node is interpolated to obtain the data of the output node, and the data under different working conditions are concatenated in a file as the original output data, and the wind speed and turbulence at the hub height of the working condition are recorded in sequence in a separate file as the input data.

5. The method for constructing a wind turbine wake model based on machine learning according to claim 1, characterized in that: The output data of each sub-model and the input data of the working condition sequence are read in, the training number threshold is set, the artificial neural network model is trained, and the model with the highest accuracy is recorded as the final machine learning wake model of each sub-model.

6. A wind farm layout optimization method, characterized in that: The wake model constructed by the method according to any one of claims 1 to 5 is used to find the optimal solution to the layout optimization problem, the steps being: Step 9: Determine the inflow conditions that need to be calculated based on the inflow conditions of the wind farm, including wind speed, turbulence, and direction; Step 10: Provide a random layout plan, determine the upstream and downstream relationship of the wind turbines according to each inflow condition, and use the machine learning wake model to solve the wake effect of the wind turbine from the upstreammost wind turbine in sequence; Step 11: For the downstream wind turbines, the wind speed and turbulence are superimposed using the corresponding wake superposition model; the wind speed and turbulence under the combined influence of the wake effects of multiple wind turbines are obtained as the inflow conditions for the downstream wind turbines; Step 12: Determine the power generation capacity of the wind turbine generator set based on its power-wind speed curve and the inflow conditions of the wind turbine generator set; Step 13: Add the generated power of all wind turbines to obtain the initial power output of the wind farm; Step 14: Calculate the next possible optimized layout scheme using the optimization algorithm, and calculate the total power output of the wind farm under the layout scheme; if the total power output is higher than the initial total power output, record the layout scheme as the optimized solution; Step 15: Repeat step 14 until the convergence condition of the optimization problem is met, then stop the optimization and take the final optimization solution as the local optimal solution of the optimization problem; Step 16: Randomize an initial distribution again and repeat steps 10-15 until the set total number of randomizations is reached to obtain multiple local optimal solutions; Step 17: The best solution among multiple local optimal solutions is used as the optimal layout.

Images