Wind power plant wake flow prediction method based on improved physical information neural network

Abstract

The invention discloses a wind power plant wake flow prediction method based on an improved physical information neural network. The wind power plant wake flow prediction method comprises the following steps: constructing a basic physical information neural network fusing a laser wind finding radar module, a two-dimensional Navier-Stokes equation and an actuating disc model; a cosine annealing algorithm and a dynamic loss function weight are adopted for the basic physical information neural network, and a step-by-step time saving strategy is introduced for training to construct an improved physical information neural network; constructing a training data set based on the radial wind speed data of the sparse measurement points of the laser wind finding radar, and training the improved physical information neural network to obtain a wind power plant wake flow prediction model; and inputting spatial position data and time data to obtain the predicted wake wind speed and air pressure of the wind power plant. According to the invention, the improved physical information neural network combines the measurement data with the physical law, and achieves the accurate and efficient prediction of the wake flow of the whole wind power plant through employing a precision improvement strategy and a step-by-step time saving strategy.

Description

Technical Field

[0001] The present invention relates to wind power generation and deep learning technology, and in particular to a method for predicting wind farm wake based on an improved physical information neural network. Background Art

[0002] As one of the most important renewable energy sources, wind power generation technology has developed rapidly around the world. With the expansion of wind farms and the large-scale development of wind turbines, the wind energy loss caused by the wake effect has become more serious. Therefore, in order to reduce the impact of the wake effect and improve the efficiency of wind turbines in utilizing wind energy, accurate and efficient wake prediction is of great significance for the optimal design and control of wind farms. The wake effect refers to the turbulent area formed behind the upstream wind turbine after extracting wind energy from the windward side, resulting in a decrease in the wind speed received by the downstream wind turbine and an increase in turbulence, thereby reducing the power generation efficiency of the wind turbine and increasing fatigue losses.

[0003] Currently, traditional wake prediction models based on deep learning technology require a large amount of data to achieve model training, but in practical applications it is often difficult to obtain the required amount of data, which hinders its widespread application in engineering practice.

[0004] Laser wind radar technology has been widely used in the field of wind power due to its advantages of high measurement accuracy, high temporal and spatial resolution, wide detection range, and fast response speed. However, laser wind radar can only measure the apparent wind speed and direction at sparse locations along the direction of the laser beam, and cannot obtain detailed information on the wake of a wind farm.

[0005] Therefore, there is an urgent need for a technical solution that can use the line-of-sight wind speed data from sparse measurement points of laser wind radar to achieve accurate and efficient prediction of wind farm wakes. Summary of the invention

[0006] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide a method for predicting wind farm wakes based on an improved physical information neural network.

[0007] Technical solution: A wind farm wake prediction method based on an improved physical information neural network of the present invention constructs a wind farm wake prediction model. The input data of the wind farm wake prediction model includes spatial position data and time data. The output data of the wind farm wake prediction model is the wind farm wake wind speed. The wind farm wake prediction model includes an improved physical information neural network and integrates a laser wind measurement radar module, a two-dimensional Navier-Stokes equation and an actuator disk model. The laser wind measurement radar module combines sparse measurement data of the laser wind measurement radar, and the two-dimensional Navier-Stokes equation and the actuator disk model are used as physical law constraints. The specific method is as follows:

[0008] Step 1: For the basic physical information neural network, the cosine annealing algorithm and the dynamic loss function weight strategy are used to integrate the laser wind radar module, the two-dimensional Navier-Stokes equations and the actuator disk model to construct an improved physical information neural network;

[0009] Step 2: Construct a training data set and introduce a step-by-step time saving strategy to train the improved physical information neural network obtained in step 1 to obtain a wind farm wake prediction model. During the training process, the laser wind radar module calculates the line-of-sight wind speed u LoS ;The training data set includes the line-of-sight wind speed data of sparse measurement points of the laser wind radar;

[0010] Step 3: Input spatial position data and time data into the trained wind farm wake prediction model, that is, input the space-time coordinates (x, y, t), and obtain the predicted wind farm wake flow direction wind speed u, span direction wind speed v and air pressure p.

[0011] The loss function of the laser wind radar module in the present invention Essentially, it refers to the comparison between the predicted apparent wind speed and the measured apparent wind speed; one of the goals of the neural network training process is to minimize the difference between the predicted apparent wind speed and the measured apparent wind speed, and its ultimate goal is to achieve the total loss function minimize.

[0012] Furthermore, the two-dimensional Navier-Stokes equation is expressed as:

[0013]

[0014] In the formula, represents partial derivative operation; u represents the flow direction wind speed; v represents the span direction wind speed; t represents the prediction time; x represents the flow direction position; y represents the span direction position; ρ represents the air density; p represents the air pressure; υ represents the air kinematic viscosity;

[0015] The expression of the actuator disc model is:

[0016]

[0017] In the formula, x f represents the flow position of the center of the actuator disk; y f is the spanwise position of the center of the actuator disk; f(x f ,y f ) means that the center position coordinate is (x f ,y f )'s body force on the actuator disk; N elem is the total number of discrete elements of the actuator disk; f irepresents the volume force at the discrete element of the i-th actuator disk; ε is the smoothing parameter of the Gaussian kernel; exp(·) represents the exponential function with the mathematical constant e as the base; is the flow position of the discrete element of the i-th actuator disk; represents the spanwise position of the discrete element of the i-th actuator disk.

[0018] Furthermore, through the loss function The laser wind radar module, two-dimensional Navier-Stokes equations and actuator disk model are integrated into the basic physical information neural network, and the loss function The formula is:

[0019]

[0020] In the formula, w 1 Represents the loss function weight of the laser wind radar module; is the loss function of the laser wind radar module; w 2 Represents the loss function weight of the Navier-Stokes equation; is the loss function of the Navier-Stokes equation; w 3 represents the loss function weight of the actuator disk model; is the loss function of the actuator disk model; finally, the neural network training is carried out with the goal of minimizing the loss function value.

[0021] Furthermore, the loss function is weighted by the dynamic loss function strategy. Make adaptive adjustments so that the loss function The three weights in are dynamically updated by gradient ascent, and the expressions are as follows:

[0022]

[0023] In the formula, and Respectively represent the loss function weights of the laser wind radar module in the k+1th and kth training rounds; is the weight w 1 The learning rate during update; Indicates the loss function is calculated with respect to weight w 1 The partial derivative of is the total loss function; Θ k is the training parameter of the neural network in the kth training round; and Represent the loss function weights of the Navier-Stokes equation at the k+1th and kth training rounds respectively; is the weight w 2 The learning rate during update; Indicates the loss function is calculated with respect to weight w 2 The partial derivative of and Respectively represent the loss function weights of the actuator disk model in the k+1th and kth training rounds; Indicates the loss function is calculated with respect to weight w 3 The partial derivative of .

[0024] Furthermore, the use of the cosine annealing algorithm is a warm restart stochastic gradient descent method. The cosine annealing algorithm is used to gradually reduce the learning rate in each warm start cycle during the training of the improved physical information neural network, thereby improving the performance of the improved physical information neural network. The formula is expressed as:

[0025]

[0026] Where η j represents the learning rate of the jth training round; η max is the maximum learning rate; η min is the minimum learning rate; T cur represents the number of rounds that the neural network has been trained since the warm restart; T e Indicates the number of training times required between two consecutive warm restarts, and its initial value is T 0 , after each warm restart, the coefficient T mult Expand, that is, T e+1 =T mult ·T e .

[0027] Furthermore, the specific method of introducing a step-by-step time saving strategy to train the improved physical information neural network in step 2 is: decomposing the entire training data set into multiple sub-training sets according to the time domain, and each sub-training set is gradually trained by the same improved physical information neural network according to the time sequence; except for the sub-training set used in the initial training, the subsequent sub-training sets will be combined with the previous sub-training set data to train the improved physical information neural network; after each step of training is completed, the improved physical information neural network will use the training parameter information as the initial condition for the next step of training.

[0028] Furthermore, the visual wind speed u predicted by the laser wind radar module during training LoS The calculation formula is:

[0029] u LoS =u train cos(θ)-v train sin(θ)(10)

[0030] In the formula, u LoS is the predicted wind speed in the direction of the laser wind radar; utrain is the flow direction wind speed output during the neural network training process; v train is the spanwise wind speed output during the neural network training process; θ is the azimuth angle of the laser wind radar measurement beam.

[0031] Beneficial effects: The present invention introduces two precision improvement strategies, namely, cosine annealing algorithm and dynamic loss function weight, and a gradual time saving strategy, into the basic physical information neural network to form an improved physical information neural network, thereby achieving more accurate and more efficient prediction of wind farm wakes; in addition, the present invention integrates the wind speed data constraints of sparse measurement points of laser wind radar, the constraints of two-dimensional Navier-Stokes equations, and the constraints of the actuator disk model, and can achieve accurate and efficient prediction of wind farm wakes by combining measurement data with physical laws. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the overall method of the present invention.

[0033] Figure 2 It is a layout diagram of a wind farm example in an embodiment of the present invention.

[0034] Figure 3 1 is a diagram of a neural network structure in an embodiment of the present invention.

[0035] Figure 4 It is a comparison diagram of the real wake and the wake predicted by the neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0037] like Figure 1 As shown, a method for predicting a wind farm wake based on an improved physical information neural network of the present invention constructs a wind farm wake prediction model, the input data of the wind farm wake prediction model includes spatial position data and time data, and the output data of the wind farm wake prediction model is the wind farm wake wind speed; the wind farm wake prediction model includes an improved physical information neural network, and integrates a laser wind radar module, a two-dimensional Navier-Stokes equation and an actuator disk model, the laser wind radar module combines the sparse measurement data of the laser wind radar, and the two-dimensional Navier-Stokes equation and the actuator disk model are used as physical law constraints; the specific method is as follows:

[0038] Step 1: For the basic physical information neural network, the cosine annealing algorithm and the dynamic loss function weight strategy are used to integrate the laser wind radar module, the two-dimensional Navier-Stokes equations and the actuator disk model to construct an improved physical information neural network;

[0039] Step 2: Construct a training data set and introduce a step-by-step time saving strategy to train the improved physical information neural network obtained in step 1 to obtain a wind farm wake prediction model. During the training process, the laser wind radar module calculates the line-of-sight wind speed u LoS ;The training data set includes the line-of-sight wind speed data of sparse measurement points of the laser wind radar;

[0040] Step 3: Input the spatial position data and time data (space-time coordinates (x, y, t)) into the trained wind farm wake prediction model to obtain the predicted wind speed u, wind speed v and air pressure p of the wind farm wake.

[0041] The spatial position data input to the wind farm wake prediction model in the present invention includes the two-dimensional coordinates (x, y) of the wake, such as Figure 2 As shown in the horizontal and vertical coordinates in ; the time data input to the wind farm wake prediction model includes time t. For example, if a wake of 400s needs to be predicted, the time data t is from 1s to 400s; finally, the spatial position data and the time data are input into the wind farm wake prediction model in the form of space-time coordinates (x, y, t), and then the wake flow direction wind speed u, span direction wind speed v and air pressure p of the corresponding coordinate point (x, y) at time t are directly output to the wind farm wake prediction model; the air pressure p is output because there is a term for partial derivative of the air pressure p in the two-dimensional Navier-Stokes equation.

[0042] In this embodiment, the two-dimensional Navier-Stokes equation is expressed as:

[0043]

[0044] In the formula, represents partial derivative operation; u is the wind speed in the stream direction; v is the wind speed in the span direction; t is the prediction time; x is the position in the stream direction; y is the position in the span direction; ρ represents the air density; p is the air pressure; υ is the kinematic viscosity of the air. The expression of the actuator disc model is:

[0045]

[0046] In the formula, x f represents the flow position of the center of the actuator disk; y f is the spanwise position of the center of the actuator disk; f(x f ,y f ) means that the center position coordinate is (x f ,y f )'s body force on the actuator disk; N elem is the total number of discrete elements of the actuator disk; f irepresents the volume force at the discrete element of the i-th actuator disk; ε is the smoothing parameter of the Gaussian kernel; exp(·) represents the exponential function with the mathematical constant e as the base; is the flow position of the discrete element of the i-th actuator disk; represents the spanwise position of the discrete element of the i-th actuator disk.

[0047] Step 1: Through the loss function The laser wind radar module, two-dimensional Navier-Stokes equations and actuator disk model are integrated into the basic physical information neural network, and the loss function The formula is:

[0048]

[0049] In the formula, w 1 Represents the loss function weight of the laser wind radar module; is the loss function of the laser wind radar module; w 2 Represents the loss function weight of the Navier-Stokes equation; is the loss function of the Navier-Stokes equation; w 3 represents the loss function weight of the actuator disk model; is the loss function of the actuator disk model.

[0050] In order to improve the prediction accuracy, the loss function is further adjusted through the dynamic loss function weight strategy. Make adaptive adjustments so that the loss function The three weights in are dynamically updated by gradient ascent, and the expressions are as follows:

[0051]

[0052] In the formula, and Respectively represent the loss function weights of the laser wind radar module in the k+1th and kth training rounds; is the weight w 1 The learning rate during update; Indicates the loss function is calculated with respect to weight w 1 The partial derivative of is the total loss function; Θ k is the training parameter of the neural network in the kth training round; and Represent the loss function weights of the Navier-Stokes equation at the k+1th and kth training rounds respectively; is the weight w 2 The learning rate during update; Indicates the loss function is calculated with respect to weight w2 The partial derivative of and Respectively represent the loss function weights of the actuator disk model in the k+1th and kth training rounds; Indicates the loss function is calculated with respect to weight w 3 The partial derivative of .

[0053] To further optimize the learning rate, the cosine annealing algorithm is used to gradually reduce the learning rate in each warm start cycle during the training of the improved physical information neural network. The expression is:

[0054]

[0055] Where η j represents the learning rate of the jth training round; η max is the maximum learning rate; η min is the minimum learning rate; T cur T represents the number of rounds that the neural network has been trained since the warm restart; e It indicates the number of training times required between two consecutive warm restarts, and its initial value is T 0 , after each warm restart, the coefficient T mult Expand, that is, T e+1 =T mult ·T e .

[0056] In order to reduce the training time and make the neural network model more efficient, step 2 also introduces a gradual time-saving strategy to train the improved physical information neural network. The specific method is: decompose the entire training data set into multiple sub-training sets according to the time domain, and each sub-training set is gradually trained by the same improved physical information neural network according to the time sequence; except for the sub-training set used in the initial training, the subsequent sub-training sets will be combined with the previous sub-training set data for improved physical information neural network training; after each step of training, the improved physical information neural network will use the training parameter information as the initial condition for the next step of training.

[0057] The visual wind speed u predicted by the laser wind radar module during the training process of this embodiment LoS The calculation formula is:

[0058] u LoS =u train cos(θ)-v train sin(θ)

[0059] In the formula, u LoS is the predicted wind speed in the direction of the laser wind radar; u train is the flow direction wind speed output during the neural network training process; v trainis the spanwise wind speed output during the neural network training process; θ is the azimuth angle of the laser wind radar measurement beam.

[0060] After completing the above steps, the streamwise wind speed and spanwise wind speed of the wake field are obtained through simulation calculation (regarded as the real streamwise wind speed and spanwise wind speed data), and then the required laser wind radar data are extracted from the obtained simulation data and a data set is constructed. Then, a wind farm wake prediction model is obtained through neural network training, and finally the streamwise wind speed and spanwise wind speed of the wake are predicted.

[0061] When evaluating the prediction accuracy, the wake flow direction wind speed and span direction wind speed predicted by the trained wake prediction model are compared with the flow direction wind speed and span direction wind speed (real wind speed data) of the wake field obtained by simulation calculation (the difference between the two).

[0062] The relevant prediction accuracy evaluation indicators are:

[0063]

[0064] In the formula, u MRMSE represents the average value of the root mean square error of the wind speed in the direction; T is the total prediction time, N pred Indicates the number of predicted spatial location points; (x k ,y k ,t) is the spatial location data and time data of the prediction point; u pred (x k ,y k ,t) represents the neural network's response to the spatial position point (x k ,y k ) is the predicted value of the streamwise wind speed at time t; u true (x k ,y k ,t) represents the spatial position point (x k ,y k ) is the true value of the streamwise wind speed at time t; v MRMSE represents the average value of the root mean square error of the spanwise wind speed; v pred (x k ,y k ,t) represents the neural network's response to the spatial position point (x k ,y k ) is the predicted spanwise wind speed at time t; v true (x k ,y k ,t) represents the spatial position point (x k ,y k ) is the true value of the spanwise wind speed at time t.

[0065] In order to verify the effect of the technical solution of the present invention, this embodiment takes a wind farm containing six NREL 5MW wind turbines as the research object and applies the technical solution of the present invention. The wind turbine rotor diameter D in the wind farm of this embodiment is 126.4m, the wind turbine flow spacing is 5D, and the span spacing is 3D. Figure 2 shown.

[0066] This example is based on the FAST.Farm medium-fidelity multi-physics engineering simulation tool, and performs a 700s simulation calculation with a wind speed of 10m / s and a turbulence of 6% inflow conditions. The laser wind radar is configured with a measurement range of 550m, an interval of 30m between adjacent measurement points, and a laser beam scanning measurement at a speed of 3.5° / s in a 42° opening angle range at the hub height. Figure 2 The black dots in the middle are schematic diagrams of sparse measurement points of the laser wind radar. To ensure that the wake is fully formed and stable, the wind speed data of the plane at the hub height is saved from the last 400s simulation as the true wind speed value. Based on the laser wind radar configuration, the line-of-sight wind speed data of the laser wind radar sparse measurement points are extracted from the saved wind speed data to construct a neural network training data set; based on the step-by-step time saving strategy, the training data set is used to train the neural network. Figure 3 The neural network model shown is trained.

[0067] Compared with the existing physical information neural network, the technical solution of the present invention has an improvement effect on the prediction accuracy and training speed of the wind farm wake (see Table 1 for the results), and a comparison between the real wake results and the wake results predicted by the method of the present invention is given, as shown in Figure 4 As shown. Among them, Figure 4 (a) in the figure is the wind farm wake obtained by simulation calculation; Figure 4 (b) in the figure shows the wake field predicted by the basic physical information neural network on the left, and the wake field error predicted by the basic physical information neural network on the right. Figure 4 In (c), the left figure is the wake field map predicted by the improved physical information neural network, and the right figure is the wake field error map predicted by the improved physical information neural network.

[0068] The above results show that the method proposed in the present invention can achieve accurate and efficient prediction of wind farm wake based on sparse wind speed data of laser wind radar measurement points.

[0069] Table 1 Comparison of prediction accuracy and training time of neural networks

[0070]

[0071] In summary, the improved physical information neural network method in the present invention combines the sparse measurement data of the laser wind radar with the physical laws, and realizes accurate and efficient prediction of the wake of the entire wind farm by adopting the precision improvement strategy and the step-by-step time saving strategy.

Claims

1. A wind farm wake prediction method based on an improved physical information neural network, characterized in that: A wind farm wake prediction model is constructed. The input data of the wind farm wake prediction model includes spatial position data and time data. The output data of the wind farm wake prediction model is the wind speed of the wind farm wake. The wind farm wake prediction model includes an improved physical information neural network and integrates a laser wind radar module, a two-dimensional Navier-Stokes equation and an actuator disk model. The laser wind radar module combines sparse measurement data of the laser wind radar, and the two-dimensional Navier-Stokes equation and the actuator disk model are used as physical law constraints. The specific method is as follows: Step 1: For the basic physical information neural network, the cosine annealing algorithm and the dynamic loss function weight strategy are used to integrate the laser wind radar module, the two-dimensional Navier-Stokes equations and the actuator disk model to construct an improved physical information neural network; Step 2: Construct a training data set and introduce a step-by-step time saving strategy to train the improved physical information neural network obtained in step 1 to obtain a wind farm wake prediction model. During the training process, the laser wind radar module predicts and calculates the line-of-sight wind speed u. LoS ;The training data set includes the line-of-sight wind speed data of sparse measurement points of the laser wind radar; Step 3: Input spatial position data and time data into the trained wind farm wake prediction model, that is, input the space-time coordinates (x, y, t), and obtain the predicted wind farm wake flow direction wind speed u, span direction wind speed v and air pressure p.

2. The wind farm wake prediction method based on improved physical information neural network according to claim 1 is characterized in that: The two-dimensional Navier-Stokes equation expression is: In the formula, represents partial derivative operation; u is the wind speed in the stream direction; v is the wind speed in the span direction; t is the prediction time; x is the position in the stream direction; y is the position in the span direction; ρ is the air density; p is the air pressure; υ is the kinematic viscosity of the air; The expression of the actuator disc model is: In the formula, x f represents the flow position of the center of the actuator disk; y f is the spanwise position of the center of the actuator disk; f(x f ,y f ) means that the center position coordinate is (x f ,y f )'s body force on the actuator disk; N elem is the total number of discrete elements of the actuator disk; f i represents the volume force at the discrete element of the i-th actuator disk; ε is the smoothing parameter of the Gaussian kernel; exp(·) represents the exponential function with the mathematical constant e as the base; is the flow position of the discrete element of the i-th actuator disk; represents the spanwise position of the discrete element of the i-th actuator disk.

3. The method for predicting wind farm wake based on improved physical information neural network according to claim 1 is characterized in that: Through the loss function The laser wind radar module, two-dimensional Navier-Stokes equations and actuator disk model are integrated into the basic physical information neural network, and the loss function The formula is: Where w1 represents the loss function weight of the laser wind radar module; is the loss function of the laser wind radar module; w2 represents the loss function weight of the Navier-Stokes equation; is the loss function of the Navier-Stokes equation; w3 represents the loss function weight of the actuator disk model; is the loss function of the actuator disk model.

4. The method for predicting wind farm wake based on improved physical information neural network according to claim 3 is characterized in that: The loss function is adjusted by the dynamic loss function weight strategy Make adaptive adjustments so that the loss function The three weights in are dynamically updated by gradient ascent, and the expressions are as follows: In the formula, and Respectively represent the loss function weights of the laser wind radar module in the k+1th and kth training rounds; is the learning rate when weight w1 is updated; Indicates the partial derivative of the loss function with respect to weight w1; is the total loss function; Θ k is the training parameter of the neural network in the kth training round; and Respectively represent the loss function weights of the Navier-Stokes equation at the k+1th and kth training rounds; is the learning rate when updating weight w2; Indicates the partial derivative of the loss function with respect to weight w2; and Respectively represent the loss function weights of the actuator disk model in the k+1th and kth training rounds; It means to find the partial derivative of the loss function with respect to the weight w3.

5. The method for predicting wind farm wake based on improved physical information neural network according to claim 1 or 3, characterized in that: The cosine annealing algorithm is used to gradually reduce the learning rate in each warm start cycle during the training of the improved physical information neural network. The expression is: Where η j represents the learning rate of the jth training round; η max is the maximum learning rate; η min is the minimum learning rate; T cur represents the number of rounds that the neural network has been trained since the warm restart; T e It indicates the number of training times required between two consecutive warm restarts. Its initial value is T0. After each warm restart, the coefficient T mult Expand, that is, T e+l =T mult ·T e .

6. The method for predicting wind farm wake based on improved physical information neural network according to claim 1, characterized in that: The specific method of introducing the stepwise time saving strategy to train the improved physical information neural network in step 2 is as follows: decomposing the entire training data set into multiple sub-training sets according to the time domain, and each sub-training set is gradually trained by the same improved physical information neural network according to the time sequence; except for the sub-training set used in the initial training, the subsequent sub-training sets will be combined with the previous sub-training set data to train the improved physical information neural network; After each step of training is completed, the improved physical information neural network uses the training parameter information as the initial condition for the next step of training.

7. The method for predicting wind farm wake based on improved physical information neural network according to claim 1, characterized in that: The visual wind speed u predicted by the laser wind radar module during training LoS The calculation formula is: you LoS =u train coS(θ)-v train sin(θ) In the formula, u LoS is the predicted wind speed in the direction of the laser wind radar; u train is the flow direction wind speed output during the neural network training process; v train is the spanwise wind speed output during the neural network training process; θ is the azimuth angle of the laser wind radar measurement beam.

Images