A wind turbine power prediction method based on dynamic inflow conditions

Abstract

The application relates to the technical field of wind power generation, in particular to a wind turbine power prediction method based on dynamic inflow conditions. The method comprises the following steps: establishing a wind turbine unsteady numerical simulation model, using CFD software to establish a wind turbine three-dimensional unsteady numerical simulation model, dividing a calculation domain and generating a grid, setting an inlet velocity boundary condition, an outlet pressure boundary condition and a symmetric boundary condition, adopting a turbulence model to solve a transient Reynolds average Navier-Stokes equation, verifying the accuracy of a numerical method through grid independence verification and comparison with public literature data, simulating dynamic inflow conditions such as gradual wind, extreme gust and disturbance wind, constructing an inlet wind speed change and output power response database, extracting characteristic parameters such as a wind speed change rate, a transition time and a power overshoot amplitude from the database, and training a power overshoot amplitude prediction model.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, specifically a method for predicting wind turbine power based on dynamic inflow conditions. Background Technology

[0002] Wind power generation, as a crucial component of clean energy, plays a vital role in driving energy structure transformation and achieving goals. With the continuous increase in installed wind power capacity and grid connection rates, the fluctuating power output characteristics of wind turbines have an increasingly prominent impact on power system dispatching, reserve capacity allocation, energy storage system planning, and the safe and stable operation of the power grid. In actual natural wind environments, the incoming wind speed is not constant but often exhibits dynamic changes such as gradually varying wind speeds, extreme gusts, and periodic disturbances. These dynamic inflow conditions cause continuous adjustments in the wind turbine's induced velocity, blade aerodynamic loads, and wake flow field, resulting in significant unsteady response characteristics and potentially leading to short-term overshoot or lag in power output during or after wind speed changes. Therefore, accurately describing the power response characteristics of wind turbines under dynamic inflow conditions and achieving rapid prediction is of great significance for optimizing wind farm operation and ensuring safe power system dispatching.

[0003] Currently, the closest existing technologies to this invention mainly fall into two categories: one is the unsteady numerical simulation method based on computational fluid dynamics (CFD), and the other is the wind power prediction model based on historical operating data or statistical learning methods. The CFD method typically establishes a three-dimensional computational model of the wind turbine, applies time-varying inlet wind speed boundary conditions using methods such as sliding meshes, moving meshes, or user-defined functions, and solves the unsteady Reynolds-averaged Navier-Stokes equations to obtain the aerodynamic loads, wake flow field, and power time series of the wind turbine under dynamic inflow conditions. This method can accurately reveal the physical relationships between wind speed changes, induced velocity lag, wake vortex structure evolution, and power overshoot, and is suitable for dynamic inflow mechanism analysis and high-fidelity numerical verification.

[0004] Another type of existing technology is the wind power prediction model, which mainly relies on meteorological forecast data, historical wind speed, historical power, or wind farm operation data. It uses physical models, statistical models, or machine learning models to establish a mapping relationship between input parameters and output power. This type of method has been widely used in conventional wind power prediction and can meet the short-term or ultra-short-term power prediction needs of wind farms to a certain extent. However, existing power prediction models typically focus on power changes under average wind speed or steady-state inflow conditions, failing to adequately consider the impact of dynamic inflow effects and making it difficult to accurately reflect the unsteady power response of wind turbines under dynamic wind conditions.

[0005] In summary, while existing technologies can study wind turbine power variation from the perspectives of high-fidelity numerical simulation or data-driven prediction, a method that can simultaneously address the dynamic inflow physics mechanism, extract key characteristic parameters, and provide rapid prediction capabilities is still lacking. To address this, this invention proposes a wind turbine power prediction method based on dynamic inflow conditions. By constructing a numerical simulation database of typical dynamic inflow conditions, it extracts characteristic parameters with clear physical meaning, such as wind speed change rate, transition time, inflow wind speed overshoot amplitude, and average wind speed. Furthermore, it establishes a wind turbine power overshoot amplitude prediction model using a neural network, thereby achieving rapid prediction of the power response of a single wind turbine and downstream wind turbines in a wind farm under dynamic wind conditions. This method can provide technical support for the unsteady performance evaluation of wind turbines under dynamic inflow conditions, rapid wind power prediction, and optimized operation of wind farms. Summary of the Invention

[0006] (1) Technical problems to be solved This invention aims to address the problems of insufficient prediction accuracy, high computational cost, low efficiency in engineering applications, and weak physical interpretability in existing wind turbine power prediction methods under dynamic inflow conditions. Specifically, this invention addresses the difficulty of traditional steady-state or quasi-steady-state power prediction methods in accurately reflecting the induced velocity adjustment lag, aerodynamic load hysteresis response, and power overshoot caused by wind speed changes under unsteady inflow conditions such as gradually changing winds, extreme gusts, and turbulent winds, thereby improving the accuracy of wind turbine power response prediction under dynamic wind conditions.

[0007] (2) Technical solution To achieve the above objectives, the present invention provides a wind turbine power prediction method based on dynamic inflow conditions, comprising: S1. Obtain the geometric parameters of the target wind turbine, establish a three-dimensional unsteady numerical simulation model of the wind turbine using CFD software, divide the computational domain and generate a mesh, set inlet velocity boundary conditions, outlet pressure boundary conditions and symmetric boundary conditions, solve the transient Reynolds-averaged Navier-Stokes equations using a turbulence model, and verify the accuracy of the numerical method by verifying mesh independence and comparing with data from published literature.

[0008] S2. Based on the validated numerical simulation model, the system systematically simulates the gradual wind speed condition, extreme gust condition, and disturbed wind condition. Unsteady numerical simulations are performed on each set of condition parameters, and the inlet wind speed-time curve and the wind turbine output power-time curve are recorded to construct a dynamic inflow condition database.

[0009] S3. Extract the wind speed change rate from the simulation results. and transition time As an input feature, the power overshoot amplitude As the output target, the power overshoot amplitude is defined as ,in This represents the maximum output power of the wind turbine during the duration of dynamic wind conditions. The instantaneous power value at the end of the dynamic wind condition; and the training sample library for the power prediction model of a single wind turbine.

[0010] S4, based on wind speed change rate and transition time Input power overshoot amplitude For the output, an error backpropagation neural network is used to establish a nonlinear mapping relationship between the input features and the output target, thus obtaining a fast prediction model for the power overshoot amplitude of a single wind turbine.

[0011] S5. Establish a numerical model of a dual-wind turbine wind farm and extract the overshoot amplitude of the incoming wind speed at the downstream wind turbine inlet. and average wind speed As a wake correction parameter, the input features are expanded to , , and With four parameters, a BP neural network was retrained to establish a prediction model for the power overshoot amplitude of downstream wind turbines in wind farms affected by upstream wakes.

[0012] Furthermore, in step S1, the computational domain is divided as follows: it extends from 6R upstream to 10R downstream of the wind turbine rotor plane, with the radial boundary 5R away from the rotor axis, where R is the rotor radius; the rotor rotation is simulated using a sliding mesh technique, and the rotating domain and the stationary domain are connected through an interface; the Realizable k-ε model is selected as the turbulence model, and the transient RANS equations are solved using the SIMPLE algorithm; the time step is set to the physical time corresponding to the rotor rotation of 0.5° to 1°; the mesh is an unstructured tetrahedral mesh, and local refinement is applied to the blade surface, leading edge, trailing edge, tip / root region, and near-wake region.

[0013] Furthermore, in step S2, the core parameters of the gradual wind speed condition include the initial wind speed, the final wind speed, and the transition time, during which the wind speed changes linearly with time; the extreme gust condition is defined according to the IEC61400-1 standard, and its core parameters include the periodic influence factor and the gust duration, simulating the process of the wind speed first rising rapidly and then recovering; the core parameters of the disturbed wind condition include the average wind speed, the disturbance amplitude, and the disturbance frequency, with the wind speed fluctuating periodically with time, used to simulate the process of the incoming wind speed continuously disturbing near the average wind speed in a natural wind environment.

[0014] Furthermore, in step S3, the wind speed change rate Defined as (final wind speed - initial wind speed) / transition time, in m / s²; the transition time The duration of the wind speed change, in seconds; in step S5, the amplitude of the incoming wind speed overshoot. Defined as ,in This represents the maximum wind speed at the downstream wind turbine inlet. The average wind speed is the instantaneous wind speed value at the monitoring point at the end of the dynamic wind condition period. The average wind speed at the downstream wind turbine inlet during dynamic wind conditions is used to characterize the speed loss level caused by the upstream wake.

[0015] Furthermore, in step S4, the BP neural network includes one input layer, at least one hidden layer, and one output layer; the activation function of the hidden layer neurons is the sigmoid function; the loss function is the mean squared error; the training algorithm is the Levenberg-Marquardt algorithm; the overfitting prevention strategy is the early stopping method, which randomly divides all samples into a training set and a validation set, and terminates the training early when the validation set error no longer decreases in multiple consecutive iterations; in a preferred embodiment, the network structure is one input layer, three hidden layers, and one output layer, the sample ratio of the training set to the validation set is approximately 85%:15%, and the model's determination coefficient reaches above 0.953.

[0016] Furthermore, in step S5, the axial spacing between the upstream and downstream wind turbines in the numerical model of the dual-wind turbine wind farm is set to 12 times the rotor diameter, and they are arranged coaxially; monitoring points are set at key locations in the incoming flow in front of the downstream wind turbine rotor to record the inlet wind speed-time variation curve; the input features of the corrected BP neural network are four-dimensional vectors, and the output target is still the power overshoot amplitude. The network structure is adjusted accordingly based on the input dimension. After introducing the trail correction parameter, the model's determination coefficient is increased to over 0.98.

[0017] Furthermore, for the second row and subsequent wind turbines in the wind farm, the overshoot amplitude of the incoming wind speed at the downstream wind turbine inlet is obtained through a simple wake model or mapping using a small amount of pre-calculated database. and average wind speed The approximate value will be , , and The corrected prediction model is input to predict its power overshoot amplitude; the predicted power overshoot amplitude is combined with the steady-state power prediction value to obtain the extreme range of power output under dynamic operating conditions, providing input for grid dispatch and energy storage system control.

[0018] Furthermore, in step S4, the machine learning method for constructing the prediction model can be replaced by any one of support vector regression, random forest, XGBoost, long short-term memory neural network, or convolutional neural network. When there is sufficient sample data, the above models can establish a nonlinear mapping relationship between input features and power overshoot amplitude, and thus replace the BP neural network to achieve the purpose of this invention.

[0019] Furthermore, in step S2, the numerical simulation method for generating training samples can be replaced by any one of the following: large eddy simulation, actuated disk model, actuated line model, vortex method, or blade element-momentum theory modified model; the wind speed input sequence for dynamic inflow conditions can also be constructed based on wind farm anemometer data, lidar anemometer data, or SCADA operation data; in step S5, the inflow wind speed overshoot amplitude at the downstream wind turbine inlet... and average wind speed The acquisition method can be replaced by an engineering wake model, a data-driven wake model, or a correction method based on measured downstream wind speed; the above alternatives are all equivalent alternatives to the technical solution of this invention.

[0020] (3) Beneficial effects Compared with existing dynamic inflow analysis methods based on CFD unsteady numerical simulation, this invention significantly improves computational efficiency while maintaining prediction accuracy. Existing CFD methods require re-establishing computational tasks for each set of dynamic inflow conditions and performing long-term transient solutions, resulting in high computational costs and making it difficult to meet the needs of rapid prediction in engineering applications. This invention constructs a dynamic inflow condition database through preliminary numerical simulations and uses neural networks to establish a mapping relationship between dynamic inflow characteristic parameters and power overshoot amplitude. After model training, only parameters such as wind speed change rate and transition time need to be input to quickly obtain power overshoot prediction results, avoiding repeated high-cost calculations for similar conditions.

[0021] Compared with traditional steady-state or quasi-steady-state power prediction methods, this invention can more accurately reflect the dynamic inflow effect caused by wind speed changes over time. Traditional methods are usually based on average wind speed or steady-state power curves, which are difficult to describe induced velocity lag, load hysteresis response, and power overshoot phenomena under conditions of rapid wind speed changes. This invention introduces dynamic characteristic parameters such as wind speed change rate, transition time, inflow wind speed overshoot amplitude, and average wind speed into the prediction model, enabling the prediction results to reflect the unsteady characteristics of wind turbine power response under dynamic wind conditions, thereby improving the prediction accuracy under dynamic inflow wind conditions.

[0022] Compared to general data-driven prediction models, the input parameters of this invention have clear physical meanings, resulting in stronger model interpretability. Existing machine learning prediction models often rely on black-box fitting of historical wind speed and power data, failing to adequately consider the formation mechanism of power overshoot under dynamic inflow conditions. This invention, based on the dynamic inflow response laws revealed by numerical simulation, selects key parameters that characterize the intensity and duration of wind speed changes and the effect of wake modulation as inputs. This enables the model to not only have better nonlinear fitting capabilities but also reflect the physical relationship between dynamic inflow and power response.

[0023] Furthermore, this invention can be extended from single-turbine prediction to prediction scenarios of downstream wind turbines in wind farms. By introducing wake correction parameters such as the downstream incoming wind speed overshoot amplitude and average wind speed, this invention can consider the impact of the dynamic wake of upstream wind turbines on the inflow conditions and power response of downstream units, reducing the prediction bias caused by the wake coupling effect in existing methods. Therefore, this invention has the advantages of fast prediction speed, low computational cost, strong physical interpretability, wider applicability, and high engineering application value. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is an overall flowchart of the wind turbine power prediction method in a wind farm based on neural networks, according to the present invention. Figure 2 This is a graph showing wind speed-time versus corresponding power-time. Figure 3 The training flowchart for the BP neural network prediction model; Figure 4 This is a schematic diagram of a BP neural network structure; Figure 5 A graph showing the comparison between model predictions and actual CFD simulations; Figure 6 A comparison chart showing the correction between the model's predictions and the actual values ​​from the CFD simulation. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Throughout this specification, references to "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "an embodiment," "an example," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] While existing dynamic inflow analysis methods based on CFD unsteady numerical simulation can accurately obtain the flow field structure, load changes, and power response results of wind turbines under dynamic wind conditions, these methods have high requirements for computational resources and time. For a single dynamic inflow condition, it is usually necessary to establish a refined three-dimensional geometric model, generate a high-quality mesh, set a sliding mesh and unsteady boundary conditions, and perform long-term transient iterative calculations. The computation cycle is long, making it difficult to meet the needs of rapid power prediction in wind farm operation and scheduling. At the same time, the CFD simulation process involves multiple steps, such as turbulence model selection, time step setting, mesh independence verification, and boundary condition definition, which requires high professional skills from users and lacks convenience for engineering applications.

[0029] Existing steady-state or quasi-steady-state power prediction methods typically use average wind speed, historical power, or weather forecast data as primary inputs, failing to adequately consider the dynamic inflow effects caused by wind speed variations over time in actual wind conditions. While these methods are applicable to conditions of stable wind speed changes or normal operation, they struggle to accurately capture dynamic characteristics such as induced velocity adjustment lag, load response lag, and power overshoot when wind turbines encounter unsteady inflows with varying wind speeds. This results in significant discrepancies between the predicted results and the actual power response.

[0030] Existing data-driven prediction models often rely on large amounts of historical operational data to establish a statistical mapping relationship between input and output. The internal physical mechanisms of these models are not clearly defined, and they do not adequately consider the intrinsic relationship between dynamic characteristic parameters of dynamic inflow wind conditions and power response. Therefore, while these models may have good fitting performance within the training data coverage area, their generalization and extrapolation abilities are limited when encountering insufficiently trained dynamic wind conditions or extreme operating conditions, making it difficult to guarantee the reliability of the prediction results.

[0031] Furthermore, in wind farm scenarios, the wakes of upstream wind turbines alter the inflow conditions for downstream turbines, subjecting them to the combined effects of velocity loss, enhanced turbulence, and dynamic wake propagation. Existing power prediction methods typically treat wind turbines within the wind farm as relatively independent entities or only consider average wake losses, failing to adequately account for the downstream inflow wind speed overshoot caused by upstream dynamic wakes and its modulation effect on power response. This leads to systematic errors in downstream wind turbine power prediction. Overall, existing technologies struggle to simultaneously meet the requirements of expressing the physical mechanism of dynamic inflow, achieving high prediction accuracy, low computational cost, and rapid engineering application.

[0032] To address the issues of insufficient accuracy, high computational cost, and difficulty in quantifying the impact of wind farm wakes in wind turbine power prediction under dynamic inflow conditions, this invention proposes a wind turbine power prediction method based on dynamic inflow conditions. This method is based on the unsteady response characteristics of wind turbines under typical dynamic wind conditions. It constructs a sample database through numerical simulation, extracts key parameters characterizing dynamic inflow features, and uses a neural network to establish a mapping relationship between input features and power overshoot amplitude, thereby achieving rapid prediction of wind turbine power response under dynamic wind conditions.

[0033] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can understand and implement it.

[0034] Figure 1 This is a flowchart illustrating the overall process of the wind turbine power prediction method in a wind farm based on neural networks, as described in this invention. This method systematically starts from the study of fundamental physical mechanisms, constructs a database through numerical simulation, and utilizes neural networks to establish a mapping relationship between key dynamic characteristic parameters and power response, ultimately achieving rapid prediction of power overshoot amplitude from a single turbine to the entire wind farm.

[0035] A wind turbine power prediction method based on dynamic inflow conditions mainly includes the following steps: S1. Establish and verify a high-fidelity numerical simulation model; First, select the target wind turbine model, such as the NREL-5MW wind turbine used in this specific embodiment. Obtain its detailed geometric parameters (such as blade airfoil spanwise distribution, rotor diameter, rated speed, etc.).

[0036] Secondly, a three-dimensional unsteady numerical simulation model of the wind turbine was established using commercial CFD software (such as ANSYS Fluent). Specific steps included: S11. Computational Domain Partitioning: Create a computational region comprising a rotating domain (around the rotor) and a stationary domain, connected by an interface. A sliding mesh technique is used to simulate rotor rotation. The computational domain is defined as extending from 6R (R, rotor radius) upstream of the wind turbine rotor plane to 10R downstream, with a radial boundary distance of 5R relative to the rotor axis.

[0037] S12. Mesh generation: The computational domain is discretized using unstructured tetrahedral meshes, and local refinement is applied to the blade surface, leading edge, trailing edge, tip / root region, and near-wake region to ensure the capture of key flow features.

[0038] S13. Boundary Conditions and Solution Settings: The inlet is set as a velocity inlet, and variable wind speed boundary conditions are applied through a user-defined function (UDF); the outlet is set as a pressure outlet; the outer wall is set as a symmetric boundary. The Realizable k-ε model is selected as the turbulence model, and the SIMPLE algorithm is used to solve the transient RANS equations. The time step is set to the physical time corresponding to the rotor rotation from 0.5° to 1°.

[0039] S14. Method Validation: Perform grid independence validation and comparison validation with publicly available literature data, including comparison validation of parameters such as wind turbine thrust and power, to ensure the accuracy of the numerical method.

[0040] S2. Define typical dynamic inflow conditions and generate a sample database; based on the validated model, systematically simulate three typical dynamic inflow conditions: 1. Gradual Wind Speed ​​Condition: Wind speed changes linearly with time. Key parameters include: initial wind speed (e.g., rated wind speed 11.4 m / s), final wind speed (e.g., 13, 15, 17, 20 m / s), and transition time (e.g., 0.5T, 1T, 2T, 4T, where T is the wind turbine's rotational period). Wind speed-time and power-time curves are shown below. Figure 2 As shown.

[0041] Extreme gust conditions: Based on the IEC 61400-1 standard definition, this simulates the process of wind speed rising rapidly and then recovering. The core parameters are: periodicity factor (e.g., 3.3, 4.8, 6.4, corresponding to different return periods) and gust duration (e.g., 1T, 2T, 3T, 4T, 5T).

[0042] Disturbance wind condition: Wind speed fluctuates periodically over time, used to simulate the continuous disturbance of incoming wind speed around the average wind speed in a natural wind environment. The core parameters are: average wind speed (e.g., rated wind speed 11.4 m / s), disturbance amplitude, and disturbance frequency (e.g., 0.5, 1.0, 1.5, etc.), where disturbance amplitude characterizes the intensity of wind speed fluctuation, and disturbance frequency characterizes the rate of wind speed change.

[0043] Unsteady numerical simulations were performed on each set of operating parameters, and complete inlet wind speed-time curves and wind turbine output power-time curves were recorded.

[0044] S3. Extract key feature parameters and construct a training dataset; extract key features for model training from the simulation results: 1. Input Features: For a single wind turbine, the core features are two parameters describing the dynamic inflow intensity: Wind speed change rate ( : Defined as (terminal wind speed - initial wind speed) / transition time, in m / s².

[0045] Transition time ( ): The duration of a change in wind speed, measured in seconds.

[0046] 2. Output target: Power overshoot amplitude ( Its definition is: in, This represents the maximum output power of the wind turbine during the duration of dynamic wind conditions. This represents the instantaneous power value at the end of the dynamic wind condition.

[0047] Under different working conditions ( , ) as input, corresponding As output, these samples collectively form the training sample library for the power prediction model of a single wind turbine. For example, in this embodiment, 41 sets of valid samples were generated based on gradually changing wind conditions.

[0048] S4. Construct and train a BP neural network prediction model; use a backpropagation (BP) neural network to establish a nonlinear mapping relationship between input features and output target. The specific training process is as follows: Figure 3 As shown.

[0049] 1. Network Structure: The optimal structure was determined through extensive comparative experiments. In a preferred embodiment, the network contains one input layer (two neurons, corresponding to...). and The network consists of 3 hidden layers (8, 8, and 4 neurons respectively) and 1 output layer (1 neuron, outputting A_p). A schematic diagram of the network structure is shown below. Figure 4 As shown.

[0050] 2. Key Settings: Activation function: The hidden layer neurons use the Sigmoid function.

[0051] Loss function: Mean squared error (MSE) is used.

[0052] Training algorithm: The Levenberg-Marquardt algorithm is used to accelerate convergence.

[0053] Overfitting prevention strategy: Employ the "early stopping" method. Randomly divide all samples into a training set (approximately 85%) and a validation set (approximately 15%). The training process is as follows: Figure 3 As shown, training is terminated early when the validation set error no longer decreases in consecutive iterations.

[0054] 3. Model Training and Validation: The network is trained using the pre-built sample library. After training, the model can quickly predict the overshoot amplitude of the wind turbine power under this dynamic condition based on the input wind speed change rate (a) and transition time (Δt). ). Figure 5 This diagram illustrates the comparison between the model's predictions and the actual values ​​from the CFD simulation. The horizontal axis represents each predicted sample point, and the vertical axis represents the corresponding result for each sample point (representing the power overshoot in this model). The top and bottom graphs compare the prediction results for the training and test sets, respectively. It can be seen that the prediction results for both the training and test sets closely match the actual values, with a high coefficient of determination (COP). The accuracy reached over 0.953, demonstrating extremely high prediction accuracy.

[0055] S5. Correction of the downstream wind turbine power prediction model in wind farms; This embodiment illustrates how to expand and correct the single-unit model to be applicable to downstream wind turbines in wind farms affected by upstream wakes.

[0056] S51. Establish a numerical model of a dual-wind turbine wind farm; establish a simplified wind farm model consisting of two NREL-5MW wind turbines. The axial spacing between the upstream and downstream wind turbines is set to 12 times the rotor diameter (12D), and they are arranged coaxially. The computational domain setting, mesh strategy, and solution method are the same as those for the single-turbine model.

[0057] S52. Obtain a dynamic operating condition sample database of the wind farm; apply a gradual wind condition similar to that in Example 1 to the above dual-turbine model. Through numerical simulation, simultaneously record the power output of the upstream and downstream wind turbines, and set monitoring points at key locations in the incoming flow in front of the downstream wind turbine rotor (such as near the blade tip height) to record its inlet wind speed-time variation curve.

[0058] S53. Analyze the input characteristics of downstream wind turbines and correct the model; the analysis revealed that the inlet wind speed of downstream wind turbines also exhibits overshoot. Therefore, the key factors affecting the power overshoot amplitude, besides the wind speed change rate (…), include… ) and transition time ( In addition to ), it should also include: Incoming air velocity overshoot amplitude ( Its definition is analogous to the power overshoot amplitude: in, This represents the maximum wind speed at the downstream wind turbine inlet. This represents the instantaneous wind speed at that point at the end of the dynamic wind condition.

[0059] Average wind speed ( ): During dynamic wind conditions, the average wind speed at the downstream wind turbine inlet reflects the speed loss level caused by the upstream wake.

[0060] S54. Construct a revised neural network prediction model; based on dual-machine simulation data, construct a new training sample library. For downstream wind turbines, each sample has: The input features are expanded to four: , , , .

[0061] The output target remains the power overshoot amplitude. .

[0062] Using the same BP neural network construction and training method as in Example 1 (the network structure needs to be adjusted according to the input dimension), a modified prediction model is established. This model can comprehensively consider the coupling effect of dynamic inflow and wake interference. Figure 6 The comparison between the predicted results of the corrected model and the actual values ​​of the CFD simulation is shown. The horizontal axis of the figure represents each predicted sample point, and the vertical axis represents the corresponding result for each sample point (representing the overshoot amplitude of power in this model). The top and bottom figures compare the prediction results for the training and test sets, respectively. The results show that after introducing the wake effect parameter, the predicted values ​​match the actual values ​​well, and the coefficient of determination (COP) is relatively high. The accuracy was improved to over 0.98, verifying the effectiveness of the correction.

[0063] S55. Application Example: For a wind farm with a known layout, when the weather forecast or actual measurement provides the dynamic wind speed curve (such as gradually changing wind) at the wind farm inlet: For the first row of wind turbines, directly extract the rate of change of the wind speed curve ( ) and transition time ( By inputting the model from Example 1, its power overshoot amplitude can be predicted.

[0064] For the second and subsequent rows of wind turbines, the inlet wind conditions must first be assessed. This can be done using a simple wake model (such as the Jensen model) or by calculating the wake conditions of the first row of turbines. and (As an approximation, or through a small amount of pre-computed database mapping). Then... , , , Input and implement a modified model applicable to simple wind farms to predict their power overshoot amplitude.

[0065] By combining the predicted power overshoot amplitude with the steady-state power prediction value, a more accurate range of power output extreme values ​​under dynamic operating conditions can be obtained, providing key input for grid dispatch and energy storage system control.

[0066] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0067] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting wind turbine power based on dynamic inflow conditions, characterized in that, Includes the following steps: S1. Obtain the geometric parameters of the target wind turbine, establish a three-dimensional unsteady numerical simulation model of the wind turbine using CFD software, divide the computational domain and generate a mesh, set inlet velocity boundary conditions, outlet pressure boundary conditions and symmetric boundary conditions, solve the transient Reynolds-averaged Navier-Stokes equations using a turbulence model, and verify the accuracy of the numerical method by verifying mesh independence and comparing with data from published literature. S2. Based on the validated numerical simulation model, the system systematically simulates the gradual wind speed condition, extreme gust condition and disturbed wind condition. Unsteady numerical simulation is performed on each set of condition parameters. The inlet wind speed-time curve and the wind turbine output power-time curve are recorded to build a dynamic inflow condition database. S3. Extract the wind speed change rate from the simulation results. and transition time As an input feature, the power overshoot amplitude As the output target, the power overshoot amplitude is defined as ,in This represents the maximum output power of the wind turbine during the duration of dynamic wind conditions. The instantaneous power value at the end of the dynamic wind condition; a training sample library for the power prediction model of a single wind turbine is constructed; S4, based on wind speed change rate and transition time Input power overshoot amplitude For the output, an error backpropagation neural network is used to establish a nonlinear mapping relationship between the input features and the output target, and a fast prediction model for the power overshoot amplitude of a single wind turbine is obtained. S5. Establish a numerical model of a dual-wind turbine wind farm and extract the overshoot amplitude of the incoming wind speed at the downstream wind turbine inlet. and average wind speed As a wake correction parameter, the input features are expanded to , , and With four parameters, a BP neural network was retrained to establish a prediction model for the power overshoot amplitude of downstream wind turbines in wind farms affected by upstream wakes.

2. The method according to claim 1, characterized in that, In step S1, the computational domain is divided as follows: it extends from 6R upstream to 10R downstream of the wind turbine rotor plane, with the radial boundary 5R away from the rotor axis, where R is the rotor radius. The rotor rotation is simulated using a sliding mesh technique, and the rotating domain and the stationary domain are connected by an interface. The Realizable k-ε model is selected as the turbulence model, and the SIMPLE algorithm is used to solve the transient RANS equations. The time step is set to the physical time corresponding to the rotor rotation of 0.5° to 1°. The mesh is an unstructured tetrahedral mesh, and local refinement is applied to the blade surface, leading edge, trailing edge, tip / root region, and near-wake region.

3. The method according to claim 1, characterized in that, In step S2, the core parameters of the gradual wind speed condition include the initial wind speed, the final wind speed, and the transition time. The wind speed changes linearly with time during the transition time. The extreme gust condition is defined according to the IEC61400-1 standard, and its core parameters include the periodic influence factor and the gust duration, simulating the process of the wind speed rising rapidly and then recovering. The core parameters of the disturbed wind condition include the average wind speed, the disturbance amplitude, and the disturbance frequency. The wind speed fluctuates periodically with time, simulating the process of the incoming wind speed continuously disturbing near the average wind speed in a natural wind environment.

4. The method according to claim 1, characterized in that, In step S3, the wind speed change rate Defined as (final wind speed - initial wind speed) / transition time, in m / s²; the transition time The duration of the wind speed change, in seconds; in step S5, the amplitude of the incoming wind speed overshoot. Defined as ,in This represents the maximum wind speed at the downstream wind turbine inlet. The average wind speed is the instantaneous wind speed value at the monitoring point at the end of the dynamic wind condition period. The average wind speed at the downstream wind turbine inlet during dynamic wind conditions is used to characterize the speed loss level caused by the upstream wake.

5. The method according to claim 1, characterized in that, In step S4, the BP neural network includes one input layer, at least one hidden layer, and one output layer; the activation function of the hidden layer neurons is the sigmoid function; the loss function is the mean squared error. The training algorithm uses the Levenberg-Marquardt algorithm; the overfitting prevention strategy uses early stopping, which randomly divides all samples into training set and validation set, and terminates training early when the validation set error no longer decreases in multiple consecutive iterations; in a preferred embodiment, the network structure is 1 input layer, 3 hidden layers and 1 output layer, the ratio of training set to validation set samples is about 85%:15%, and the model's determination coefficient reaches above 0.

953.

6. The method according to claim 1, characterized in that, In step S5, the axial spacing between the upstream and downstream wind turbines in the numerical model of the dual-wind turbine wind farm is set to 12 times the rotor diameter, and they are arranged coaxially. Monitoring points are set at key locations in the incoming flow in front of the downstream wind turbine rotor to record the inlet wind speed-time variation curve. The input features of the corrected BP neural network are four-dimensional vectors, and the output target is still the power overshoot amplitude. The network structure is adjusted accordingly based on the input dimension. After introducing the trail correction parameter, the model's determination coefficient is increased to over 0.

98.

7. The method according to claim 1, characterized in that, For the second row and subsequent wind turbines in a wind farm, the overshoot amplitude of the incoming wind speed at the downstream wind turbine inlet is obtained through a simple wake model or mapping from a small amount of pre-calculated database. and average wind speed The approximate value will be , , and The corrected prediction model is input to predict its power overshoot amplitude; the predicted power overshoot amplitude is combined with the steady-state power prediction value to obtain the extreme range of power output under dynamic operating conditions, providing input for grid dispatch and energy storage system control.

8. The method according to claim 1, characterized in that, In step S4, the machine learning method for constructing the prediction model can be replaced by any one of support vector regression, random forest, XGBoost, long short-term memory neural network or convolutional neural network. When there is sufficient sample data, the above models can establish a nonlinear mapping relationship between input features and power overshoot amplitude, and thus replace the BP neural network to achieve the purpose of this invention.

9. The method according to claim 1, characterized in that, In step S2, the numerical simulation method for generating training samples can be replaced by any one of the following: large eddy simulation, actuated disk model, actuated line model, vortex method, or blade element-momentum theory modified model; the wind speed input sequence for dynamic inflow conditions can also be constructed based on wind farm anemometer data, lidar anemometer data, or SCADA operation data; in step S5, the inflow wind speed overshoot amplitude at the downstream wind turbine inlet... and average wind speed The acquisition method can be replaced by an engineering wake model, a data-driven wake model, or a correction method based on measured downstream wind speed; the above alternatives are all equivalent alternatives to the technical solution of this invention.

Images