Wind power generation prediction method and system based on digital twin energizing, and computer equipment

Abstract

The invention relates to the technical field of wind power generation, in particular to a wind power generation prediction method and system based on digital twin energizing and computer equipment. Comprising the following steps: establishing a blade stress-strain physical model, and quantifying a blade load and fatigue damage by combining aerodynamics and a thin plate bending theory; integrating weather, equipment parameters and operation data, and building a digital twinborn platform of a multi-task learning framework; executing the main task and the auxiliary task in parallel; theoretical generating capacity is calculated through a wind energy conversion equation, and a predicted value is corrected by fusing RNN; and selecting an optimal predicted value based on a confidence interval, and dynamically adjusting a yaw angle and a pitch angle in combination with an extreme weather risk level to form closed-loop control. According to the scheme, the power generation prediction error is reduced by 18%, the service life of the blade is prolonged by 28%, extreme weather early warning is brought forward by 72 hours, and the power generation efficiency and the equipment safety are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and specifically relates to a wind power generation prediction method, system and computer device empowered by digital twin. Background Art

[0002] In the process of accelerating the global energy transformation, wind power generation, as a key representative of clean energy, has become increasingly important. In recent years, the installed capacity of wind power generation has continued to grow rapidly, becoming an important force in alleviating the energy crisis and addressing climate change. However, the current wind power generation technology still faces many challenges in practical applications.

[0003] The existing wind power generation power prediction methods have obvious deficiencies. On the one hand, most traditional prediction models are based on simple statistical analysis or single data sources, and do not fully consider the complexity of meteorological factors. For example, only relying on historical wind speed data for prediction ignores the comprehensive influence of atmospheric temperature, humidity, air pressure, etc. on wind speed and power generation efficiency. The meteorological conditions vary greatly in different regions, and these models are difficult to adjust adaptively, resulting in limited prediction accuracy. During periods of variable weather, the prediction error can be as high as 20%-30%, seriously affecting the accuracy of power dispatching. On the other hand, when predicting, the existing technology often ignores the impact of the equipment's own state on the power generation power. Hidden faults such as wear and fatigue of key components of the wind turbine will reduce the wind energy conversion efficiency, but traditional methods cannot monitor the equipment's health status in real time and incorporate it into the prediction model, making it difficult to give early warnings of faults, resulting in an increased probability of sudden equipment failures and high maintenance costs. The power generation losses caused by equipment failures reach hundreds of millions of yuan every year. In addition, in the face of extreme weather, the existing technology is difficult to accurately predict its path and intensity, and the wind farm cannot take preventive measures in advance, easily causing equipment damage and power supply interruption, resulting in serious economic losses and safety hazards. Summary of the Invention

[0004] The purpose of the present invention is to provide a wind power generation prediction method, system and computer device empowered by digital twin, improve the accuracy of wind power generation power prediction, realize real-time monitoring and optimal control of the wind power generation system, reduce the risk of equipment failure, and improve power generation efficiency and economic benefits.

[0005] The technical solution of the present invention is as follows:

[0006] One of the technical solutions of the present invention is to provide a wind power generation prediction method empowered by digital twin, including:

[0007] Based on the thin plate bending theory and the aerodynamic equation, establish a stress-strain physical model of the wind turbine blade;

[0008] Considering the meteorological characteristics of the wind power generation area, the structure, performance, and operational characteristics of the wind power generation system, a multi-task learning framework is built to form a digital twin platform for the wind power generation system;

[0009] The following parallel tasks are performed through the multi-task learning framework of the digital twin platform of the wind power system:

[0010] Main task: Use the digital twin platform to predict the wind farm power generation in the next 0-24 hours. The input is the node characteristics of the spatiotemporal graph network and the historical power data of wind turbines. The output is the predicted power generation value.

[0011] Auxiliary tasks:

[0012] a. Predict the fatigue life of turbine blades based on a particle filter algorithm combined with a stress-strain physical model;

[0013] b. Calculate wind speed prediction noise and equipment status uncertainty using a particle filter algorithm to obtain a probability confidence interval for the revised power generation prediction value;

[0014] c. Use Bezier curves to fit historical wind field movement trajectories, combine them with meteorological forecast data to predict the path of extreme weather, and generate a probability distribution map of risk areas.

[0015] The theoretical power generation is calculated using the wind energy conversion efficiency equation. This is combined with the predicted power generation value and the fatigue life data of the turbine blades and input into the recurrent neural network of the digital twin platform. The output layer is the revised predicted power generation value.

[0016] According to the probability confidence interval, the power generation prediction value with the highest confidence is selected, and the turbine yaw angle and blade pitch angle are adjusted in real time based on the probability distribution map of the risk area and the power generation prediction value with the highest confidence.

[0017] As a further optimization of the first aspect, the stress-strain physical model of the wind turbine blade is:

[0018] D▽ 4 w(x,y)=q(x,y);

[0019] Where D is the bending stiffness of the plate,

[0020] in,

[0021] The load q(x,y) exerted by the wind on the blade is calculated by the lift force F on the blade. L and resistance F D Decomposed to obtain the lift force F on the blade L and resistance F D for:

[0022]

[0023] Among them, C L and C D are the lift coefficient and the drag coefficient respectively, which are determined by methods such as wind tunnel experiments and numerical simulations. S is the blade projected area, v is the wind speed, and ρ 空 is the air density;

[0024] The stress-strain physical model of the wind turbine blade solves the deflection distribution by the finite element method.

[0025] As a further optimization of the first aspect, the boundary conditions solved by the stress-strain physical model of the wind turbine blade include simply supported edges, fixed edges and free edges;

[0026] Among them, the fixed boundary is at the blade root of the wind turbine blade, where the deflection w = 0 at the blade root, and the rotations around the x and y directions

[0027] Among them, the free edge is at the blade tip of the wind turbine blade, and the boundary condition expression is M x = 0, M y = 0, Q x = 0, Q y = 0, where M x , M y are bending moments, and Q x , Q y are shear forces. The relationships between the bending moments and shear forces and the deflection w(x, y) are obtained from the internal force-displacement relationships in the thin plate bending theory.

[0028] As a further optimization of the first aspect, the digital twin platform adopts a microservices architecture, including a data layer, a model layer, an application layer, a virtual-real interaction interface and a cloud-edge collaboration architecture;

[0029] Among them, the data layer processes real-time stream data by an Apache Kafka cluster and stores historical data in a columnar manner by HBase;

[0030] The model layer is deployed in a containerized manner through Docker Swarm, and TensorFlow Serving provides gRPC inference services; the application layer realizes 3D visualization based on D3.js and WebGL, and supports dynamic rendering of the blade stress nephogram;

[0031] The virtual-real interaction interface calls the ABAQUS / CAE script through REST API for finite element calculation, and defines the data interaction format using JSONSchema;

[0032] The edge computing node deploys a lightweight model to preprocess high-frequency vibration data, and the cloud server performs complex inference tasks to achieve a cloud-edge collaborative architecture.

[0033] As a further optimization of the first aspect, the input layer of the main task includes meteorological data, system parameters of the wind power system, and operation data; it also includes the historical sequence of power generation, that is, the power generation data of the previous 24 hours; the encoder of the main task uses a bidirectional LSTM layer to extract time series features. The decoder of the main task uses a Transformer decoder to generate future 24-step predictions; the output layer of the main task uses a Dense layer to output the power prediction value, and the activation function is linear;

[0034] The output result of the main task is the power generation power prediction value with a 15-minute time step in the future The prediction error index, that is, the root mean square error, needs to be controlled at ≤0.15 MW.

[0035] As a further optimization of the first aspect, the input layer of the auxiliary task a includes obtaining the blade stress distribution from the stress-strain model, as well as the wind speed fluctuation and temperature change in the environmental parameters, and the fatigue limit σ of the blade -1 ;

[0036] Apply the particle filter algorithm to calculate the estimated value of the remaining fatigue life of the blade, that is:

[0037]

[0038] Among them, is the estimated value of the remaining fatigue life of the blade, ω i is the weight of the i-th particle, L i The fatigue life value represented by the i-th particle, and M is the number of particles.

[0039] As a further optimization of the first aspect, the input data of the auxiliary task b includes wind speed prediction noise, equipment state uncertainty, and the mean and standard deviation of historical prediction errors;

[0040] Apply the particle filter algorithm to estimate the probability distribution of the corrected power generation power prediction value; according to the estimated probability distribution, determine the confidence interval of the power generation power prediction value at a specified confidence level;

[0041] The output result of the auxiliary task b is the 95% confidence interval of the corrected power generation power prediction value Among them, the width of the confidence interval needs to satisfy

[0042] As a further optimization of the first aspect, the input data of the auxiliary task c includes the historical trajectory of extreme weather, meteorological forecast data, and terrain data around the wind farm;

[0043] The Bezier curve fitting algorithm is used to combine meteorological forecast data with the Bezier curve fitting results to predict the future path of extreme weather. According to the geographical location of the wind power generation system and the influence range of extreme weather, risk areas are divided.

[0044] The output result of auxiliary task c is the extreme weather path based on Bezier curve fitting and generates a risk area probability distribution map.

[0045] As a further optimization of the first aspect, the theoretical power generation P th The calculation formula is:

[0046]

[0047] Among them, P th is the theoretical power generation, A is the swept area of the fan, which is determined by the blade radius R, that is, A = πR 2 , v is the wind speed, and C p is the wind energy utilization coefficient;

[0048] The predicted power generation value of the main task The estimated remaining fatigue life of the blade of auxiliary task a And the theoretical power generation P th Are integrated to form a feature vector:

[0049]

[0050] The feature vector is input into the trained RNN model, and the corrected predicted power generation value is output

[0051] As a further optimization of the first aspect, an error compensation mechanism is added to the theoretical power generation, and dynamic compensation is performed in combination with the output of the RNN. The calculation formula is:

[0052]

[0053] Among them, is the predicted power generation value after compensation, and γ is the compensation coefficient.

[0054] The second technical solution of the present invention is to provide a fan intelligent early warning system based on an adaptive scheme, including:

[0055] Physical modeling module: used to establish a stress-strain physical model of the wind turbine blade based on the thin plate bending theory and the aerodynamic equation;

[0056] Digital Twin Platform: It includes a multi-task learning framework that integrates a main task branch and an auxiliary task branch. The main task branch uses a spatio-temporal sequence model to predict power generation, and the auxiliary task branch performs blade fatigue life prediction, confidence interval calculation, and extreme weather path prediction through the particle filter algorithm and Bezier curve fitting respectively;

[0057] Data Fusion Module: It calculates the theoretical power generation using the wind energy conversion efficiency equation, combines the predicted power generation value and the fatigue life data of the turbine blades, and inputs them into the recurrent neural network of the digital twin platform. The output layer is the corrected predicted power generation value;

[0058] Control Optimization Module: It is used to select the predicted power generation value with the highest confidence according to the probability confidence interval, and adjust the yaw angle and blade pitch angle of the turbine in real time based on the risk area probability distribution map and the predicted power generation value with the highest confidence.

[0059] The third aspect of the technical solution of the present invention is to provide a computer device, which includes: a processor and a memory. At least one instruction, at least one program, a code set, or an instruction set is stored in the memory, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the wind power generation prediction method based on digital twin empowerment as described in the first aspect of the technical solution.

[0060] The beneficial effects brought by the technical solutions provided in the embodiments of the present application at least include the following

[0061] Beneficial Effects:

[0062] Through the deep integration of digital twin and multi-task learning, the prediction accuracy of wind power generation and the equipment management level have been significantly improved. Its core advantages are reflected in three aspects: First, the prediction ability breakthrough. Through the multi-task learning framework built by digital twin, the 24-hour power generation prediction error is reduced from 0.23MW of the traditional method to 0.15MW, providing an accurate basis for power dispatching; Second, the innovation of equipment health management. Based on the blade stress-strain model established by the thin plate bending theory, combined with the particle filter algorithm to predict the fatigue life in real time, dynamically adjust the yaw angle and pitch angle, extend the blade life by 28%, and reduce the number of fault shutdowns by 75%; Third, the intelligent response to extreme weather. Using Bezier curve fitting to fit the wind field trajectory, combined with GIS to generate a risk probability distribution map, shorten the path prediction error from 50km to 25km, and advance the warning time to 72 hours, significantly reducing the equipment losses caused by extreme weather.

[0063] In terms of economic benefits, through multi-task collaborative optimization, this technology has achieved a 15% increase in power generation efficiency, a 22% reduction in operation and maintenance costs, and an annual income increase of approximately 300,000 yuan per single wind turbine. At the same time, based on the dynamic decision-making mechanism of the confidence interval, the curtailment of wind power due to prediction deviation has been reduced, and the annual power generation has increased by 5%. Overall, a closed-loop system of "accurate prediction - intelligent control - benefit optimization" has been formed, providing a systematic solution for the construction of intelligent wind farms. Description of the Drawings

[0064] Figure 1 It is a schematic diagram of the overall process of the wind power generation prediction method empowered by digital twin;

[0065] Figure 2 It is a detailed flowchart of step S100 of the wind power generation prediction method empowered by digital twin;

[0066] Figure 3 It is a detailed flowchart of step S200 of the wind power generation prediction method empowered by digital twin;

[0067] Figure 4 It is a detailed flowchart of step S300 of the wind power generation prediction method empowered by digital twin;

[0068] Figure 5 It is a detailed flowchart of step S400 of the wind power generation prediction method empowered by digital twin;

[0069] Figure 6 It is a detailed flowchart of step S500 of the wind power generation prediction method empowered by digital twin. Detailed Embodiments

[0070] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0071] In the field of wind power generation, problems such as low accuracy of power generation prediction, insufficient correlation prediction of equipment status, and difficulty in coping with extreme weather are faced. To solve the above problems, please refer to Figure 1 , which shows a wind power generation prediction method empowered by digital twin provided by an embodiment of the present invention. The method includes:

[0072] S100: Based on the thin plate bending theory and the aerodynamic equation, establish a stress-strain physical model of the wind turbine blade;

[0073] S200: Consider the meteorological characteristics of the wind power generation area, the structure, performance, and operation characteristics of the wind power generation system, build a multi-task learning framework, and form a digital twin platform for the wind power generation system;

[0074] S300: Execute the following parallel tasks through the multi-task learning framework of the digital twin platform for the wind power generation system:

[0075] Main task: Use the digital twin platform to predict the power generation power of the wind farm in the next 0-24 hours. The input is the node characteristics of the spatio-temporal graph network and the historical power data of the wind turbines, and the output is the predicted value of the power generation power;

[0076] Auxiliary tasks:

[0077] a. Based on the particle filter algorithm, combine the stress-strain physical model to predict the fatigue life of the turbine blades;

[0078] b. Calculate the wind speed prediction noise and equipment status uncertainty through the particle filter algorithm to obtain the probability confidence interval of the corrected power generation power prediction value;

[0079] c. Use Bezier curve fitting to fit the historical wind field movement trajectory, combine meteorological forecast data to predict the path of extreme weather, and generate a probability distribution map of the risk area.

[0080] S400: Calculate the theoretical power generation by using the wind energy conversion efficiency equation, and combine the power generation power prediction value and the fatigue life data of the turbine blades, and input them into the recurrent neural network of the digital twin platform. The output layer is the corrected power generation power prediction value;

[0081] S500: According to the probability confidence interval, select the power generation power prediction value with the highest confidence level, and adjust the yaw angle and blade pitch angle of the turbine in real time based on the probability distribution map of the risk area and the power generation power prediction value with the highest confidence level.

[0082] This technical solution establishes a stress-strain physical model of the wind turbine blade based on the thin plate bending theory and the aerodynamic equation, builds a digital twin platform by combining meteorological and power generation system characteristics, uses its multi-task learning framework to predict power generation power, blade fatigue life, confidence interval, and extreme weather path respectively, then calculates the theoretical power generation and corrects the power generation power prediction value by combining relevant data, and finally adjusts the yaw angle and blade pitch angle of the turbine according to the calculation results, thereby improving the accuracy of power generation power prediction, optimizing equipment operation, reducing equipment failure risk, reducing extreme weather damage, improving power generation efficiency and economic benefits, and ensuring the safe and stable operation of the power generation system. Each step will be elaborated in detail below.

[0083] According to the thin plate bending theory, clarify the geometric parameters and material properties of the blade. Combining with the aerodynamic equation, considering factors such as different wind speeds, wind directions, and blade rotation angles, establish a mathematical model of blade stress-strain to quantify the deformation and fatigue damage of the blade under aerodynamic loads.

[0084] Please refer to Figure 2 , which shows a flowchart of an exemplary wind power generation prediction method S100 empowered by digital twin according to the present application, and its content includes:

[0085] S110: Considering the wind force, calculate the load acting on the blade by combining with the aerodynamic equation.

[0086] The load acting on the blade by the wind force is q(x, y), and q(x, y) is related to factors such as wind speed v and air density ρ 空 and so on. According to the aerodynamic principle, the lift force F L and drag force F D on the blade are expressed as:

[0087]

[0088] Among them, C L and C D are the lift coefficient and drag coefficient respectively, which are determined by methods such as wind tunnel experiments and numerical simulations. S is the projected area of the wind turbine blade, v is the wind speed, and ρ 空 is the air density.

[0089] Decompose the lift force and drag force into each direction of the blade to obtain the distributed load q(x, y) acting on the blade.

[0090] S120: Establish a blade stress-strain model by using the thin plate bending theory.

[0091] When establishing the stress-strain model of the wind turbine blade, select the thin plate bending theory as the basis. The thin plate bending theory is a mechanical analysis theory for thin plates under transverse loads. The wind turbine blade can be approximately regarded as a thin plate in terms of structure and force characteristics.

[0092] Simplify the blade into an anisotropic thin plate, and use the Mindlin plate theory to establish a blade stress-strain model. The middle surface of the blade is in the x-y plane, and the direction perpendicular to the middle surface is the z direction. The load acting on the blade by the wind force is q(x, y), and the deflection generated under the action of q(x, y) is w(x, y). The blade stress-strain model established based on the thin plate bending theory is:

[0093] D▽ 4 w(x, y) = q(x, y);

[0094] where D is the bending stiffness of the thin plate, E is the elastic modulus of the wind turbine blade, h is the thickness distribution of the wind turbine blade, and μ is the Poisson's ratio of the wind turbine blade.

[0095] where

[0096] By solving this equation, the deflection distribution of the blade under different loads can be obtained.

[0097] To solve the above equation, the boundary conditions of the blade need to be clarified.

[0098] In a possible implementation, the set boundary conditions include simply supported edges, fixed edges, and free edges, etc. For a wind turbine blade, the root of the blade is usually approximately a fixed boundary condition, that is, the deflection w at the root of the blade is 0, and the rotations around the x and y directions The tip of the blade is generally approximately a free edge, that is, the blade edge is not subjected to transverse forces and bending moments. The corresponding boundary condition expressions are M x = 0, M y = 0, Q x = 0, Q y = 0, where M x , M y are bending moments, Q x , Q y are shear forces. The relationships between the bending moment and shear force and the deflection w(x,y) can be obtained from the internal force-displacement relationship in the thin plate bending theory.

[0099] Due to the complex shape and actual stress conditions of the blade, it is generally difficult to obtain an analytical solution for the above blade stress-strain model. In a possible implementation, numerical methods such as the finite difference method and the finite element method are used to solve the equation.

[0100] Exemplarily, when solving with the finite element method, first divide the blade into a grid, dividing it into a large number of small elements, such as triangular elements or quadrilateral elements. Each element has several nodes, and the relationship between the deflection w of each point in the element and the node displacement is represented by an interpolation function. Then, based on the principle of virtual work or the variational principle, establish the stiffness equation of each element, that is, the relationship between the element nodal forces and the node displacements. Assemble the stiffness equations of all elements to form the overall stiffness equation of the entire blade structure. Finally, combined with the set boundary conditions, solve the overall stiffness equation to obtain the deflection values of each node of the blade. Through these deflection values, and then using the geometric equation and physical equation in the thin plate bending theory, further calculate the strain and stress distributions of each point of the blade. During the finite element solution process, the density of the grid will affect the calculation accuracy and calculation efficiency, and it needs to be reasonably selected according to the actual situation.

[0101] S200 constructs a digital twin platform for wind power generation under a multi-task learning framework by integrating meteorological data, power generation system parameters, and operation data. This platform adopts a microservices architecture to form an intelligent decision-making platform integrating data fusion, multi-task collaborative optimization, and real-time monitoring.

[0102] Please refer to Figure 3 , which shows a flowchart of an exemplary wind power generation prediction method S200 empowered by digital twin according to this application. Its content includes:

[0103] S210: Obtain meteorological data, system parameters of the wind power generation system, and operation data.

[0104] Specifically, collect meteorological data of the wind power generation area through means such as meteorological stations and satellite remote sensing, including wind speed v, wind direction α, temperature T, humidity H, and air pressure P. Collect system parameters of the wind power generation system, such as tower barrel height H t , number of blades n, blade geometric parameters such as blade radius R, etc.; operation data, such as generator output power P g , rotational speed n g , yaw angle θ yaw , pitch angle β, etc.

[0105] Integrate meteorological data, structural parameters, performance parameters, and operation data of the wind power generation system to form a unified data set.

[0106] Clean the collected data to remove outliers and missing values. Use methods such as interpolation and smoothing to preprocess the data to improve the quality and usability of the data

[0107] S220: Construct a multi-task learning framework.

[0108] When constructing the multi-task learning framework of the wind power generation system digital twin platform, clarifying the learning objectives is the primary task. The learning objectives are interrelated and each has important significance, jointly serving to improve the operation efficiency, reliability, and safety of the wind power generation system.

[0109] The multi-task learning objectives include tasks such as power generation power prediction, blade fatigue life prediction, calculation of the confidence interval of power generation power prediction values, and extreme weather path prediction.

[0110] Adopt a deep learning framework to design the architecture of the multi-task learning model. The model includes an input layer, a hidden layer, and an output layer. The input layer receives meteorological data, system parameters, and operation data. The hidden layer performs feature extraction and transformation, and the output layer outputs the prediction results of each task respectively.

[0111] In a possible implementation manner, the designed platform architecture includes:

[0112] Underlying Shared Layer: Convolutional Neural Network (CNN) or Transformer is used to process spatio-temporal sequence data (such as historical wind speed sequence, blade vibration signal).

[0113] Task-Specific Layer: Branch networks process different tasks respectively;

[0114] Among them, the main task branch: LSTM or Temporal Fusion Transformer (TFT) predicts power generation. Blade fatigue life branch: Combines stress-strain physical model with Recurrent Neural Network (RNN) to predict fatigue damage accumulation. Confidence interval branch: Bayesian Neural Network (BNN) quantifies prediction uncertainty. Extreme weather branch: Graph Neural Network (GNN) models the spatial correlation between wind field movement trajectory and meteorological elements.

[0115] Multi-Task Loss Function: Adopts weighted joint loss, and balances the training priorities of each task through hyperparameter optimization.

[0116] S230: Construction of a digital twin platform for a wind power generation system.

[0117] The digital twin platform adopts a microservices architecture, including a data layer, a model layer, an application layer, a virtual-real interaction interface, and a cloud-edge collaboration architecture.

[0118] In a possible implementation, the data layer processes real-time stream data by an Apache Kafka cluster, and stores historical data in HBase columnar storage. The model layer is deployed in a containerized manner through Docker Swarm, and TensorFlow Serving provides gRPC inference services. The application layer realizes 3D visualization based on D3.js and WebGL, and supports dynamic rendering of blade stress nephogram. The virtual-real interaction interface calls ABAQUS / CAE scripts through REST API for finite element calculation, and defines the data interaction format using JSON Schema. Edge computing nodes deploy lightweight models to preprocess high-frequency vibration data, and cloud servers execute complex inference tasks to achieve a cloud-edge collaboration architecture.

[0119] S300 relies on the multi-task learning framework of the digital twin platform to execute four major tasks in parallel: the main task predicts the power generation for the next 24 hours; auxiliary task a updates the blade fatigue life; auxiliary task b generates a confidence interval; auxiliary task c predicts the extreme weather path and divides the probability of risk areas. Each task realizes collaborative optimization through the attention mechanism and the weighted loss function.

[0120] Please refer to Figure 4 , which shows a flowchart of an exemplary wind power generation prediction method S300 empowered by digital twin of the present application, and its content includes:

[0121] S310: Execute the main task, i.e., wind farm power generation prediction, through the multi-task learning framework of the digital twin platform.

[0122] Achieve accurate prediction of the wind farm power generation for the next 0 - 24 hours, providing a decision-making basis for power dispatching, energy trading, and equipment operation and maintenance.

[0123] In a possible implementation, predict the power generation for the next 24 hours based on the TFT of the digital twin platform, capture spatio-temporal dependencies through the attention mechanism, and combine meteorological forecast data to improve the prediction accuracy.

[0124] The input layer of the main task includes meteorological data, system parameters of the wind power generation system, and operation data. It also includes the historical power generation sequence, i.e., the power generation data for the previous 24 hours. The encoder of the main task uses a bidirectional LSTM layer to extract time series features. The decoder of the main task uses a Transformer decoder to generate predictions for the next 24 steps. The output layer of the main task uses a Dense layer to output the power prediction value, and the activation function is linear.

[0125] The output result of the main task is the power generation prediction value with a 15 - minute time step. The prediction error index, i.e., the root mean square error, needs to be controlled at ≤0.15 MW.

[0126] S320: Execute the auxiliary task a, i.e., blade fatigue life prediction, through the multi-task learning framework of the digital twin platform.

[0127] Prevent overload damage and optimize the maintenance strategy by real-time evaluating the remaining fatigue life of the blade. Combine the particle filter algorithm with the stress-strain model of S130 to dynamically predict the blade fatigue life. The core lies in the construction of the state space model and the measurement update mechanism to achieve accurate tracking of the material degradation process.

[0128] In a possible implementation, the input layer of the auxiliary task a includes the blade stress distribution obtained through the stress-strain model established by S100, the wind speed fluctuation and temperature change in the environmental parameters, and the fatigue limit σ of the blade. -1 。

[0129] According to the characteristics and stress state of the blade material, select the Miner linear cumulative damage theory, and at the same time, determine the fatigue life curve (S-N curve) of the blade material through fatigue tests or referring to relevant literature.

[0130] Apply the particle filter algorithm, regard the fatigue life of the blade as a state variable, and establish a state space model. The state equation in the state space model describes the dynamic change of the fatigue life, and the observation equation in the state space model describes the relationship between the observable stress-strain data and the fatigue life. Initialize a set of particles, where each particle represents a possible fatigue life value. Assign an initial weight to each particle, usually with equal initial weights. At each time step, update the state of the particles according to the state equation and calculate the weights of the particles according to the observation equation. Through the resampling operation, remove the particles with smaller weights and replicate the particles with larger weights to maintain the effectiveness of the particles. Calculate the estimated value of the remaining fatigue life of the blade based on the states and weights of the particles. That is

[0131]

[0132] where is the estimated value of the remaining fatigue life of the blade, ω i is the weight of the i-th particle, L i is the fatigue life value represented by the i-th particle, and M is the number of particles.

[0133] The output result of auxiliary task a, the estimated value of the remaining fatigue life of the blade (unit: equivalent operating hours).

[0134] S330: Execute auxiliary task b, that is, confidence interval calculation, through the multi-task learning framework of the digital twin platform.

[0135] Confidence interval calculation is to quantify the uncertainty of power generation prediction and provide a basis for risk assessment of the power system.

[0136] In a possible implementation, the input data of auxiliary task b includes wind speed prediction noise, equipment state uncertainty, mean and standard deviation of historical prediction errors. Among them, wind speed prediction noise and equipment state uncertainty are the sources of uncertainty. The wind speed prediction noise follows a normal distribution, and the equipment state uncertainty describes the component degradation state through a hidden Markov model (HMM).

[0137] Similar to auxiliary task a, initialize a set of particles, where each particle represents a possible power generation prediction value and related uncertainty parameters. At each time step, update the state and weights of the particles according to the extended state space model and perform the resampling operation. Estimate the probability distribution of the corrected power generation prediction value based on the states and weights of the particles. Determine the confidence interval of the power generation prediction value at the specified confidence level according to the estimated probability distribution.

[0138] The output result of auxiliary task b is the 95% confidence interval of the corrected power generation prediction value where the width of the confidence interval needs to satisfy

[0139] S340: Execute the auxiliary task c, i.e., extreme weather warning, through the multi-task learning framework of the digital twin platform.

[0140] The extreme weather warning is to predict the paths of extreme weather such as typhoons and thunderstorms 72 hours in advance and generate a probability distribution map of risk areas.

[0141] In a possible implementation, the input data of the auxiliary task c includes the historical trajectories of extreme weather, meteorological forecast data, and the terrain data around the wind farm. Among them, the historical trajectories of extreme weather are the path data of extreme weather in the past 10 years.

[0142] According to the shape of the historical wind field trajectory, select appropriate control points. The number and position of the control points will affect the fitting effect of the Bezier curve. Adopt the Bezier curve fitting algorithm, and determine the parameters of the Bezier curve according to the control points and the historical wind field trajectory data. Obtain the meteorological forecast data of extreme weather released by the meteorological department, combine the meteorological forecast data with the fitting result of the Bezier curve, predict the future path of extreme weather, and divide the risk areas according to the geographical location of the wind power generation system and the influence range of extreme weather.

[0143] The output result of the auxiliary task c is the extreme weather path based on the Bezier curve fitting and generate a probability distribution map of risk areas.

[0144] Based on the main task, auxiliary task a, auxiliary task b, and auxiliary task c, design the architecture of the multi-task learning model using the deep learning framework on the digital twin platform to achieve collaborative learning of multiple tasks. The deep learning framework adopted by the digital twin platform includes convolutional neural network (CNN), recurrent neural network (RNN) and its variants (such as LSTM, GRU), and graph neural network (GNN).

[0145] S400 calculates the theoretical power generation through the wind energy conversion equation, fuses the predicted value of the main task, the fatigue life and the theoretical value of the auxiliary task a into a feature vector, and inputs it into the RNN model for dynamic correction. Combine the theoretical and measured deviations to improve the prediction accuracy, and finally output the corrected power generation value.

[0146] Please refer to Figure 5 , which shows a flowchart of an exemplary wind power generation prediction method S400 based on digital twin empowerment of the present application, and its content includes:

[0147] S410: Calculate the theoretical power generation.

[0148] Establish a theoretical power generation benchmark through the wind energy conversion efficiency equation and dynamically update the benchmark value in combination with real-time meteorological parameters. According to the wind energy conversion efficiency equation, the theoretical power generation P th is expressed as:

[0149]

[0150] where P th is the theoretical power generation, A is the swept area of the wind turbine, which is determined by the blade radius R, i.e., A = πR 2 , v is the wind speed, and C p is the wind energy utilization coefficient.

[0151] S420: Perform data feature fusion on the data obtained from the main task and each auxiliary task, as well as the theoretical power generation.

[0152] Integrate the predicted power generation value of the main task the estimated remaining fatigue life of the blade of auxiliary task a and the theoretical power generation P th to form a feature vector:

[0153]

[0154] Arrange the feature vectors in a time series to form the sequence data input to the recurrent neural network (RNN).

[0155] S430: Correct the power generation prediction based on the recurrent neural network of the digital twin platform.

[0156] Based on the RNN network, capture the long-term dependence relationship in historical data and perform dynamic correction in combination with the theoretical power generation. Input the preprocessed time series data into the trained RNN model, and output the corrected predicted power generation value

[0157] In a possible implementation, design an error compensation mechanism, and perform dynamic compensation by calculating the deviation between the theoretical power generation and the predicted value of the main task in combination with the output of the RNN:

[0158]

[0159] where γ is the compensation coefficient.

[0160] S500 Select the optimal predicted value based on the confidence interval, dynamically adjust the turbine operation parameters in combination with the extreme weather risk level; optimize the pitch angle using sequential quadratic programming based on the fatigue life constraint, and set the safety threshold at the same time. The control instructions are sent down through the OPCUA protocol, including three-level verification and blockchain evidence storage, forming a closed-loop control system of "prediction - decision - execution - feedback".

[0161] Please refer to Figure 6 , which shows a flowchart of an exemplary wind power prediction method S500 empowered by digital twin according to the present application, and its content includes:

[0162] S510: Select the optimal prediction value by analyzing the confidence interval.

[0163] Based on the 95% confidence interval generated by the auxiliary task b, combined with real-time meteorological data and equipment status, select the predicted power generation value with the highest confidence.

[0164] In a possible implementation manner, the selection method of the predicted power generation value with the highest confidence includes:

[0165] Evaluate the interval coverage according to the confidence interval width , that is, verify the probability that the interval contains the true value through historical data, and the requirement is ≥95%.

[0166] If the confidence interval is symmetric, take the mean value; if there is asymmetric uncertainty, use weighted average to obtain the optimal value.

[0167] Combined with the extreme weather probability of the auxiliary task c, when the risk area probability > 30%, automatically increase the confidence to 99% and shrink the interval width.

[0168] S520: Analyze the dynamic probability distribution of the risk area and determine the priority of the current wind farm affected by extreme weather.

[0169] Determine the priority of the current wind farm affected by extreme weather by analyzing the risk area map generated by the auxiliary task c.

[0170] In a possible implementation manner, the determination method of the priority of the current wind farm affected by extreme weather includes:

[0171] Based on the path probability density predicted by the Bezier curve, divide into three levels of risks, including: low risk (probability < 10%): normal operation. Medium risk (10% ≤ probability < 30%): start the pre-adjustment strategy. High risk (probability ≥ 30%): trigger emergency protection.

[0172] Adjust the control parameter weights according to the risk level, where the low risk weight is 1.0, the medium risk weight is 0.8, and the high risk weight is 0.5. The parameter weights are used for the adjustment amplitude of the subsequent yaw angle and pitch angle.

[0173] S530: Based on the risk area probability distribution map and the predicted power generation value with the highest confidence, adjust the yaw angle and blade pitch angle of the turbine in real time.

[0174] The digital twin platform predicts the future 6-hour wind field movement path based on the Bezier curve generated by auxiliary task c, and optimizes the control point parameters by combining historical data and real-time weather forecasts. The center position of the wind field is fitted by a cubic Bezier curve, and the path probability is corrected considering wind shear and terrain shielding effects. When calculating the target yaw angle, the system obtains the main wind direction of the wind farm in real time, combines it with the current azimuth angle of the tower barrel, and determines the optimal upwind angle to maximize the swept area. An adaptive PID controller is used to track the target yaw angle, and the control parameters are dynamically adjusted according to the risk level: conventional parameters (proportional coefficient 0.5, integral coefficient 0.1) are used at low risk, and damping is increased (derivative coefficient is increased to 0.1) at high risk to reduce mechanical vibration. The yaw system adopts a dual-redundancy drive design, and automatically switches to the standby motor when a single motor fails, ensuring that the angle adjustment accuracy is within the range of ±0.5°.

[0175] The digital twin platform obtains the remaining fatigue life data of the blades of auxiliary task a in real time, and calculates the fatigue damage rate by combining the current wind speed and the tip speed ratio. An optimization model with the goal of maximizing power generation is established, and the constraint conditions include that the fatigue damage rate does not exceed the safety threshold (0.001 / hour) and the pitch angle adjustment rate ≤ 2° / minute. The sequential quadratic programming algorithm is used to solve the nonlinear optimization problem, and the pitch angle is dynamically adjusted to balance power generation efficiency and equipment safety. When it is detected that the blade stress is close to the fatigue limit (>85%), a load reduction strategy is automatically triggered: the pitch angle is increased by 5°, and at the same time the tip speed ratio is reduced by 0.5, ensuring that the damage rate is reduced by more than 30%. The system monitors the blade strain in real time through fiber Bragg grating sensors, updates the optimization parameters every 100 ms, and forms a closed-loop control.

[0176] The digital twin platform converts the optimized yaw angle and pitch angle parameters into control commands, and issues them to the wind turbine controller through the OPCUA protocol via the digital twin platform. The command includes a three-level verification mechanism: first, check whether the parameters are within the safe range, second, compare the deviation between the theoretical calculation value and the actual execution value, and finally verify the matching of the control command and the equipment status. During the execution process, status data such as the current of the yaw motor and the pressure of the hydraulic system are collected in real time. When the current mutation exceeds 20% or the pressure is abnormal, an emergency stop procedure is automatically triggered. The system sets double safety thresholds: when the wind speed exceeds 25 m / s, the pitch angle is forced to be adjusted to 90°; when the remaining life of the blade is less than 100 hours, it automatically switches to the standby mode. All control commands and execution data are stored in the blockchain evidence storage system to ensure the traceability and immutability of operations.

[0177] This application also provides a wind turbine intelligent early warning system based on an adaptive scheme, including:

[0178] A physical modeling module: used to establish a stress-strain physical model of a wind turbine blade based on the thin plate bending theory and the aerodynamic equation;

[0179] Digital twin platform: including a multi-task learning framework, which integrates a main task branch and an auxiliary task branch. The main task branch uses a spatio-temporal sequence model to predict the power generation, and the auxiliary task branch performs blade fatigue life prediction, confidence interval calculation, and extreme weather path prediction through the particle filter algorithm and Bezier curve fitting respectively;

[0180] Data fusion module: calculates the theoretical power generation using the wind energy conversion efficiency equation, combines the predicted power generation value and the fatigue life data of the turbine blade, and inputs them into the recurrent neural network of the digital twin platform. The output layer is the corrected predicted power generation value;

[0181] Control optimization module: is used to select the predicted power generation value with the highest confidence according to the probability confidence interval, and adjust the yaw angle and blade pitch angle of the turbine in real time based on the risk area probability distribution map and the predicted power generation value with the highest confidence

[0182] This application also provides a computer device, which includes: a processor and a memory. At least one instruction, at least one program, a code set, or an instruction set is stored in the memory. The at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement a wind power generation prediction method empowered by digital twin.

[0183] The basic principles of this application have been described above in combination with specific embodiments. However, it should be noted that the advantages, benefits, effects, etc. mentioned in this application are only examples and not limitations. It cannot be considered that these advantages, benefits, effects, etc. are essential for each embodiment of this application. In addition, the above-disclosed specific details are only for illustrative and easy-to-understand purposes, rather than limitations. These details do not limit this application to necessarily adopt the above specific details to implement.

[0184] The block diagrams of the devices, apparatuses, equipment, and systems involved in this application are only illustrative examples and do not intend to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any way. Words such as "including", "comprising", "having", etc. are open-ended words, meaning "including but not limited to", and can be used interchangeably with each other. The words "or" and "and" used here refer to the word "and / or" and can be used interchangeably with each other, unless the context clearly indicates otherwise. The word "such as" used here refers to the phrase "such as but not limited to" and can be used interchangeably with each other.

[0185] It should also be noted that in the devices, equipment, and methods of the present application, each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent solutions of the present application.

[0186] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of the present application. Therefore, the present application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0187] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A wind power generation prediction method empowered by digital twin, characterized in that, Including: Based on the thin plate bending theory and aerodynamic equations, establish the stress-strain physical model of the wind turbine blade; Considering the meteorological characteristics of the wind power generation area, the structure, performance and operation characteristics of the wind power generation system, build a multi-task learning framework to form a digital twin platform for the wind power generation system; Execute the following parallel tasks through the multi-task learning framework of the digital twin platform of the wind power generation system: Main task: Use the digital twin platform to predict the power generation power of the wind farm in the next 0-24 hours. The input is the node characteristics of the spatio-temporal graph network and the historical power data of the wind turbine, and the output is the predicted value of the power generation power; Auxiliary tasks: a. Based on the particle filter algorithm, combine the stress-strain physical model to predict the fatigue life of the turbine blade; b. Calculate the wind speed prediction noise and equipment state uncertainty through the particle filter algorithm to obtain the probability confidence interval of the corrected power generation power prediction value; c. Use Bezier curve to fit the historical wind field movement trajectory, combine with meteorological forecast data to predict the path of extreme weather, and generate a probability distribution map of the risk area; Calculate the theoretical power generation by using the wind energy conversion efficiency equation, and combine with the power generation power prediction value and the fatigue life data of the turbine blade, and input it into the recurrent neural network of the digital twin platform. The output layer is the corrected power generation power prediction value; According to the probability confidence interval, select the power generation power prediction value with the highest confidence, and adjust the yaw angle and blade pitch angle of the turbine in real time based on the probability distribution map of the risk area and the power generation power prediction value with the highest confidence.

2. The wind power generation prediction method based on digital twin empowerment according to claim 1, wherein, The stress-strain physical model of the wind turbine blade is: D▽ 4 w(x, y) = q(x, y); where D is the bending stiffness of the thin plate, Among them, The load q(x,y) exerted by the wind on the blade is obtained by decomposing through the lift force F L and the drag force F D on the blade. The lift force F L and the drag force F D are as follows: Among them, C L and C D are the lift coefficient and the drag coefficient respectively, which are determined by methods such as wind tunnel experiments and numerical simulations. S is the projected area of the blade, v is the wind speed, and ρ 空 is the air density; The stress-strain physical model of the wind turbine blade solves the deflection distribution by the finite element method.

3. The wind power generation prediction method based on digital twin empowerment according to claim 2, wherein, The boundary conditions solved by the stress-strain physical model of the wind turbine blade include simply supported edges, fixed edges and free edges; Among them, the clamped boundary is at the blade root of the wind turbine blade, where the deflection w at the blade root is 0, and the rotations around the x and y directions Among them, the free edge is at the tip of the wind turbine blade, and the boundary condition expression is M x = 0, M y = 0, Q x = 0, Q y = 0, where M x , M y are bending moments, Q x , Q y are shear forces, and the relationships between the bending moment and the shear force and the deflection w(x, y) are obtained from the internal force-displacement relationships in the thin plate bending theory.

4. The wind power generation prediction method based on digital twin empowerment according to claim 1, wherein, The digital twin platform adopts a microservice architecture, including a data layer, a model layer, an application layer, a virtual-real interaction interface and a cloud-edge collaborative architecture; Among them, the data layer processes real-time stream data by the Apache Kafka cluster, and stores historical data in a columnar manner by HBase; The model layer is deployed in a containerized manner through Docker Swarm, and TensorFlow Serving provides gRPC inference services; the application layer realizes 3D visualization based on D3.js and WebGL, and supports dynamic rendering of the blade stress nephogram; The virtual-real interaction interface calls the ABAQUS / CAE script through the REST API for finite element calculation, and defines the data interaction format by JSON Schema; The edge computing node deploys a lightweight model to preprocess high-frequency vibration data, and the cloud server executes complex inference tasks to realize the cloud-edge collaborative architecture.

5. The wind power generation prediction method based on digital twin empowerment according to claim 4, characterized in that, The input layer of the main task includes meteorological data, system parameters of the wind power generation system, and operation data; it also includes the historical sequence of power generation, that is, the power generation data of the previous 24 hours; the encoder of the main task uses a bidirectional LSTM layer to extract time series features; the decoder of the main task uses a Transformer decoder to generate future 24-step predictions; the output layer of the main task uses a Dense layer to output the power prediction value, and the activation function is linear; The output result of the main task is the predicted power generation value with a 15-minute time step in the future. The prediction error index, i.e., the root mean square error, needs to be controlled at ≤0.15 MW.

6. The wind power generation prediction method based on digital twin empowerment according to claim 4, wherein The input layer of the auxiliary task a includes a stress-strain model to obtain the stress distribution of the blade, as well as the wind speed fluctuation and temperature change in the environmental parameters, and the fatigue limit σ of the blade -1 ; Apply the particle filter algorithm to calculate the estimated value of the remaining fatigue life of the blade, that is: Among them, is the estimated remaining fatigue life of the blade, ω i is the weight of the i-th particle, L i is the fatigue life value represented by the i-th particle, and M is the number of particles.

7. The wind power generation prediction method based on digital twin empowerment according to claim 4, characterized in that The input data of the auxiliary task b includes wind speed prediction noise, equipment state uncertainty, the mean and standard deviation of historical prediction errors; Apply the particle filter algorithm to estimate the probability distribution of the corrected power generation prediction value; according to the estimated probability distribution, determine the confidence interval of the power generation prediction value at a specified confidence level; The output result of the auxiliary task b is the 95% confidence interval of the corrected predicted power generation value wherein, the confidence interval width needs to satisfy 8. The wind power generation prediction method based on digital twin empowerment according to claim 4, wherein, The input data of the auxiliary task c includes the historical trajectory of extreme weather, meteorological forecast data, and topographic data around the wind farm; Adopt the Bezier curve fitting algorithm to combine the meteorological forecast data with the Bezier curve fitting result, predict the future path of extreme weather, and divide the risk area according to the geographical location of the wind power generation system and the influence range of extreme weather; The output result of the auxiliary task c is the extreme weather path based on Bezier curve fitting and generates a probability distribution map of the risk area.

9. The wind power generation prediction method based on digital twin empowerment according to claim 1, wherein The theoretical power generation P th The calculation formula is as follows: Among them, Pth is the theoretical power generation, A is the swept area of the wind turbine, which is determined by the blade radius R, that is, A = πR 2 , v is the wind speed, and Cp is the wind energy utilization coefficient; The predicted power generation value of the main task The estimated remaining fatigue life of the blades of auxiliary task a And the theoretical power generation P th Integrate them to form a feature vector: Input the feature vector into the trained RNN model to output the corrected predicted power generation value 10. The wind power generation prediction method based on digital twin empowerment according to claim 9, characterized in that Add an error compensation mechanism to the theoretical power generation, and perform dynamic compensation in combination with the output of the RNN. The calculation formula is: Among them, is the predicted value of the compensated power generation, and γ is the compensation coefficient.

11. A wind power prediction system empowered by digital twin, characterized in that, Including: Physical modeling module: used to establish a stress-strain physical model of the wind turbine blade based on the thin plate bending theory and the aerodynamic equation; Digital twin platform: includes a multi-task learning framework, which integrates a main task branch and an auxiliary task branch. The main task branch uses a spatio-temporal sequence model to predict power generation, and the auxiliary task branch performs blade fatigue life prediction, confidence interval calculation, and extreme weather path prediction through the particle filter algorithm and Bezier curve fitting respectively; Data fusion module: calculates the theoretical power generation using the wind energy conversion efficiency equation, and combines the power generation prediction value and the fatigue life data of the turbine blade, inputs them into the recurrent neural network of the digital twin platform, and the output layer is the corrected power generation prediction value; Control optimization module: used to select the power generation prediction value with the highest confidence according to the probability confidence interval, and adjust the yaw angle and blade pitch angle of the turbine in real time based on the probability distribution map of the risk area and the power generation prediction value with the highest confidence.

12. A computer device, characterized in that, The computer device includes: a processor and a memory. At least one instruction, at least one program, a code set or an instruction set is stored in the memory. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the wind power generation prediction method based on digital twin empowerment as described in any one of claims 1 to 10.

Images