Vertical axis wind turbine power prediction method based on multi-dimensional environment data

Abstract

The invention provides a vertical axis wind turbine power prediction method based on multi-dimensional environment data, and relates to the technical field of vertical axis wind turbine power based on the multi-dimensional environment data. Multi-dimensional data are collected in real time based on a sensor network, and the multi-dimensional data comprise environment parameters such as wind speed, turbulence intensity and temperature and humidity, and operation states such as rotor rotating speed and pitch angle; secondly, performing empirical mode decomposition on a vibration acceleration signal, constructing a physical information neural network, taking environmental data and an aerodynamic load frequency domain energy spectrum as input, outputting a real-time wind energy utilization coefficient and an aerodynamic coefficient, and finally, determining the aerodynamic load based on an aerodynamic load time domain signal. A load cyclic spectrum is extracted by adopting online rain flow counting of dynamic window length, an accumulated damage value is calculated in combination with a fatigue damage model of humidity correction, and when the residual life is lower than a threshold value, load peak suppression and power generation efficiency balance are realized through cooperative control of magnetic resistance torque linear attenuation and flap fuzzy PID (Proportion Integration Differentiation).

Description

Technical Field

[0001] The present invention relates to the technical field of vertical axis wind turbine power based on multi-dimensional environmental data, and in particular to a vertical axis wind turbine power prediction method based on multi-dimensional environmental data. Background Art

[0002] Vertical axis wind turbines are increasingly used in distributed power generation and urban environments due to their compact structure and good wind direction adaptability. However, the unsteady characteristics of their aerodynamic loads, such as dynamic stall and turbulent transients, pose significant challenges to fatigue life prediction and operational control. Traditionally, aerodynamic load monitoring relies on simplified mechanism models or purely data-driven black box predictions. The former lacks accuracy due to the neglect of dynamic stall effects and three-dimensional flow interference, while the latter lacks physical constraints and is prone to overfitting problems. Both methods make it difficult to achieve real-time, high-precision monitoring under complex working conditions. At the same time, existing fatigue life assessment methods are mostly based on static load spectra or simplified cycle counting, without considering the transient correction of turbulence intensity to dynamic angle of attack and the coupled effect of humidity on material crack propagation, resulting in a remaining life prediction deviation of more than 30%, causing the active load reduction strategy to respond lag or be overly conservative, seriously affecting unit reliability and power generation efficiency.

[0003] Existing technologies mainly address this problem through two approaches: first, based on the modified blade element momentum theory combined with the NREL air density correction formula, but without considering the nonlinear effect of turbulent transients on dynamic stall; second, using time-series neural networks such as LSTM to directly fit SCADA data. Although the error can be reduced to 15% under steady-state conditions, the lack of physical constraints causes the generalization ability to drop sharply under extreme turbulence, and the error rebounds to more than 35%.

[0004] The shortcomings of existing technologies are mainly reflected in the following aspects: 1) The aerodynamic model is separated from the data-driven method, and it is impossible to take into account the coupling of dynamic stall effects and real-time environmental parameters, such as the humidity-fatigue life correlation; 2) Fatigue life assessment relies on static rain flow counting and standard material parameters, ignoring the humidity-induced crack growth acceleration effect. Experiments have shown that increasing humidity from 30% to 90% can shorten fatigue life by 40%; 3) The load reduction strategy is triggered by a fixed threshold and is not dynamically coupled with real-time life indicators, resulting in the coexistence of power generation efficiency loss and mechanical overload risks.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The object of the present invention is to provide a method for predicting vertical axis wind turbine power based on multi-dimensional environmental data to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for predicting vertical axis wind turbine power based on multi-dimensional environmental data, comprising the following steps:

[0009] Step 1: Acquire real-time multi-dimensional data of the vertical axis wind turbine through a sensor network, including environmental data, operating data, and structural response data. The environmental data includes wind speed, turbulence intensity, air temperature, and humidity; the operating data includes rotor speed and pitch angle; and the structural response data includes blade root strain and vibration acceleration.

[0010] Step 2: Perform modal decomposition on the vibration acceleration signal to extract the aerodynamic load components. Based on the historical multidimensional data and the aerodynamic load components as input, the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are used as labels to construct and train a physical information neural network model. The neural network output is constrained by the vertical axis aerodynamic equation residual and the dynamic stall hysteresis effect deviation in the loss function.

[0011] Step 3: Based on the trained physical information neural network model, the real-time normal force coefficient and tangential force coefficient are output. Combined with the real-time air density, wind speed, blade azimuth angle, and rotor swept area, the aerodynamic load time domain signal is calculated using the aerodynamic load formula.

[0012] Step 4: Perform online rain flow counting on the aerodynamic load time domain signal, calculate the fatigue damage value based on the Miner cumulative damage criterion, and output the remaining life index. When the output remaining life index is less than the set threshold, the load reduction strategy is triggered, including adjusting the reluctance torque and flap deflection angle;

[0013] The formula of the loss function is:

[0014] L=||P pred -P meas ||2+α·||R BEM (C p ,C L ,C D )||2+β·||R stall (α eff )||2

[0015] Among them, P pred is the theoretical power value predicted by the physical information neural network based on the input multi-dimensional environmental data, P meas is the real-time output power of the vertical axis wind turbine directly measured by the sensor during actual operation, α and β are weight coefficients, and R BEM is the residual term of the modified BEM equation of the vertical axis wind turbine double-disc multi-tube model, R stallis the residual term of the Leishman-Beddoes dynamic stall model, α eff is the dynamic angle of attack correction factor, C p ,C L ,C D are the instantaneous wind energy utilization coefficient, normal force coefficient and tangential force coefficient respectively.

[0016] Furthermore, the wind speed in the environmental data is set to V, the turbulence intensity is TI, and the temperature is T air The humidity is H, the rotor speed in the operating data is ω, the pitch angle is θ, the blade root strain in the structural response data is ε, and the vibration acceleration is a g ;

[0017] According to the temperature T air The real-time air density ρ is calculated based on the following formula:

[0018]

[0019] Among them, P atm is atmospheric pressure, R air is the gas constant of air, and e is the water vapor partial pressure.

[0020] Furthermore, the modal decomposition of the vibration acceleration signal and the logic of extracting the aerodynamic load component are as follows:

[0021] Vibration acceleration a g Perform empirical mode decomposition to obtain several intrinsic mode functions. Based on the rotor rotation frequency, filter out the frequency band where the frequency domain energy is concentrated in the aerodynamic load dominant frequency band;

[0022] The calculation formula of the wind wheel rotation frequency is:

[0023]

[0024] Where ω is the rotor speed;

[0025] Among them, the screening rules of the intrinsic mode function are:

[0026] f aero ∈[1f rot ,2f rot ,3f rot ]

[0027] Perform fast Fourier transform on the selected intrinsic mode functions and calculate their frequency domain energy spectrum E(f) as the frequency domain representation of the aerodynamic load;

[0028] The intrinsic mode function screening condition of the empirical mode decomposition is: the energy concentration is set to The resolution of the frequency domain energy spectrum E(f) is less than or equal to 0.5 Hz, and the frequency band f aero The boundary error is less than or equal to 5%.

[0029] Furthermore, a physical information neural network model is constructed and trained based on historical multidimensional data and aerodynamic load components, and the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are used as labels. The logic is as follows:

[0030] The historical multidimensional data includes wind speed, turbulence intensity, rotor speed, pitch angle, frequency domain energy spectrum and blade root strain;

[0031] The corresponding instantaneous wind energy utilization coefficient is obtained by inverse calculation of the measured power. The normal force coefficient and the tangential force coefficient are calculated based on the decomposition of the aerodynamic load, and correspond to the component of the aerodynamic force acting on the blade perpendicular to the direction of movement of the blade and the component along the direction of movement, respectively.

[0032] A physical information neural network architecture is used, with historical multidimensional data and aerodynamic load components after modal decomposition as input, and the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient as output. The mean square error between the predicted value and the label is minimized, and the aerodynamic equation and dynamic stall model are forced to be satisfied through a loss function. The predicted instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are substituted into the BEM equation modified by the double-disc multi-flow tube model, and the theoretical residual is calculated. Based on the Leishman-Beddoes dynamic stall model, the residual of the aerodynamic coefficient hysteresis effect and the measured value is calculated.

[0033] The predicted power P in the loss function pred The formula is:

[0034]

[0035] Among them, ρ is the real-time air density, A is the swept area of the wind wheel, V is the real-time wind speed, C p is the instantaneous wind energy utilization coefficient;

[0036] Correction of angle of attack based on turbulence intensity TI Where ζ is the turbulence sensitivity coefficient calibrated by wind tunnel experiments, α geo is the geometric angle of attack, which is calculated from the blade azimuth angle φ and the incoming flow direction, U is the relative wind speed, and c is the blade chord length;

[0037] The formula of the turbulence sensitivity coefficient is ζ = 0.25·(TI / 0.15) 0.6 ; The dynamic angle of attack correction term The relative wind speed of the vertical axis wind turbine blades should be calculated based on the azimuth angle: Where R is the radius of the wind wheel;

[0038] The residual term R of the modified BEM equation of the double-disc multi-flow tube model is BEM The formula is:

[0039]

[0040] in, is the blade azimuth angle, φ is the cumulative angle obtained by integrating the real-time angular velocity ω from the initial moment to the current moment t, is the solidity, N is the number of blades, c is the blade chord length, R is the radius of the wind wheel, a is the axial induction factor, C L ,C D are the normal force coefficient and tangential force coefficient output by the physical information neural network;

[0041] The dynamic stall residual term R stall The formula is:

[0042]

[0043] Among them, C L,invi is the inviscid lift, ΔC L,vort is the lift increment caused by vortex generation, ΔC L,sep is the lift loss caused by flow separation, τ1 and τ2 are the time constants of vortex convection and reattachment, which are related to the blade chord length c and the relative wind speed U.

[0044] Furthermore, the normal force coefficient C output by the physical information neural network model after training is L and the tangential force coefficient C D , and the real-time air density ρ, wind speed V, blade azimuth φ and rotor swept area A, the aerodynamic load time domain signal F is calculated by the aerodynamic load formula aero The formula is:

[0045]

[0046] Among them, F aero As a time domain signal input to the online rain flow counting;

[0047] The time domain signal F of the aerodynamic load aero Perform Kalman filtering with a cutoff frequency of 3f rot , retain the aerodynamic dominant frequency band and filter out high-frequency noise.

[0048] Furthermore, the online rain flow counting is performed on the aerodynamic load time domain signal, and the logic of calculating the fatigue damage value in combination with the Miner cumulative damage criterion is as follows:

[0049] From the time domain signal F aeroExtract the peak-valley sequence. If the segment composed of four consecutive peak-valley points meets the following constraints, then the segment is regarded as a closed loop, so as to extract all the closed loops in the time domain signal.

[0050] The constraint condition is: within this section, if the previous peak value is not less than the next peak value, and the previous valley value is not greater than the next valley value, or if the previous peak value is not greater than the next peak value, and the previous valley value is not less than the next valley value, then this area is defined as a closed loop.

[0051] The online rain flow counting is based on a sliding time window to detect the time domain signal F in real time. aero , filter closed cycles and calculate the stress amplitude of each closed cycle. The calculation formula is as follows:

[0052]

[0053] Where Δσ i is the stress amplitude corresponding to the i-th closed cycle, P i is the highest peak value of the i-th closed cycle, V u is the lowest valley value of the i-th closed loop, i is the index of the closed loop, and i∈[1,N], N represents the time domain signal F aero The number of closed loops in

[0054] The fatigue life is calculated by combining the stress amplitude and the humidity correction term according to the formula:

[0055]

[0056] Among them, N f,i is the fatigue life of the ith closed cycle, σ fat ,m is the material fatigue parameter, γ H is the humidity influence coefficient;

[0057] The formula for the fatigue damage value D is:

[0058]

[0059] Where N is the total number of closed loops.

[0060] Furthermore, the formula for outputting the remaining life indicator is:

[0061] L remain =L design (1-D)

[0062] Among them, L design is the design life, which is the rated operating time;

[0063] When L remain <0.3Ldesign When the load reduction strategy is triggered, including adjusting the reluctance torque where k em is the reluctance gain coefficient, and the flap deflection angle δ = K p ·D;

[0064] The magnetoresistance gain coefficient k em The dynamic adjustment formula is:

[0065]

[0066] in, is the decay rate coefficient;

[0067] The flap deflection angle δ is controlled by a fuzzy PID algorithm, and the proportional coefficient where σ crit is the yield strength of the material, is the rated reluctance torque, A ref is the cross-sectional area of the blade.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] The present invention proposes a vertical axis wind turbine power-life collaborative monitoring and control method based on physical-data fusion to systematically solve the three major technical problems of vertical axis wind turbines in complex environments: inaccurate power prediction, distorted life assessment and delayed load reduction control: First, in order to address the problem that traditional aerodynamic models ignore the coupling of dynamic stall and multi-dimensional environment, the BEM residual constraint corrected by the double-disk multi-flow tube model is embedded through the physical information neural network, and the angle of attack is dynamically corrected in combination with real-time turbulence intensity and humidity data, so that the prediction error of the wind energy utilization coefficient is reduced under extreme turbulent conditions. At the same time, the accuracy of capturing transient fluctuations of the normal force coefficient is significantly improved, which significantly optimizes the power output stability.

[0070] The present invention addresses the defect that the existing life assessment ignores the acceleration of crack propagation due to humidity. Through the Miner criterion of online rain flow counting and humidity correction, the cyclic damage of the aerodynamic load time domain signal is calculated in real time. Combined with the humidity influence factor calibrated by the salt spray experiment, the fatigue life prediction deviation is reduced to achieve accurate early warning of life attenuation. Finally, to address the shortcomings of the fixed threshold load reduction strategy, the multi-level load reduction control triggered by the remaining life is used to dynamically adjust the magnetic resistance torque and the flap deflection angle. The closed loop of the sensor network collecting multi-dimensional data in real time → the physical constraint neural network predicts the aerodynamic load → the dynamic life assessment triggers the adaptive load reduction, which overcomes the industry problems of model mismatch, life assessment distortion and control response hysteresis under environmental coupling. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of the overall method flow of the present invention;

[0072] Figure 2 Schematic diagram of wind speed-theoretical power curve of the present invention;

[0073] Figure 3 This is a schematic diagram of the fatigue damage-remaining life curve of the present invention;

[0074] Figure 4 This is a fitting curve diagram of the remaining life of the vertical axis wind turbine of the present invention;

[0075] Figure 5 This is a schematic diagram of the fatigue damage-theoretical power curve of the present invention. DETAILED DESCRIPTION

[0076] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0077] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0078] Example:

[0079] See also Figure 1-Figure 5 , the present invention provides a technical solution:

[0080] A method for predicting vertical axis wind turbine power based on multi-dimensional environmental data, comprising the following steps:

[0081] Step 1: Acquire real-time multi-dimensional data of the vertical axis wind turbine through a sensor network, including environmental data, operating data, and structural response data. The environmental data includes wind speed, turbulence intensity, air temperature, and humidity; the operating data includes rotor speed and pitch angle; and the structural response data includes blade root strain and vibration acceleration.

[0082] The performance and lifespan of vertical-axis wind turbines are affected by a combination of the environment, operating status, and structural response. Environmental data quantifies wind resource characteristics and atmospheric conditions, providing input for aerodynamic models. Operational data monitors the real-time control status of wind turbines and optimizes power output. Structural response data assesses blade fatigue damage, predicts remaining life, and triggers protection strategies.

[0083] Set the wind speed in the environmental data to V, the turbulence intensity to TI, and the temperature to T air Assume that the humidity is H, the rotor speed in the operating data is ω, the pitch angle is θ, the blade root strain in the structural response data is ε, and the vibration acceleration is a g ;

[0084] ω is the angular velocity of the rotor, which determines the tip speed ratio. An increase in the rotational speed leads to an increase in the tip speed ratio. The pitch angle θ is the angle between the blade and the plane of rotation, which is used to adjust the aerodynamic load and power output. An increase in the pitch angle leads to a decrease in the lift coefficient, which in turn leads to a decrease in power output.

[0085] The blade root strain ε is the deformation of the blade root material, which directly reflects the aerodynamic load and structural stress. The increase in strain leads to an increase in the amplitude of the alternating stress, which in turn leads to accelerated accumulation of fatigue damage; the vibration acceleration a g The blade vibration intensity is used to identify the aerodynamic load components and structural resonance frequencies. Increased acceleration amplitude leads to an increased risk of dynamic stall or mechanical failure, triggering load reduction control.

[0086] According to the temperature T air The real-time air density ρ is calculated based on the following formula:

[0087]

[0088] Among them, P atm is atmospheric pressure, R air is the air gas constant, e is the water vapor partial pressure;

[0089] ρ reflects the air mass per unit volume and directly affects the wind energy conversion efficiency. The higher the density, the greater the kinetic energy captured at the same wind speed. It changes with the temperature T air rises and falls, that is, the gas expands, and falls slightly with increasing humidity H, because water vapor displaces dry air;

[0090] Atmospheric pressure P atm It is the static atmospheric pressure, which is related to the altitude. The standard sea level value is 101.325kPa. It can be set as a constant at a specific geographical location or measured in real time by a pressure sensor. The air gas constant R air It is the physical property parameter of dry air, which is only related to the air composition and does not change with the environment;

[0091] Temperature Tair is the ambient temperature, which determines the speed of thermal motion of gas molecules. An increase in temperature increases the distance between molecules, which in turn leads to a decrease in density. Humidity H is the ratio of the water vapor content in the air to the saturation content, ranging from [0,1]. An increase in humidity leads to an increase in the water vapor partial pressure e, which in turn leads to a slight decrease in the air density ρ.

[0092] The water vapor partial pressure e is the pressure generated by water vapor in the air, reflecting the effect of humidity on air density, which is determined by the temperature T air It is determined by the humidity H, and the increase in temperature leads to an exponential increase in the saturated vapor pressure, while the increase in humidity causes e to increase linearly.

[0093] This formula converts air temperature and humidity into air density, correcting for environmental factor deviations in wind power calculations. A larger ρ indicates denser air, resulting in greater kinetic energy capture and higher theoretical power output at the same wind speed. A smaller ρ indicates thinner air, requiring higher wind speeds to achieve rated power, a common occurrence in areas with high temperatures, high altitudes, or high humidity.

[0094] Temperature T air : For every 10°C increase in temperature, the density decreases by about 3%, significantly affecting the power output in the low wind speed area; when the humidity H increases from 0% to 100%, the density decreases by at most 2%, which is a smaller impact but needs to be considered in high-precision models.

[0095] Step 2: Perform modal decomposition on the vibration acceleration signal to extract the aerodynamic load components. Based on the historical multidimensional data and the aerodynamic load components as input, the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are used as labels to construct and train a physical information neural network model. The neural network output is constrained by the vertical axis aerodynamic equation residual and the dynamic stall hysteresis effect deviation in the loss function.

[0096] The logic of performing modal decomposition on the vibration acceleration signal and extracting the aerodynamic load component is as follows:

[0097] Vibration acceleration a g Perform empirical mode decomposition to obtain several intrinsic mode functions. Based on the rotor rotation frequency, filter out the frequency band where the frequency domain energy is concentrated in the aerodynamic load dominant frequency band;

[0098] The calculation formula of the wind wheel rotation frequency is:

[0099]

[0100] Where ω is the rotor speed;

[0101] Among them, the screening rules of the intrinsic mode function are:

[0102] f aero ∈[1f rot ,2frot ,3f rot ]

[0103] Perform fast Fourier transform on the selected intrinsic mode functions and calculate their frequency domain energy spectrum E(f) as the frequency domain representation of the aerodynamic load;

[0104] Vibration acceleration a g Reflects the dynamic response of the blade under the action of aerodynamic loads, mechanical vibration and noise. The single component signal is obtained after IMF (Intrinsic Mode Function) decomposition. Each IMF represents a vibration mode of different time scales. The frequency of IMF is arranged from low to high. High-frequency IMF corresponds to noise, and low-frequency IMF corresponds to mechanical or aerodynamic loads.

[0105] f rot It is directly determined by the rotor speed ω, which reflects the speed of the wind wheel rotation. When the speed ω increases, the rotation frequency f rot Linear increase; f aero The aerodynamic load energy is concentrated in the fundamental frequency, second harmonic and third harmonic of the rotation frequency. rot Real-time update, dynamic adaptation to the current working conditions; the purpose of this formula is to lock the characteristic frequency of the aerodynamic load to avoid mechanical vibration, such as interference from the bearing fault frequency. If the energy of a certain IMF is concentrated on f aero , indicating that it is dominated by aerodynamic loads;

[0106] The purpose of performing empirical mode decomposition on the vibration acceleration signal is to separate the aerodynamic load component from other interferences, such as mechanical resonance and noise. The larger the amplitude of the IMF, the greater the contribution of the frequency component to the vibration.

[0107] The intrinsic mode function screening condition of the empirical mode decomposition is: the energy concentration is defined as The resolution of the frequency domain energy spectrum E(f) is less than or equal to 0.5 Hz, and the frequency band f aero The boundary error is less than or equal to 5%.

[0108] E(f) represents the energy distribution of the signal at different frequencies, the energy density corresponding to frequency f. The higher the energy, the more significant the frequency component. E(f) represents the energy density corresponding to frequency f. aero A peak appears at , and the aerodynamic load is significant;

[0109] η is the ratio of the energy of the aerodynamic load frequency band to the total energy. The larger η is, the higher the proportion of the aerodynamic load component in the IMF is. This ensures that the selected IMF is dominated by the aerodynamic load, that is, the proportion is greater than or equal to 70%, eliminating noise interference. If η is less than 0.7, the IMF may contain excessive noise or mechanical vibration components and needs to be discarded. The resolution is 0.5Hz to distinguish adjacent harmonics. The margin error is 5% to ensure that the frequency band covers the true aerodynamic frequency.

[0110] The physical information neural network model is constructed and trained based on historical multidimensional data and aerodynamic load components, and the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are used as labels. The logic is as follows:

[0111] The historical multidimensional data includes wind speed, turbulence intensity, rotor speed, pitch angle, frequency domain energy spectrum and blade root strain;

[0112] The corresponding instantaneous wind energy utilization coefficient is obtained by inverse calculation of the measured power, and the calculation formula is: Where ρ is the real-time air density, A is the swept area of the wind wheel, and V is the real-time wind speed;

[0113] The normal force coefficient and the tangential force coefficient are calculated based on the decomposition of the aerodynamic load, and correspond to the component of the aerodynamic force acting on the blade perpendicular to the direction of movement of the blade and the component along the direction of movement, respectively;

[0114] A physical information neural network architecture is used, with historical multidimensional data and aerodynamic load components after modal decomposition as input, and the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient as output. The mean square error between the predicted value and the label is minimized, and the aerodynamic equation and dynamic stall model are forced to be satisfied through a loss function. The predicted instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are substituted into the BEM equation modified by the double-disc multi-flow tube model, and the theoretical residual is calculated. Based on the Leishman-Beddoes dynamic stall model, the residual of the aerodynamic coefficient hysteresis effect and the measured value is calculated.

[0115] The formula of the loss function is:

[0116] L=||P pred -P meas ||2+α·||R BEM (C p ,C L ,C D )||2+β·||R stall (α eff )||2

[0117] Among them, P pred is the theoretical power value predicted by the physical information neural network based on the input multi-dimensional environmental data, P meas is the real-time output power of the vertical axis wind turbine directly measured by the sensor during actual operation, which is obtained through the wind turbine output power sensor. α and β are weight coefficients used to balance the importance of the data error term and the physical residual term. R BEM is the residual term of the modified BEM equation of the vertical axis wind turbine double-disc multi-tube model, R stall is the residual term of the Leishman-Beddoes dynamic stall model, α effis the dynamic angle of attack correction factor, reflecting the effect of turbulent transients on aerodynamic loads, C p ,C L ,C D are the instantaneous wind energy utilization coefficient, normal force coefficient and tangential force coefficient respectively;

[0118] P meas It is the real-time output power of the vertical axis wind turbine directly measured by the sensor during actual operation. It is usually installed at the generator output terminal or grid connection point to directly record the instantaneous value of the generated power. pred It is the theoretical power value predicted by the physical information neural network based on the input multi-dimensional environmental data, such as wind speed, turbulence intensity, rotation speed, etc.

[0119] Wind energy utilization coefficient C p Reflects the efficiency of converting wind energy into mechanical energy, C p The larger it is, the higher the aerodynamic efficiency is and the closer the power output is to the theoretical maximum value; C L ,C D is the normal force and tangential force coefficient, which characterizes the aerodynamic load distribution and affects the blade stress and fatigue life, where C L Increase, lift increases, may cause dynamic stall; C D As it increases, the resistance loss increases;

[0120] R BEM Measures the deviation between the neural network output and the theoretical aerodynamic model. The smaller the residual, the more consistent the neural network prediction is with the physical law. stall Quantify the difference between the actual lift and the stall model. The smaller the residual, the more accurately the dynamic stall effect is modeled.

[0121] This formula combines data-driven and physical constraints to improve C p ,C L ,C D The credibility of the prediction, the data error term P in the loss function pred -P meas Reflects the model fitting ability, the physical residual term R BEM 、R stall Ensure predictions are consistent with aerodynamic principles;

[0122] The predicted power P in the loss function pred The formula is:

[0123]

[0124] Among them, ρ is the real-time air density, A is the swept area of the wind wheel, V is the real-time wind speed, C p is the instantaneous wind energy utilization coefficient;

[0125] P predReflects the theoretical output power of the wind turbine under the current working conditions. The power varies with ρ, V 3 、C p The wind speed increases significantly with the increase of the wind speed. ρ is calculated by temperature and humidity, which affects the efficiency of wind energy capture. The higher the density, the linear increase of power. For example, the power decreases in low temperature and high altitude areas. A is the effective frontal area of the vertical axis wind turbine, which is a geometric parameter. The larger the area, the linear increase of power. V is the main energy source and has a cubic effect on power. The power increases significantly with the increase of wind speed. For example, if the wind speed doubles, the power increases by 8 times. C p It is an index of aerodynamic efficiency, with a maximum theoretical value of 0.593 Betz limit, predicted by neural network and constrained by aerodynamic model. p The closer to the Betz limit, the higher the efficiency;

[0126] The formula is based on the Betz theory, quantifying the theoretical power output of the vertical axis wind turbine, and embedding the real-time environmental parameters ρ, V and aerodynamic efficiency C. p ,provides a physical benchmark for power prediction and guides the optimization of wind turbine control strategies;

[0127] Correction of angle of attack based on turbulence intensity TI Among them, ζ is the turbulence sensitivity coefficient calibrated by wind tunnel experiments, α geo is the geometric angle of attack, calculated from the blade azimuth angle φ and the incoming flow direction, U is the relative wind speed, and c is the blade chord length;

[0128] α eff is the corrected blade angle of attack, reflecting the transient effect of turbulence on aerodynamic loads. As the angle of attack increases, there is a stall critical value, and the lift coefficient C L Increase first and then decrease; it is the theoretical angle of attack calculated by the blade azimuth angle φ and the incoming flow direction;

[0129] ζ is the correction amplitude of the turbulence intensity to the angle of attack. The stronger the turbulence, the more significant the correction. TI is the ratio of the standard deviation of wind speed fluctuation to the average wind speed, reflecting the degree of flow field turbulence. As TI increases, α eff As the speed increases, the risk of dynamic stall increases; U is the vector result of the composite incoming wind speed and blade movement speed. The larger the speed ω, the larger U is, resulting in α eff This formula is used to correct the angle of attack fluctuation caused by turbulence and improve the prediction accuracy of the dynamic stall model;

[0130] The formula of the turbulence sensitivity coefficient is ζ = 0.25·(TI / 0.15) 0.6 ; The dynamic angle of attack correction term The relative wind speed of the vertical axis wind turbine blades should be calculated based on the azimuth angle: Where R is the radius of the wind wheel;

[0131] The vertical axis aerodynamic model is an aerodynamic model designed specifically for the asymmetric flow field characteristics of vertical axis wind turbines. Its core is the blade element momentum theory modified by the double-disc multi-duct model. Unlike the traditional horizontal axis BEM, the DMS model divides the flow field into two upper and lower disk regions, and calculates the blade aerodynamic loads at different azimuth angles separately to more accurately reflect the periodic aerodynamic characteristics of the vertical axis wind turbine.

[0132] The rotor rotation plane is divided into the upper half (blade moving against the wind) and the lower half (blade moving with the wind), and the induction factor a is calculated for each up and a down ;

[0133] The residual term R of the modified BEM equation of the double-disc multi-flow tube model is BEM The formula is:

[0134]

[0135] in, is the blade azimuth angle, φ is the cumulative angle obtained by integrating the real-time angular velocity ω from the initial moment to the current moment t, is the solidity, N is the number of blades, c is the blade chord length, R is the rotor radius, a is the axial induction factor, and the momentum conservation of the upper and lower disk areas and the balance of blade forces are ensured through iterative solution. L ,C D are the normal force coefficient and tangential force coefficient output by the physical information neural network;

[0136] The t in the formula represents the time when real-time data is collected, that is, the time span from the initial moment, such as the system startup time t = 0, to the current data point. In real-time monitoring, t is the timestamp of the sensor data, and the integration interval [0,1] represents the time accumulation from the initial moment to the current moment;

[0137] R BEM is a measure of the C of the neural network prediction L ,C D The smaller the residual deviation from the aerodynamic model, the more consistent the neural network output is with the momentum-blade element theory, and the higher the model credibility. σ is the ratio of the total blade area to the swept area. The higher the solidity, the stronger the aerodynamic interference and the larger the axial induction factor b. a is the speed reduction ratio of the airflow through the wind rotor. The larger a is, the lower the flow velocity behind the wind rotor and the greater the thrust. φ is the rotation angle of the blade relative to the incoming flow direction. It changes periodically with the rotation speed ω and affects the direction of lift and drag. This formula corrects the neural network output through physical constraints, namely momentum conservation and blade element force balance, to ensure that the prediction conforms to aerodynamic principles.

[0138] The dynamic stall constraint is used to describe the transient aerodynamic effect of airflow separation caused by sudden change in angle of attack during the periodic rotation of vertical axis wind turbine blades. The residual term is constructed by the Leishman-Beddoes dynamic stall model to ensure that the lift coefficient C predicted by the neural network is L Includes the effects of dynamic stall;

[0139] The dynamic stall residual term R stall The formula is:

[0140]

[0141] Among them, C L,invi is the inviscid lift, calculated by potential flow theory, ΔC L,vort is the lift increment caused by vortex generation, ΔC L,sep is the lift loss caused by flow separation, τ1, τ2 are the time constants of vortex convection and reattachment, which are related to the blade chord length c and wind speed U;

[0142] R stall Quantify the difference between the actual lift and the stall model. The smaller the residual, the more accurate the dynamic stall effect is modeled. L,invi is the ideal lift calculated by potential flow theory, ignoring the viscous effect, which is calculated by the panel method or vortex grid method and depends on α eff ;ΔC L,vort It is the lift enhancement brought by the leading edge vortex in dynamic stall, which is measured by wind tunnel transient experiments and is positively correlated with the vortex strength;

[0143] ΔC L,sep It is the drop in lift caused by airflow separation. The larger the separation area, the more significant the loss. τ1 and τ2 characterize the delayed effects of vortex convection and reattachment, respectively. τ1 = c / κ1U, τ2 = c / κ2U, and κ1 and κ2 are empirical coefficients that distinguish the vortex generation and reattachment processes. This formula captures the unsteady effects of dynamic stall and improves the accuracy of transient load prediction.

[0144] Step 3: Based on the trained physical information neural network model, the real-time normal force coefficient and tangential force coefficient are output. Combined with the real-time air density, wind speed, blade azimuth angle, and rotor swept area, the aerodynamic load time domain signal is calculated using the aerodynamic load formula.

[0145] The normal force coefficient C is output by the physical information neural network model after training. L and the tangential force coefficient C D , and the real-time air density ρ, wind speed V, blade azimuth φ and rotor swept area A, the aerodynamic load time domain signal F is calculated by the aerodynamic load formula aero The formula is:

[0146]

[0147] Among them, F aero As a time domain signal input to the online rain flow counting;

[0148] The time domain signal F of the aerodynamic load aero Perform Kalman filtering with a cutoff frequency of 3f rot , retain the aerodynamic dominant frequency band and filter out high-frequency noise;

[0149] F aero Reflects the aerodynamic force exerted on the blade during rotation, directly affecting the structural stress and fatigue life. It changes dynamically with air density, wind speed, lift and drag coefficient, and azimuth angle. The greater the density ρ, the more linearly the aerodynamic load increases. For example, the load decreases in low temperature and high altitude areas. The larger the area A, the more linearly the aerodynamic load increases. When the wind speed V doubles, the aerodynamic load increases by 4 times.

[0150] C L is the lift coefficient, which is predicted by the neural network and represents the aerodynamic efficiency of the blade in the normal direction (perpendicular to the incoming flow direction), C L Increase, lift dominates, especially when φ≈90°, aerodynamic load increases; C D is the drag coefficient, which is predicted by the neural network and represents the aerodynamic resistance of the blade in the tangential direction (parallel to the incoming flow direction), C D As the azimuth angle φ increases, the drag offsets the lift, especially when φ≈0°, the aerodynamic load decreases; the change of the azimuth angle φ causes sinφ and cosφ to periodically modulate the contribution ratio of lift to drag;

[0151] Indicates the incoming flow pressure, which is the energy source of aerodynamic force, C L sinφ is the component of lift perpendicular to the rotor plane, C D cosφ is the component of resistance parallel to the wind rotor plane; F aero It is the vector sum of the normal and tangential forces, reflecting the periodic alternating load on the blade during rotation;

[0152] When φ=90°, the blade is perpendicular to the incoming flow, sinφ=1, lift dominates, F aero Increase significantly; when φ=0°, the blade is parallel to the incoming flow, cosφ=1, the drag offsets the lift, F aero Reduce; when the blades of the vertical axis wind turbine rotate one circle, φ changes periodically from 0 to 2π, F aero It presents sinusoidal fluctuations, and the frequency is the wind wheel rotation frequency f rot and its harmonics.

[0153] Step 4: Perform online rain flow counting on the aerodynamic load time domain signal, calculate the fatigue damage value based on the Miner cumulative damage criterion, and output the remaining life index. When the output remaining life index is less than the set threshold, the load reduction strategy is triggered, including adjusting the reluctance torque and flap deflection angle;

[0154] The logic of online rain flow counting of the time domain signal of the aerodynamic load component and calculation of fatigue damage value in combination with the Miner cumulative damage criterion is as follows:

[0155] The logic of online rain flow counting of aerodynamic load time domain signals and calculation of fatigue damage value in combination with Miner cumulative damage criterion is as follows:

[0156] From the time domain signal F aero Extract the peak-valley sequence. If the segment composed of four consecutive peak-valley points meets the following constraints, then the segment is regarded as a closed loop, so as to extract all the closed loops in the time domain signal.

[0157] The constraint condition is: within this section, if the previous peak value is not less than the next peak value, and the previous valley value is not greater than the next valley value, or if the previous peak value is not greater than the next peak value, and the previous valley value is not less than the next valley value, then this area is defined as a closed loop.

[0158] The online rain flow counting is based on a sliding time window to detect the time domain signal F in real time. aero , filter closed cycles and calculate the stress amplitude of each closed cycle. The calculation formula is as follows:

[0159]

[0160] Where Δσ i is the stress amplitude corresponding to the i-th closed cycle, P i is the highest peak value of the i-th closed cycle, V i is the lowest valley value of the i-th closed loop, i is the index of the closed loop, and i∈[1,N], N represents the time domain signal F aero The number of closed loops in

[0161] The fatigue life is calculated by combining the stress amplitude and the humidity correction term according to the formula:

[0162]

[0163] Among them, σ fat ,m is the material fatigue parameter, is the SN curve parameter, N f,i is the fatigue life of the ith closed cycle, γ H is the humidity influence coefficient; N f,i The smaller it is, the higher the damage rate of this stress amplitude is;

[0164] Δσ i is the stress variation caused by alternating load, Δσ i The larger the N f,i The smaller it is, the higher the damage rate is; fat is the reference stress value of the SN curve, corresponding to N f,i =1, which is determined by material fatigue testing and has nothing to do with humidity; m is the slope of the SN curve, reflecting the sensitivity of stress amplitude to life. The larger m is, the more significant the effect of stress amplitude on life.

[0165] γ H Quantify the accelerated effect of humidity on fatigue life. The higher the humidity H, the faster N f,i The greater the reduction, the salt spray test is used for calibration. The experimental method includes: applying alternating loads in the range of humidity H∈[30%,90%] and recording the crack growth rate; γ H When γ H ≤1 / H max , such as H max =0.9,γ H ≤1.11;

[0166] The formula for the fatigue damage value D is:

[0167]

[0168] Where N is the total number of closed cycles;

[0169] D reflects the cumulative degree of fatigue damage caused by alternating stress in the material. D∈[0,1]. The closer it is to 1, the shorter the remaining life. When D=1, it indicates failure.

[0170] The formula for outputting the remaining life indicator is:

[0171] L remain =L design (1-D)

[0172] Among them, L design is the design life, i.e. the rated operating time;

[0173] L remain is the remaining safe operating time predicted based on the current damage value. The larger D is, the greater L remain The smaller it is, the more urgent the life warning is; L design It is the expected service life under rated conditions, such as 20 years, determined by the manufacturer based on materials and operating conditions;

[0174] When L remain <0.3L design When the load reduction strategy is triggered, including adjusting the reluctance torque where k em is the reluctance gain coefficient, and the flap deflection angle δ = K p ·D;

[0175] The magnetoresistance gain coefficient k em The dynamic adjustment formula is:

[0176]

[0177] in, is the attenuation rate coefficient, which is optimized by historical failure data; ζ controls the sensitivity of gain to the change of remaining life, which is optimized by historical failure data. The larger the gain is, the more sensitive it is to lifespan changes. 0.2 is used to adjust the critical point for triggering load reduction. remain / L design <0.3, the gain decreases rapidly;

[0178] The flap deflection angle δ is controlled by a fuzzy PID algorithm, and the proportional coefficient where σ crit is the yield strength of the material, is the rated reluctance torque, A ref is the cross-sectional area of the blade;

[0179] σ crit is the ultimate tensile / compressive strength of the blade material, which is determined by material testing. The higher the strength, the greater the allowable flap deflection; is the maximum output torque of the generator design. The greater the torque, the greater the proportional coefficient K. p The smaller the value, the smoother the control; A ref is the blade cross-sectional area, used to make the proportionality factor dimensionless and ensure that K p Dimensionless to avoid unit conflicts.

[0180] Table 1 Statistics of factors affecting power

[0181]

[0182]

[0183] During the operation of wind turbines, the acquisition of environmental data is fundamental. Table 1 shows data from 30 different operating conditions, where wind speeds range from 5 m / s to 34 m / s and turbulence intensities range from 0.1 to 0.2. These changes can reflect the working environment of wind turbines under different meteorological conditions. For example, the wind speed in scenario 1 is 5 m / s and the turbulence intensity is 0.1, while in scenario 30, the wind speed reaches 34 m / s and the turbulence intensity is 0.2. This diverse environmental data provides a basis for training and validation of the model, ensuring its adaptability and accuracy under various operating conditions.

[0184] The calculation of air density is a key step in the model. Table 1 lists the air density of each set of data, ranging from 1.171 kg / m 3 to 1.204kg / m 3 These data are calculated based on the air temperature and humidity and meet the requirements of the technical plan. Air density directly affects the prediction of theoretical power. In each scenario, the theoretical power calculation is based on the current environmental data and the design parameters of the wind turbine. For example, in scenario 1, the wind speed is 5m / s and the air density is 1.204kg / m 3 , assuming the wind wheel area is 10m 2 and instantaneous wind energy utilization coefficient C p The theoretical power can be calculated as 0.3. This calculation method ensures the accuracy of the model under different wind speeds and environmental conditions. Accordingly, as wind speed increases, the theoretical power reaches 67.5 kW in scenario 30, showing a clear growth trend, which shows the direct impact of wind speed on power output.

[0185] The actual power values reported in Table 1 serve as validation criteria for the model's output. Comparing theoretical and actual power allows us to assess the model's predictive capabilities. For example, in Scenario 2, the theoretical power is 2.0 kW, while the actual power is 1.85 kW. The relative closeness between the two indicates that the model performs well under these conditions. Actual power varies with wind speed and turbulence intensity, demonstrating the model's effectiveness in dynamic environments.

[0186] Fatigue damage and remaining life are important indicators of wind turbine safety. Fatigue damage is calculated based on stress amplitudes under actual operating conditions and evaluated in conjunction with the design life. For example, in scenario 10, the fatigue damage is 0.23 and the remaining life is 830 hours. This data helps maintenance personnel take timely measures to avoid equipment failures caused by excessive fatigue.

[0187] The data presented in Table 1 comprehensively demonstrates the practicality and effectiveness of this technical solution through the acquisition of multi-dimensional environmental parameters, calculation of air density, comparison of theoretical and actual power, and assessment of fatigue damage and remaining life. This data not only provides a reliable basis for model training and validation but also promotes wind turbine performance optimization and safe operation, ensuring the sustainable use of wind energy.

[0188] Table 2 Theoretical power and remaining life statistics

[0189]

[0190]

[0191] In each scenario in Table 2, the comparison of theoretical and actual power reveals the actual performance of the wind turbine. For example, in Scenario 1, the theoretical power is 1.5 kW, while the actual power is 1.4 kW. The difference is small, indicating that the model performs well under this operating condition. As wind speed and rotational speed increase, the theoretical power also increases. In Scenario 30, the theoretical power reaches 67.5 kW, while the actual power is 64.5 kW. The difference is still within an acceptable range (only 3 kW), demonstrating the rationality and accuracy of the model under high loads. This accuracy reflects the model's ability to effectively predict wind turbine performance and ensure the efficient use of wind energy under different environmental conditions. This result is consistent with the concept of power prediction based on multidimensional environmental data, as emphasized in the technical solution.

[0192] Fatigue damage is a key indicator for assessing wind turbine operational reliability. The table shows that the fatigue damage value gradually increases from 0.1 in Scenario 1 to 0.64 in Scenario 30. This indicates that as power output increases, the wind turbine's workload also increases, leading to a gradual increase in fatigue damage. In Scenario 10, the fatigue damage value reaches 0.23, which may indicate the need for inspection and maintenance to avoid potential failures.

[0193] Remaining life is an important indicator for evaluating the usable time of a wind turbine in its current working state. In the table, the remaining life gradually decreases from 990 hours to 430 hours, and this process is closely related to the increase in fatigue damage. When fatigue damage is low, the remaining life of the wind turbine is relatively long, indicating that the equipment is in good operating condition. For example, the remaining life in scenario 12 is 790 hours, showing that under this environmental condition, the equipment still has sufficient usable time. On the contrary, in scenario 30, the remaining life drops to 430 hours, which suggests that the equipment requires more frequent maintenance and attention. This approach not only reduces the risk of equipment failure, but also effectively extends the overall service life of the wind turbine, meeting the requirements of reliability and maintenance optimization in the technical solution.

[0194] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0195] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0196] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0197] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for predicting vertical axis wind turbine power based on multi-dimensional environmental data, characterized in that: The specific steps include: Step 1: Acquire real-time multi-dimensional data of the vertical axis wind turbine through a sensor network, including environmental data, operating data, and structural response data. The environmental data includes wind speed, turbulence intensity, air temperature, and humidity; the operating data includes rotor speed and pitch angle; and the structural response data includes blade root strain and vibration acceleration. Step 2: Perform modal decomposition on the vibration acceleration signal to extract the aerodynamic load components. Based on the historical multidimensional data and the aerodynamic load components as input, the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are used as labels to construct and train a physical information neural network model. The neural network output is constrained by the vertical axis aerodynamic equation residual and the dynamic stall hysteresis effect deviation in the loss function. Step 3: Based on the trained physical information neural network model, the real-time normal force coefficient and tangential force coefficient are output. Combined with the real-time air density, wind speed, blade azimuth angle, and rotor swept area, the aerodynamic load time domain signal is calculated using the aerodynamic load formula. Step 4: Perform online rain flow counting on the aerodynamic load time domain signal, calculate the fatigue damage value based on the Miner cumulative damage criterion, and output the remaining life index. When the output remaining life index is less than the set threshold, the load reduction strategy is triggered, including adjusting the reluctance torque and flap deflection angle; The formula of the loss function is: L=||P pred -P meas ||2+a·||R BEM (C p ,C L ,C D )||2+β·||R stall (a eff )||2 Among them, P pred is the theoretical power value predicted by the physical information neural network based on the input multi-dimensional environmental data, P meas is the real-time output power of the vertical axis wind turbine directly measured by the sensor during actual operation, α and β are weight coefficients, and R BEM is the residual term of the modified BEM equation of the vertical axis wind turbine double-disc multi-tube model, R stall is the residual term of the Leishman-Beddoes dynamic stall model, α eff is the dynamic angle of attack correction factor, C p ,C L ,C D are the instantaneous wind energy utilization coefficient, normal force coefficient and tangential force coefficient respectively.

2. The method for predicting vertical axis wind turbine power based on multi-dimensional environmental data according to claim 1, characterized in that: Set the wind speed in the environmental data to V, the turbulence intensity to TI, and the temperature to T air The humidity is H, the rotor speed in the operating data is ω, the pitch angle is θ, the blade root strain in the structural response data is ε, and the vibration acceleration is a g ; According to the temperature T air The real-time air density ρ is calculated based on the following formula: Among them, P atm is atmospheric pressure, R air is the gas constant of air, and e is the water vapor partial pressure.

3. The method for predicting vertical axis wind turbine power based on multi-dimensional environmental data according to claim 2, characterized in that: The logic of performing modal decomposition on the vibration acceleration signal and extracting the aerodynamic load component is as follows: Vibration acceleration a g Perform empirical mode decomposition to obtain several intrinsic mode functions. Based on the rotor rotation frequency, filter out the frequency band where the frequency domain energy is concentrated in the aerodynamic load dominant frequency band; The calculation formula of the wind wheel rotation frequency is: Where ω is the rotor speed; Among them, the screening rules of the intrinsic mode function are: in aero ∈[1f rot ,2f rot ,3f rot ] Perform fast Fourier transform on the selected intrinsic mode functions and calculate their frequency domain energy spectrum E(f) as the frequency domain representation of the aerodynamic load; The intrinsic mode function screening condition of the empirical mode decomposition is: the energy concentration is set to The resolution of the frequency domain energy spectrum E(f) is less than or equal to 0.5 Hz, and the frequency band f aero The boundary error is less than or equal to 5%.

4. The method for predicting vertical axis wind turbine power based on multi-dimensional environmental data according to claim 3, characterized in that: The physical information neural network model is constructed and trained based on historical multidimensional data and aerodynamic load components, and the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are used as labels. The logic is as follows: The historical multidimensional data includes wind speed, turbulence intensity, rotor speed, pitch angle, frequency domain energy spectrum and blade root strain; The corresponding instantaneous wind energy utilization coefficient is obtained by inverse calculation of the measured power. The normal force coefficient and the tangential force coefficient are calculated based on the decomposition of the aerodynamic load, and correspond to the component of the aerodynamic force acting on the blade perpendicular to the direction of movement of the blade and the component along the direction of movement, respectively. A physical information neural network architecture is used, with historical multidimensional data and aerodynamic load components after modal decomposition as input, and the corresponding instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient as output. The mean square error between the predicted value and the label is minimized, and the aerodynamic equation and dynamic stall model are forced to be satisfied through a loss function. The predicted instantaneous wind energy utilization coefficient, normal force coefficient, and tangential force coefficient are substituted into the BEM equation modified by the double-disc multi-flow tube model, and the theoretical residual is calculated. Based on the Leishman-Beddoes dynamic stall model, the residual of the aerodynamic coefficient hysteresis effect and the measured value is calculated. The predicted power P in the loss function pred The formula is: Among them, ρ is the real-time air density, A is the swept area of the wind wheel, V is the real-time wind speed, C p is the instantaneous wind energy utilization coefficient; Correction of angle of attack based on turbulence intensity TI Where ζ is the turbulence sensitivity coefficient calibrated by wind tunnel experiments, α geo is the geometric angle of attack, which is calculated from the blade azimuth angle φ and the incoming flow direction, U is the relative wind speed, and c is the blade chord length; The formula of the turbulence sensitivity coefficient is ζ = 0.25·(TI / 0.15) 0.6 ; The dynamic angle of attack correction term The relative wind speed of the vertical axis wind turbine blades should be calculated based on the azimuth angle: Where R is the radius of the wind wheel; The residual term R of the modified BEM equation of the double-disc multi-flow tube model is BEM The formula is: in, is the blade azimuth angle, φ is the cumulative angle obtained by integrating the real-time angular velocity ω from the initial moment to the current moment t, is the solidity, N is the number of blades, c is the blade chord length, R is the radius of the wind wheel, a is the axial induction factor, C L ,C D are the normal force coefficient and tangential force coefficient output by the physical information neural network; The dynamic stall residual term R stall The formula is: Among them, C L,invi is the inviscid lift, ΔC L,vort is the lift increment caused by vortex generation, ΔC L,sep is the lift loss caused by flow separation, τ1 and τ2 are the time constants of vortex convection and reattachment, which are related to the blade chord length c and the relative wind speed U.

5. The method for predicting vertical axis wind turbine power based on multi-dimensional environmental data according to claim 4, characterized in that: The normal force coefficient C is output by the physical information neural network model after training. L and the tangential force coefficient C D , and the real-time air density ρ, wind speed V, blade azimuth φ and rotor swept area A, the aerodynamic load time domain signal F is calculated by the aerodynamic load formula aero The formula is: Among them, F aero As a time domain signal input to the online rain flow counting; The time domain signal F of the aerodynamic load aero Perform Kalman filtering with a cutoff frequency of 3f rot , retain the aerodynamic dominant frequency band and filter out high-frequency noise.

6. The method for predicting vertical axis wind turbine power based on multi-dimensional environmental data according to claim 5, characterized in that: The logic of online rain flow counting of aerodynamic load time domain signals and calculation of fatigue damage value in combination with Miner cumulative damage criterion is as follows: From the time domain signal F aero Extract the peak-valley sequence. If the segment composed of four consecutive peak-valley points meets the following constraints, then the segment is regarded as a closed loop, so as to extract all the closed loops in the time domain signal. The constraint condition is: within this section, if the previous peak value is not less than the next peak value, and the previous valley value is not greater than the next valley value, or if the previous peak value is not greater than the next peak value, and the previous valley value is not less than the next valley value, then this area is defined as a closed loop. The online rain flow counting is based on a sliding time window to detect the time domain signal F in real time. aero , filter closed cycles and calculate the stress amplitude of each closed cycle. The calculation formula is as follows: Where Δσ i is the stress amplitude corresponding to the i-th closed cycle, P i is the highest peak value of the i-th closed cycle, V i is the lowest valley value of the i-th closed loop, i is the index of the closed loop, and i∈[1,N], N represents the time domain signal F aero The number of closed loops in The fatigue life is calculated by combining the stress amplitude and the humidity correction term according to the formula: Among them, N f,i is the fatigue life of the ith closed cycle, σ fat ,m is the material fatigue parameter, γ H is the humidity influence coefficient; The formula for the fatigue damage value D is: Where N is the total number of closed loops.

7. The method for predicting vertical axis wind turbine power based on multi-dimensional environmental data according to claim 6, characterized in that: The formula for outputting the remaining life indicator is: L remain =L design ·(1-D) Among them, L design is the design life, which is the rated operating time; When L remain <0.3L design When the load reduction strategy is triggered, including adjusting the reluctance torque where k em is the reluctance gain coefficient, and the flap deflection angle δ = K p ·D; The magnetoresistance gain coefficient k em The dynamic adjustment formula is: in, is the decay rate coefficient; The flap deflection angle δ is controlled by a fuzzy PID algorithm, and the proportional coefficient where σ crit is the yield strength of the material, is the rated reluctance torque, A ref is the cross-sectional area of the blade.

Images