Floating type fan load quantification method considering deep and far sea complex constraints

Abstract

The invention relates to a floating fan load quantification method considering deep and far sea complex constraints, and the method comprises the following steps: obtaining short-time-scale environmental data under wind wave flow coupling constraints, setting ocean water depth constraints, and obtaining short-time-scale loads of all parts of a floating fan through an FAST-AQWA model; the method comprises the following steps: constructing a load-environment sample, inputting the load-environment sample into a multi-component differential load prediction model based on MMACM-BiLSTM, extracting and fusing multi-scale feature maps by using MMACM, extracting bidirectional time sequence features by using BiLSTM, and finally outputting predicted loads of each component in various environments; and constructing a floating type fan joint reliability analysis framework comprising multiple components, and evaluating the reliability of each component and the overall reliability of the floating type fan based on the predicted load of each component in multiple environments. Compared with the prior art, the method has the advantages of realizing overall reliability evaluation of a single component and the floating fan and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ocean engineering and artificial intelligence, and particularly relates to a floating wind turbine load quantification method considering complex constraints in deep sea. BACKGROUND

[0002] With the increasing depletion of offshore resources, the offshore wind power industry is gradually expanding from the offshore area to the deep sea area. The deep sea environment makes it difficult to build columnar wind turbines, and the deployment and operation of floating wind turbines not only face more severe wind, wave and current combined effects, but also are significantly affected by multiple deep sea complex constraints such as water depth changes, sea state changes, operation and maintenance window changes. These constraints significantly increase the difficulty of wind turbine normal operation and fault maintenance, especially the time window limited operation and maintenance under typhoon in severe sea conditions. Therefore, developing a rapid load quantification method suitable for the characteristics of deep sea environment has important significance for improving the operation efficiency of wind turbines, reducing costs and ensuring the safe operation of wind turbines.

[0003] Currently, in the load analysis of floating wind turbines, the commonly used tools include the aero-hydro-servo-elastic analysis software (FAST) and the computational fluid dynamics (CFD) method. These tools can simulate the load characteristics of wind turbines in a short time scale, but they face great challenges in handling events that require rapid load analysis under small time window constraints. At the same time, due to the differences in structural characteristics and stress positions of various components of floating wind turbines, the load response to environmental factors such as wind, wave and current shows different sensitivities.

[0004] The existing reliability analysis of floating wind turbines is mostly based on the analysis of the load borne by external components, without considering the coupling relationship between the transmission of external structural responses to internal components. The dramatic changes in external environmental loads can easily cause damage to internal components through mechanical paths. Therefore, it is necessary to develop new technologies or improve existing methods to realize rapid load analysis of internal and external components of floating wind turbines under multiple working conditions, which has important guiding significance for the design, construction and operation of wind turbines. At the same time, the constraint of external limit load has important reference significance for the reliable operation of floating wind turbines, and the reliability constraint calculated from the limit load constraint can help to develop corresponding operation strategies for wind turbines. By improving the accuracy of load prediction and the reliability analysis of multiple components, the maintenance plan of wind turbines can be more effectively planned to ensure the safe and stable operation of floating wind turbines in severe sea conditions. SUMMARY

[0005] The purpose of the present application is to provide a floating wind turbine reliability evaluation method that improves prediction accuracy.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for quantifying the load of a floating wind turbine considering complex constraints in deep-sea environments includes the following steps:

[0008] We acquire short-timescale environmental data under wind, wave and current coupling constraints, set ocean depth constraints, and use the FAST-AQWA model to obtain the short-timescale loads of each component of the floating wind turbine through iterative calculation.

[0009] The short-timescale loads of each component and their corresponding environmental data are used to form load-environment samples, which are then input into a multi-component differentiated load prediction model based on MMACM-BiLSTM. Multi-scale feature maps are extracted and fused using MMACM, and bidirectional temporal features are further extracted using BiLSTM. Finally, the predicted loads of each component under various environments are output, thus completing the load quantization process.

[0010] Furthermore, the components include blades, tower, mooring, and main shaft, wherein the blades, main shaft, and tower constitute the upper structure of the floating wind turbine.

[0011] Furthermore, the FAST-AQWA model includes a FAST sub-model and an AQWA sub-model, and the iterative calculation process includes:

[0012] The dynamic equations for the floating wind turbine are constructed as follows:

[0013]

[0014] In the formula, M is the mass matrix composed of inertial mass and added mass; C and K are the damping matrix and stiffness matrix after assembly, respectively; Let be the unknown acceleration vector at time t; Let X(t) be the unknown velocity vector at time t; X(t) be the unknown displacement at time t; and h(t) be the acceleration impulse function matrix, used to test the radiation memory effect defined by the following equation:

[0015]

[0016] In the formula, B(ω) is the radiation damping corresponding to the wave frequency ω;

[0017] Using the short-timescale wind, wave, and flow environment data as input, the tower base load is obtained by solving the dynamic equations using the FAST sub-model;

[0018] The tower base load is transmitted as an external force to the AQWA sub-model to calculate the floating platform position, floating platform velocity, and floating platform acceleration.

[0019] The pontoon position, pontoon velocity and pontoon acceleration are further transferred to the FAST sub-model for solving, and the loop calculation is performed until the loop ends, so that the pontoon position, pontoon velocity and pontoon acceleration calculated finally by the AQWA sub-model are obtained, and the dynamic response of the wind turbine superstructure is further determined, and the load data of each component of the floating wind turbine under short time scale are obtained according to the dynamic response results, including the blade in-plane bending moment, the tower in-wind direction bending moment, the mooring tension and the main shaft torque.

[0020] Further, in each loop calculation process, the following steps are further included:

[0021] When the tower foundation load obtained by the FAST sub-model is transferred to the AQWA sub-model, the translation load and the rotation load are converted, and the conversion expressions are respectively:

[0022]

[0023] In the formula, F AQWA and F FAST are the translation force vectors in the AQWA sub-model and the FAST sub-model respectively; is the inverse matrix of T mat ; T mat is a transformation matrix for correcting the platform position vector; M AQWA is the torque vector of the pontoon mass center relative to the inertial coordinate system of the AQWA sub-model, and the origin of the inertial coordinate system is the tower foundation; M FAST is the torque vector acting on the tower foundation obtained in the FAST sub-model relative to the local coordinate system of the tower foundation; and COG is the position vector from the reference point to the pontoon mass center.

[0024] Further, in each loop calculation process, the following steps are further included:

[0025] When the pontoon velocity calculated by the AQWA sub-model is transferred to the FAST sub-model, the pontoon velocity is corrected, and the correction expression is:

[0026] V = U - T mat · COG × ω

[0027] In the formula, V is the translation velocity vector used in the FAST sub-model; U and ω are the translation velocity vector and the rotation velocity vector of the pontoon obtained in the AQWA sub-model respectively; T mat is a transformation matrix for correcting the platform position vector; and COG is the position vector from the reference point to the pontoon mass center.

[0028] Further, the MMACM is introduced on multiple different scale convolution channels on a multi-scale convolutional neural network, and the step of extracting and fusing the multi-scale feature maps by the MMACM includes:

[0029] Taking the load-environment sample of each component as input, for each scale convolution channel, different scale convolution kernels are used for parallel convolution operation to obtain local or global features under different receptive fields on each convolution channel as the initial feature map of each convolution channel;

[0030] The attention mechanism of each convolution channel is used to perform global average pooling and global maximum pooling respectively to obtain channel-level description vectors x avg,n and x max,n respectively, and then the channel-level description vectors x avg,n and x max,n are fused to obtain global feature value expressions of each convolution channel.

[0031] The global feature value expressions of each convolution channel are taken as the channel weight vectors of each convolution channel, and are multiplied element by element to each channel of the initial feature map of the corresponding convolution channel to obtain the feature map corresponding to each convolution channel, and further fusion is performed to obtain the feature map of each component after fusion.

[0032] Further, the operation expression of the global average pooling is:

[0033]

[0034] In the formula, x avg, is the value obtained after the global average pooling and the global maximum pooling of the nth channel, that is, the channel-level description vector, H and W represent the height and width of the feature map respectively, and x(n,i,j) is the feature value of the ith row and jth column on the channel n.

[0035] Further, the operation expression of the global maximum pooling is:

[0036]

[0037] In the formula, x max,n is the value obtained after the global maximum pooling of the nth channel, that is, the channel-level description vector, H and W represent the height and width of the feature map respectively, and x(n,i,j) is the feature value of the ith row and jth column on the channel n.

[0038] Further, the operation expression for fusing the channel-level description vectors x avg,n and x max,n is:

[0039] x n = xavg,n ·x max,n

[0040] wherein x n is the global feature value expression of the nth channel.

[0041] Further, the multi-component differentiated load prediction model further comprises an attention layer connected with the BiLSTM, and the step of outputting the predicted load of each component under multiple environments comprises:

[0042] Based on the feature map fused by each component, a bidirectional time sequence feature is extracted by using BiLSTM, wherein the operation expression of LSTM in BiLSTM is:

[0043] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0044] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0045] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0046] c t =f t *c t-1 +i t *tanh(W c ·[h t-1 ,x t ]+b c )

[0047] h t =o t *tanh(c t )

[0048] wherein i t is an input gate, o t is an output gate, f t is a forget gate, c t is the state of a memory unit, used to save the long-time dependence information of the network, h t is the hidden state of time step t, representing the memory of LSTM at the current time step, h t-1 is the hidden state of time step t-1, W i , Wf o c is a weight matrix, x t is the input feature of time step t, i.e., the fused feature extracted by MMACM, b i f o c is a bias term for adjusting the output result, and σ is a sigmoid activation function.

[0049] The bidirectional time sequence feature is taken as an input of an attention layer, and a hidden state h t An attention score is calculated to measure the importance of a time step to load prediction, and the calculation expression of the attention score is as follows:

[0050]

[0051] In the formula, e t is an attention score, h t is a hidden state of time step t, W att is an attention weight matrix, and is a learnable parameter.

[0052] The attention score is normalized to obtain an attention weight, and the expression is as follows:

[0053]

[0054] The hidden state of time step t is weighted and summed with the corresponding attention weight to obtain a context vector that fuses key information, and the expression is as follows:

[0055]

[0056] In the formula, α t is an attention weight, c is a context vector, represents the most important information in a sequence, e k is a kth attention score, α k is an attention weight of time step t, h k is a kth output hidden state.

[0057] The context vector is mapped to a target load dimension to output predicted loads of each component in multiple environments.

[0058] Further, the method further comprises the following steps: constructing a floating wind turbine combined reliability analysis framework containing multiple components, and evaluating the reliability of each component and the overall reliability of the floating wind turbine based on the predicted loads of each component in multiple environments, wherein the step of constructing the floating wind turbine combined reliability analysis framework containing multiple components comprises:

[0059] ​​​​​The main shaft torque is taken as the input torque of the floating wind turbine gearbox, and the input torque expression of the gearbox is:

[0060]

[0061] In the formula, T gear is the input torque of the gearbox, T main represents the main shaft torque; T friction represents the torque of bearing friction loss; i is the gear ratio;

[0062] Based on the input torque of the gearbox, the bending stress of the gear is calculated, and the expression is:

[0063]

[0064] In the formula, σ B is the bending stress of the gear, b is the tooth width, m n is the normal modulus, and K is the shape coefficient;

[0065] Based on the bending stress of the gear, the Paris crack propagation model considering the marine corrosion factor is constructed to represent the evolution relationship of crack length with load cycle number, and the expression is:

[0066]

[0067] In the formula, a is the crack length, N is the load cycle number, α eff is the corrosion factor considering the influence of corrosion environment, L M is the material constant representing the crack propagation characteristics of the material, Δσ B (t) is the stress amplitude of the gear at time t, which can be determined according to the extreme value of each time period during the typhoon, and p is the exponential parameter of the Paris equation;

[0068] For the crack length a in the evolution relationship with the load cycle number N, the integral is carried out in the range of initial crack length to critical crack length and load cycle number, and the following expression is obtained:

[0069]

[0070] In the formula, N f is the load cycle number experienced from the initial crack to the critical crack, a c is the critical crack length, and a0 is the initial crack length;

[0071] According to the reliability definition and Weibull distribution, the relationship between the reliability and failure probability at time t and the relationship between the failure probability and load cycle number are obtained, which are respectively represented as:

[0072] R(t) = 1 - P(t)

[0073]

[0074] In the formula, R(t) is the reliability at time t, P(t) is the failure probability at time t, N t is the number of load cycles, P(N t ) is the failure probability of the gearbox, and β is a shape parameter.

[0075] Based on the evolution relationship of the crack length with the number of load cycles, the failure probability P(N t ) of the gearbox is further expressed as:

[0076]

[0077] In the formula, N t is the number of load cycles at time t, a t is the crack length at time t.

[0078] For the three external components of the blade, the tower and the mooring, the reliability is calculated by using the limit state function, the failure probability of each external component is calculated, and the expression is:

[0079] P(t) = P{G max -G(t) < 0}

[0080] In the formula, P(t) is the failure probability of each component, G max is the maximum allowable load of the blade, the tower and the mooring, and G(t) is the current load.

[0081] According to the failure probability of each external component and the failure probability of the gearbox, a blade-tower-mooring-gear joint reliability calculation model is constructed, and the construction process of the joint reliability analysis framework of the floating wind turbine containing multiple components is completed, wherein the blade-tower-mooring-gear joint reliability calculation model is expressed as:

[0082]

[0083] In the formula, R total (t) is the blade-tower-mooring-gear joint reliability calculation model, P i (t) is P i , and each component failure probability at time t is expressed.

[0084] Compared with the prior art, the present application has the following beneficial effects:

[0085] (1) The application considers the coupling relationship of the influence of the external structure response of the floating wind turbine on the transmission of the internal components and the influence of environmental factors through the FAST-AQWA model in the cyclic calculation process, and constitutes a load-environment sample which can accurately predict the load of each component in various environments by using a multi-component differentiated load prediction model.

[0086] (2) The multi-component differentiated load prediction model of the application can predict the load of the blades, tower, mooring and gear of the four floating wind turbines according to different environmental parameters of wind and wave flow, and analyze the influence degree of different environmental parameters on the load of the four components by fusing convolution channel operation with attention mechanism and adding bidirectional long short-term memory network and attention layer. The model can reduce the calculation cost at the expense of small precision, reduce the dependence on high-cost simulation means in the long-time scale scenario of typhoon, and realize rapid load analysis.

[0087] (3) Based on MSCNN, the application further fuses global maximum pooling based on global average pooling in the traditional squeeze-excitation channel attention mechanism, integrates attention and convolution feature channels, constructs MMACM to realize the capture of key features and overall trends, and explicitly captures the degree and feature expression of the influence of different environmental variables on the load of each component; then Bi-LSTM is introduced to capture the context features of the load change over time, and the extracted features are input into the attention layer to enhance the focusing ability on the load mutation moment.

[0088] (4) On the basis of load quantification, the application also constructs a floating wind turbine joint reliability analysis framework to evaluate the reliability of each component and the overall reliability of the floating wind turbine, which can significantly improve the accuracy of reliability evaluation and more effectively plan the maintenance plan of the wind turbine to ensure the safe and stable operation of the floating wind turbine in severe sea conditions.

[0089] (5) In the floating wind turbine joint reliability analysis framework of the application, the reliability of the blades, tower and mooring three external components is described by the limit state function, and the reliability calculation method of the internal component gear box is based on the Paris crack formula to deduce from the main shaft torque to the gear load. The framework can not only calculate the overall reliability of the floating wind turbine, but also calculate the reliability of a single component. In the actual application stage, each component can be monitored and corresponding decisions can be made according to different reliability requirements. BRIEF DESCRIPTION OF DRAWINGS

[0090] Figure 1 is a method flowchart of the application;

[0091] Figure 2 is a MMACM network structure diagram of the application;

[0092] Figure 3 The environmental parameter values ​​selected for this invention, taking into account the coupling constraints of wind, waves, and current.

[0093] Figure 4 The one-hour working condition prediction effect of the multi-component differentiated load prediction model of the present invention;

[0094] Figure 5 This is the distribution of the relative importance of the six environmental parameters of this invention to each component;

[0095] Figure 6 Figure 1 shows the reliability curves of the floating wind turbine as a whole and its components, where (a) shows the reliability of each component, (b) shows the overall reliability of the floating wind turbine, and (c) shows the reliability of the gear component. Detailed Implementation

[0096] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0097] Example 1

[0098] This embodiment provides a method for quantifying the load of a floating wind turbine that takes into account the complex constraints of deep-sea environments, such as... Figure 1 As shown, the method includes the following steps:

[0099] S1: The loads of the four components of the floating wind turbine—blade, tower, mooring, and main shaft—are solved using the FAST-AQWA model and used as load prediction samples.

[0100] The step takes the environmental data at each time under the wind wave current coupling constraint as the mean value, sets the ocean water depth constraint, generates short-time scale wind wave current environmental data by using the Kaimal wind spectrum, the JONSWAP wave spectrum and the Garrett-Munk current spectrum, and takes the data as the data input of the FAST-AQWA model for load calculation. In the embodiment, the wind wave current coupling constraint is taken as a complex constraint in the deep sea. The wind field, the wave field and the current field are not isolated physical processes in the marine environment, but are coupled and influenced each other in the time and space scales. The wind generates wind waves through friction with the sea surface. The greater the wind speed and the longer the duration, the stronger the waves. The wind-driven flow forms the wind-driven current, and the change of the wind direction causes the reconstruction of the flow field. The radiation stress of the sea wave can cause the surface current and change the structure of the tidal current. The current lengthens the wave length and increases the wave steepness, thereby changing the wave propagation characteristics. In short, it means that the wind speed, the wind direction, the significant wave height, the spectral peak period, the current speed and the current direction are not independent in the time, the space and the statistics, and the six environmental parameters influence each other. The ocean water depth constraint mainly reflects in the arrangement and installation of the floating wind turbine, and the operation and maintenance. In the floating wind power project, the water depth will significantly change the engineering boundary conditions and the operation and maintenance of the system. The deeper the water depth, the longer the anchor chain, the stronger the nonlinearity of the mooring system, the larger the motion range, and the more complex the load distribution. The deeper the water depth, the more difficult the operation and maintenance, the smaller the operation and maintenance time window, and the more dependent on the climate window and the high-precision load quantification.

[0101] The FAST-AQWA model includes a FAST sub-model and an AQWA sub-model, and the responses of the blades, the tower, the main shaft, the floating platform and the mooring of the wind turbine are solved by means of cyclic calculation. In the cyclic calculation process, the FAST sub-model obtains the tower base load by solving the dynamic equation, the tower base load is taken as the external force to pass the calculation of the floating platform position, the floating platform speed and the floating platform acceleration to the sub-model, and the calculation results of the AQWA sub-model are passed to the FAST sub-model, so as to cyclically calculate until the cycle is completed, the floating platform position, the floating platform speed and the floating platform acceleration results calculated by the AQWA sub-model are obtained, the dynamic response of the upper structure of the wind turbine is determined, and the in-plane bending moment of the blade of the floating wind turbine, the tower fore-and-aft bending moment, the mooring tension and the main shaft torque are obtained according to the dynamic response results.

[0102] The specific steps of step S1 include:

[0103] Considering the uncertainty of the typhoon-induced wave current, Kaimal wind spectrum, JONSWAP wave spectrum and Garrett-Munk current spectrum are used to generate 1000s of input data, i.e. environmental samples in short time scale, as the data input of the FAST-AQWA model. The FAST-AQWA model solves the responses of the blades, tower, floating platform and mooring of the wind turbine by means of cyclic calculation, and further calculates the blade bending moment, tower bending moment, mooring tension and main shaft torque of the floating wind turbine in short time scale. In the process of cyclic calculation, the tower base load is obtained by solving the dynamic equation of the FAST sub-model, and the tower base load is transmitted to the AQWA sub-model as an external force to calculate the floating platform position, floating platform speed and floating platform acceleration. The calculation results of the AQWA sub-model are transmitted to the FAST sub-model to calculate the dynamic response of the wind turbine superstructure (blades, main shaft and tower), and the in-plane blade bending moment, tower bending moment, mooring tension and main shaft torque of the floating wind turbine in short time scale are obtained according to the dynamic response results.

[0104] The dynamic equation of the floating wind turbine solved is as follows:

[0105]

[0106] In the formula, M is the mass matrix composed of the inertia mass and the added mass; C and K are the assembled damping matrix and stiffness matrix, respectively; is the unknown acceleration vector at time t; h(t) is the acceleration impulse function matrix for verifying the radiation memory effect defined by the following formula:

[0107]

[0108] In the formula, B(ω) is the radiation damping corresponding to the wave frequency ω.

[0109] In each cyclic calculation process, the translation speed of the floating platform needs to be corrected when the floating platform speed obtained by the AQWA sub-model is transmitted to the FAST sub-model. When the reference point of the floating platform defined in the FAST sub-model is the origin of the inertial coordinate system, i.e. (0, 0, 0), the translation speed of the floating platform can be corrected by the following formula:

[0110] V = U - T mat · COG x ω

[0111] In the formula, V is the translation speed vector used in FAST; U and ω are the translation speed vector and rotation speed vector of the floating platform obtained by the AQWA sub-model, respectively; COG is the position vector from the reference point to the center of mass of the floating platform.

[0112] During each cycle of the calculation process, the load vector in the FAST sub-model also needs to be corrected when passed to the AQWA sub-model. When the tower base is the origin of the inertial coordinate system, the translation load and the rotation load need to be converted by the following formula:

[0113]

[0114] where F AQWA and F FAST are the translation force vectors in the AQWA sub-model and the FAST sub-model, respectively; T -1mat is the inverse matrix of T mat ; M AQWA is the moment vector of the buoy mass center relative to the AQWA sub-model inertial coordinate system; and M FAST is the moment vector acting on the tower base obtained in the FAST sub-model relative to the tower base local coordinate system.

[0115] Based on the above calculation steps, four kinds of load data samples of the floating wind turbine blade in-plane bending moment, the tower along-wind bending moment, the mooring tension and the main shaft torque are obtained, and the corresponding wind speed, wind direction, effective wave height, spectral peak period, flow velocity and flow direction data samples are combined to perform load prediction.

[0116] S2: Establish a multi-component differentiated load prediction model that can reflect the different influences of various environmental factors. The model includes a multi-scale multi-attention channel convolution (Multi-scale Multi-attention Channel Convolution Module, MMACM), a BiLSTM (bidirectional long short-term memory network) and an attention layer connected in turn.

[0117] This step inputs the above four kinds of load data and the corresponding environmental data as load-environment samples into the prediction model.

[0118] The specific steps of step S2 include:

[0119] The load data and the corresponding environmental data are input into the prediction model as load-environment samples.

[0120] First, the MMACM is improved on the multi-scale convolutional neural network (Multi-scale Multi-attention Channel Convolution Module, MSCNN), as shown in Figure 2The multi-channel attention mechanism is introduced in the MSCNN, and the attention mechanism is introduced on three convolution channels of 3x3, 5x5 and 7x7 to realize the extraction of multi-scale features. The squeeze-excitation method in the channel attention mechanism uses global information to adaptively learn the weight of each channel. In the global average pooling (GAP) operation based on it, the global maximum pooling (GMP) is further introduced, and the attention mechanism is combined with the convolution channel to construct the MMACM, so as to retain the overall feature trend while more effectively capturing key information.

[0121] Specifically combined Figure 2 The execution steps of the MMACM include the following:

[0122] (1) Multi-scale convolution: The input data is subjected to parallel convolution operation through multiple convolution kernels of different scales (3x3, 5x5, 7x7), and each convolution path is responsible for extracting local or global features under different receptive fields as the initial feature map of each convolution channel. Each convolution path generates a set of initial feature maps representing the feature information extracted at the current scale. These initial feature maps will be modeled separately in the subsequent attention modeling instead of being uniformly processed.

[0123] (2) Multi-attention channel: For each initial feature map corresponding to a scale (scale 1, scale 2, scale 3), a channel attention mechanism is introduced. Compression: Perform global pooling operation (GAP) on each channel to obtain a channel-level description vector x avg,n and x max,n . Excitation: Connect or weight fuse the vectors obtained by GAP and GMP to obtain the global feature value expression x n of each convolution channel. Re-weighting: Multiply the output channel weight vector element by element to each channel of the initial feature map to obtain the feature map corresponding to each convolution channel, realizing adaptive adjustment at the channel level.

[0124] (3) Multi-scale fusion: fuse all the feature maps corresponding to each convolution channel to obtain the feature map after the fusion of each part, and send it to the subsequent network structure.

[0125] Among them, in the compression stage, GAP and GMP operations are performed on the extracted feature maps, and the channel weight is calculated by dot product as follows:

[0126]

[0127] x n =x avg,n ·x max,n

[0128] In the formula, x avg,n x max,n and x n Let H and W represent the values ​​obtained by global average pooling and global max pooling for the nth channel, respectively, and the global feature value representation for that channel; where H and W represent the height and width of the feature map, respectively, and x(n,i,j) is the feature value of the i-th row and j-th column of channel n.

[0129] Secondly, the fused feature map output from the MMACM above is used as input to Bi-LSTM for temporal modeling to extract bidirectional temporal features. The calculation formula for the Long Short-Term Memory (LSTM) network can be expressed as:

[0130] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0131] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0132] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0133] c t =f t *c t-1 +i t *tanh(W c ·[h t-1 ,x t ]+b c )

[0134] h t =o t *tanh(c t )

[0135] In the formula, x t The input features at time step t are the fused features extracted by MSCNN; h t Let c be the hidden state at time step t, representing the network's memory of the current time step; t The state of the memory unit is used to store long-term dependency information of the network.

[0136] Bi-LSTM consists of a forward LSTM and a backward LSTM, which process different vectors of the input sequence respectively. The output hidden state is a concatenation of the two. In this way, the Bi-LSTM layer can capture both past and future information of the input sequence simultaneously.

[0137] Finally, the attention layer is employed to focus on key moments in the bidirectional temporal features. By assigning different weights to each time step, the model can selectively focus on time points that have a significant impact on the load. Its calculation is divided into attention score calculation, attention weight calculation, and weighted context vector, which are represented as follows:

[0138]

[0139] In the formula, h t The hidden state output by the Bi-LSTM; W att Let e ​​be the attention weight matrix, which consists of learnable parameters. t α is the attention score; α is the attention weight; c is the context vector that integrates key information, representing the most important information in the sequence.

[0140] The context vector is mapped to the target load dimension, and the predicted load of each component is output under various environments.

[0141] Based on the above steps, a multi-component differentiated load prediction model based on MMACM-BiLSTM is established to obtain prediction data for four loads—in-plane bending moment of floating wind turbine blades, windward bending moment of tower, mooring tension, and main shaft torque—under six environmental parameters: response wind speed, wind direction, significant wave height, spectral peak period, flow velocity, and flow direction. This provides a foundation for reliability analysis.

[0142] The simulation analysis is as follows:

[0143] MMACM-BiLSTM Prediction Performance Analysis:

[0144] This invention uses wind speed, wind direction, significant wave height, spectral peak period, current velocity, and current direction data for a certain typhoon at every moment from 12:00 on August 7, 2019 to 0:00 on August 10, 2019. This data represents environmental data under the constraint of wind-wave-current coupling. The values ​​of each environmental parameter are as follows: Figure 3 As shown, the ocean depth constraint is set at 100-111 meters, taking the OO-Star semi-submersible platform designed by the EU project Lifes50+, equipped with a DTU 10MW wind turbine, as an example.

[0145] This invention selects three indicators—correlation coefficient, bias, and root mean square error—to evaluate the predictive performance of the model. The calculation formula is as follows:

[0146]

[0147] like Figure 4 The figure shows the performance of MMACM-BiLSTM in predicting one-hour load conditions. Firstly, from a quantitative perspective of predictive performance, the MMACM-BiLSTM model exhibits high fitting ability across all load types. The model performs best in mooring tension prediction, with a correlation coefficient RC = 0.9753, a root mean square error (RMSE) of 0.1495, and a bias (BIAS) of 0.0147. This indicates that the model can effectively capture the dynamic changes in mooring system tension under one-hour environmental parameter disturbances, and the predicted values ​​are highly consistent with the actual values. This reflects the model's good time-dependent modeling ability for mid-frequency oscillation signals. For tower bending moment prediction, the model also shows good predictive ability, with a correlation coefficient R = 0.9656, an RMSE of 0.1655, and a BIAS of -0.0171. Although there is a slight systematic underestimation trend, the overall error control is good. From the magnified area in the figure, it can be observed that in the load abrupt change segment around 2700s, the model can effectively perceive the amplitude and trend of change, demonstrating good disturbance resistance. Regarding blade bending moment prediction, the signal exhibits higher frequency disturbances due to the significant impact of wind speed variations and dynamic lift at the blade tip. Nevertheless, the model still achieved a high correlation (R = 0.950), RMSE of 0.1744, and BIAS of 0.0228, indicating its robustness in handling high-frequency loads. In contrast, the prediction accuracy of the main shaft torque is slightly lower than the previous three. Main shaft torque variations are significantly affected by hub inertia, instantaneous wind shear, and system coupling, resulting in more complex nonlinear and high-frequency disturbance characteristics. However, the correlation coefficient still reaches R = 0.9479, RMSE of 0.1926, and BIAS of 0.0268, demonstrating good predictive ability.

[0148] Based on the attention weights in MMACM, the relative importance analysis is used to determine the contribution of each environmental parameter b to the load. The calculation formula can be expressed as:

[0149]

[0150] Figure 5The distribution of the relative importance of wind speed, wind direction, significant wave height, spectral peak period, flow velocity, and flow direction in the load prediction of four key components of a floating wind turbine is shown across all training samples. From the component reliability and overall reliability, it is evident that in blade load modeling, wind direction and wind speed rank among the top with weights of 38.853% and 32.121%, respectively, significantly higher than other parameters. This result is consistent with the physical mechanism of blades being dominated by aerodynamic loads; wind speed directly determines the dynamic intensity of the rotor, while changes in wind direction cause dynamic adjustments in the angle of attack and the angle of flow, thus affecting the stress distribution across different blade profiles. For the tower system, the relative importance ranking also shows the dominance of wind direction (37.985%) and wind speed (24.245%). This indicates that the tower primarily bears wind-induced thrust and the shear effect caused by its coupling. Changes in wind direction not only affect the overall thrust direction but also cause a shift in the inertial force at the tower tip, exacerbating the risk of lateral shear and local buckling. In the mooring system, the significance of the significant wave height reaches 35.896%, significantly higher than other variables. The peak period (10.902%), current velocity (12.547%), and current direction (4.211%) also show significant influence. This result accurately characterizes the hydrodynamic loading mechanism dominated by waves: changes in wave period and wave height significantly regulate the heave and pitch motion of the floating body, thereby amplifying the range of mooring line tension fluctuations. Wind speed has an attention weight of only 8.382% in the mooring system, indicating that wind disturbances have a relatively indirect impact on the tension system, mainly manifested through platform motion coupling. The relative importance distribution of the main shaft shows that wind speed (44.443%) is the most important factor, followed by wind direction (23.850%). As the main component in the wind turbine torque transmission path, changes in wind speed directly lead to drastic changes in torque amplitude, thus affecting the gear meshing load distribution and fatigue life.

[0151] Example 2

[0152] This embodiment builds upon step S2 in Embodiment 1 to construct a joint reliability analysis framework for floating wind turbines that includes multiple components. This framework is used to evaluate the reliability of each component and the overall reliability of the floating wind turbine, including the following:

[0153] S3: Based on the Paris fatigue crack propagation model, the relationship between the transmission and response of gear mechanical loads under external loads is derived, and a joint reliability analysis framework for a floating wind turbine system including multiple key internal and external components is further constructed.

[0154] This step involves a reliability analysis of multiple internal and external components of the floating wind turbine based on the predicted blade bending moment, tower bending moment, mooring tension, and main shaft torque. The gear input torque and bending stress can be calculated sequentially from the main shaft torque. This stress serves as the input stress in the Paris crack propagation model, allowing for the calculation of gear reliability based on the amount of crack growth. The reliability of the blade, tower, and mooring components is represented by limit state functions. Finally, a joint reliability calculation model for the blade-tower-mooring-gear system is established.

[0155] The specific steps of step S3 include:

[0156] If the main shaft torque is used as the input torque of the floating wind turbine gearbox, then the gearbox input torque can be expressed as:

[0157]

[0158] In the formula, T main Indicates spindle torque; T friction The torque represents the frictional loss of the bearing; i is the gear ratio.

[0159] After calculating the input torque, the bending stress of the gear can be calculated:

[0160]

[0161] In the formula, b is the tooth width, m n is the normal phase modulus, and K is the shape factor.

[0162] In marine environments, long-term exposure to salt spray can easily lead to internal corrosion in gearboxes, thus accelerating crack propagation. To more accurately assess gearbox reliability, this paper further incorporates the influence of marine corrosion factors into the traditional Paris crack propagation model. The evolution of crack length *a* as a function of the number of load cycles *N* can be expressed as:

[0163]

[0164] In the formula, α eff Corrosion factors considering the influence of corrosive environments; L M Δσ is a material constant representing the crack propagation characteristics of a material. B (t) represents the stress amplitude of the gear at time t, which can be determined based on the extreme values ​​of each time period during the typhoon; p is the exponential parameter of the Paris equation.

[0165] Applying the above equation to a and N from the initial crack length a0 to the critical crack length a c Integral over the range and under the load cycle number:

[0166]

[0167] Based on the definition of reliability and the Weibull distribution, the reliability R and failure probability P at time t, and the failure probability and load cycle number N, are related. t The relationship between them can be represented as:

[0168] R(t) = 1 - P(t)

[0169]

[0170] In the formula, β is the shape parameter. Combining the relationship between crack length and load cycle number calculated above, the failure probability of the gearbox can be expressed as:

[0171]

[0172] For the three external components—blades, towers, and moorings—reliability is calculated using limit state functions, and the failure probability of each component is expressed as follows:

[0173] P(t)=P{G max -G(t)<0}

[0174] In the formula, G max G(t) represents the maximum allowable load for the blades, tower, and mooring; G(t) represents the load at the current moment.

[0175] Finally, the combined reliability calculation model of blade-tower-mooring-gear can be expressed as:

[0176]

[0177] In the formula, P i This represents the failure probability of each component at time t.

[0178] The following simulation analysis is based on Example 1:

[0179] Joint reliability analysis of multiple internal and external components of floating wind turbines:

[0180] like Figure 6Figures (a), (b), and (c) show the reliability of floating wind turbine components, overall reliability, and gear reliability (corrosion factor sensitivity) under different corrosion factor values, respectively. The reliability curves reveal significant differences in the performance of each component under different operating conditions. The tower exhibits strong stability, with minimal change in reliability over time, indicating a robust structural design. While the blades show a slight decrease, they maintain high overall reliability, reflecting strong tolerance under high wind loads. In contrast, the mooring system and gears show significant degradation trends, especially the gears, whose reliability drops sharply over time, potentially becoming a major weak point in the floating wind turbine system. The overall reliability curve shows that under wind, wave, and current conditions, the wind turbine system remains relatively stable for the first 30 hours, but its reliability decreases rapidly over time, dropping below 0.95 after 60 hours, indicating significant structural impact. Gear reliability decreases over time, with the rate of decrease positively correlated with the corrosion factor. When the corrosion factor is 0.5 × 10⁻⁶... -8 At the initial stage, the reliability decayed the slowest, remaining above 0.9 even after 60 hours, indicating that minor environmental corrosion had a minimal impact on the system. However, as the corrosion factor increased to 6×10, the reliability degradation slowed. -8 The reliability decreased exponentially, dropping to nearly 0.1 after 60 hours, indicating that the gearbox structure would rapidly lose its functional integrity under highly corrosive conditions. It is noteworthy that the corrosion factor was at 1.0 × 10⁻⁶. -8 Up to 3.0×10 -8 Significant reliability differences have been observed within the range, indicating that even moderate corrosion levels can have a significant adverse impact on the service performance of critical wind turbine components.

[0181] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0183] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0184] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0185] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0186] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0187] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for quantifying the loads of a floating wind turbine taking into account complex constraints in deep offshore waters, characterized in that, The method comprises the following steps: obtaining short-time scale environmental data under the constraint of wind wave flow coupling, setting a marine water depth constraint, and obtaining short-time scale loads of each component of the floating wind turbine by a loop calculation method through a FAST-AQWA model; constructing a load-environment sample of each component by using the short-time scale load of each component and the corresponding environmental data of each component as input, inputting the load-environment sample into a multi-component differential load prediction model based on MMACM-BiLSTM, extracting a multi-scale feature map by using MMACM and fusing the multi-scale feature map, further extracting a bidirectional time sequence feature by using BiLSTM, and finally outputting the predicted load of each component under multiple environments to complete the load quantification process.

2. The method of claim 1, wherein, The components include blades, a tower, a mooring and a main shaft, wherein the blades, the main shaft and the tower are the upper structure of the floating wind turbine.

3. The method of claim 2, wherein, The FAST-AQWA model comprises a FAST sub-model and an AQWA sub-model, and the loop calculation process comprises: a dynamic equation of the floating wind turbine is constructed and is expressed as: where M is the mass matrix combined from the inertia mass and the added mass; C and K are the assembled damping matrix and stiffness matrix, respectively; is the unknown acceleration vector at time t; is the unknown velocity vector at time t; X(t) is the unknown displacement at time t; h(t) is the acceleration impulse function matrix for checking the radiation memory effect defined by the following equation: wherein B(ω) is a radiation damping corresponding to a wave frequency ω; the short-time scale wind wave flow environmental data are used as input, the FAST sub-model is used to obtain tower base loads by solving the dynamic equation; the tower base loads are transmitted to the AQWA sub-model as external forces to calculate a pontoon position, a pontoon speed and a pontoon acceleration; the pontoon position, the pontoon speed and the pontoon acceleration are further transmitted to the FAST sub-model for solving, the loop calculation is performed until the loop calculation is completed, the pontoon position, the pontoon speed and the pontoon acceleration calculated by the AQWA sub-model are obtained, and the dynamic response of the upper structure of the wind turbine is further determined, and the load data of each component of the floating wind turbine under the short-time scale are obtained according to the dynamic response result, including a blade in-plane bending moment, a tower along-wind bending moment, a mooring tension and a main shaft torque.

4. The method of claim 3, wherein, In each loop calculation process, the following steps are further included: when the tower base loads obtained by the FAST sub-model are transmitted to the AQWA sub-model, translation loads and rotation loads are converted, and the conversion expressions are respectively: where F AQWA and F FAST are the translational force vectors in the AQWA submodel and the FAST submodel, respectively; is the inverse matrix of T mat ; T mat is the transformation matrix used to correct the platform position vector; M AQWA is the moment vector about the buoy mass center relative to the AQWA submodel inertial coordinate system, whose origin is at the tower base; M FAST is the moment vector about the tower base local coordinate system, which is obtained from the FAST submodel; and COG is the position vector from the reference point to the buoy mass center.

5. The method of claim 3, wherein, In each loop calculation process, the following steps are further included: when the pontoon speed calculated by the AQWA sub-model is transmitted to the FAST sub-model, the pontoon speed is corrected, and the correction expression is: V = U - T mat • COG x ω where V is the translational velocity vector used in the FAST submodel; U and ω are the translational and rotational velocity vectors of the platform obtained from the AQWA submodel; T mat is the transformation matrix used to correct the platform position vector; and COG is the position vector from the reference point to the center of mass of the platform.

6. The method of claim 1, wherein, The MMACM is introduced on multiple different scale convolution channels on a multi-scale convolutional neural network, and the step of extracting and fusing a multi-scale feature map by using the MMACM comprises: each component load-environment sample is used as input, for each scale convolution channel, different scale convolution kernels are used for parallel convolution operation to obtain local or global features under different receptive fields on each convolution channel as initial feature maps of each convolution channel; The attention mechanism of each convolution channel is used to respectively perform global average pooling and global maximum pooling to obtain a channel-level description vector x of each convolution channel avg,n and x max,n The channel-level description vectors x avg,n and x max,n are fused to obtain a global feature value expression of each convolution channel; the global feature value expression of each convolution channel is used as a channel weight vector of each convolution channel, and is multiplied to each channel of the initial feature map of the corresponding convolution channel element by element to obtain the corresponding feature map of each convolution channel, and the feature maps of each component are further fused to obtain a fused feature map of each component.

7. The method of claim 6, wherein, The operation expression of the global average pooling is: In the formula, x avg, is the value obtained after global average pooling and global maximum pooling of the nth channel, that is, the channel-level description vector, H and W represent the height and width of the feature map, respectively, and x(n, i, j) is the feature value of the ith row and jth column in the nth channel.

8. The method of claim 6, wherein, The operation expression of the global maximum pooling is: In the formula, x max,n is the value obtained after global maximum pooling of the nth channel, that is, the channel-level description vector, H and W represent the height and width of the feature map, respectively, and x(n, i, j) is the feature value of the ith row and jth column in channel n.

9. The method of claim 6, wherein, The channel level description vector x avg,n and x max,n The operation expression for fusing is: x n = x avg,n • x max,n In the formula, x n is the global feature value expression of the nth channel.

10. The method of claim 1, wherein, The multi-component differentiated load prediction model further comprises an attention layer connected with the BiLSTM, and the step of outputting the predicted load of each component under multiple environments comprises: Based on the fused feature map of each component, a bidirectional time sequence feature is extracted by using BiLSTM, wherein the operation expression of LSTM in BiLSTM is as follows: i t = σ(W i · [h t-1 , x t ]+ b i ) f t = σ(W f · [h t-1 , x t ]+ b f ) o t = σ(W o · [h t-1 , x t ]+ b o ) c t = f t * c t-1 + i t * tanh(W c · [h t-1 , x t ] + b c ) h t = o t tanh(c t ) In the formula, i t is the input gate, o t is the output gate, f t is the forget gate, c t is the state of the memory cell, used to save the long-time dependence information of the network, h t is the hidden state of time step t, representing the memory of the LSTM at the current time step, h t-1 is the hidden state of time step t-1, W i , W f , W o , W c is the weight matrix, x t is the input feature of time step t, that is, the fused feature extracted by MMACM, b i , b f , b o , b c is the bias term, used to adjust the output result, and sigma is the sigmoid activation function; The bidirectional temporal features are used as input to the attention layer, and the hidden state h at each time step is... t An attention score is calculated to measure the importance of a time step to load prediction, wherein the expression for calculating the attention score is: where e t is the attention score, h t is the hidden state at time step t, W att is the attention weight matrix, which is a learnable parameter; The attention score is normalized to obtain an attention weight, and the expression is as follows: The hidden state at the time step t is weighted and summed with the corresponding attention weight to obtain a context vector fused with key information, and the expression is as follows: where α t is the attention weight, c is the context vector, representing the most important information in the sequence, e k is the kth attention score, α k is the attention weight for time step t, h k is the kth output hidden state; The context vector is mapped to a target load dimension to output the predicted load of each component under multiple environments.

Images