Optimization design method for anti-fatigue and lightweight III-type hydrogen storage container

Abstract

The invention discloses an optimal design method for an anti-fatigue and lightweight III-type hydrogen storage container. The method takes liner fatigue life prolonging and container mass reduction as optimization objectives, the liner wall thickness, the end socket ellipsoid ratio and the composite layer laying scheme are selected as key design variables, the fatigue characteristics of the container under different design parameters are analyzed through a finite element model, and liner fatigue life and container mass data are collected to construct a database; and constructing a mapping relationship between design parameters and performance indexes by adopting a deep neural network, and performing Pareto frontier search by taking the fatigue life of the liner predicted by the network and the quality parameters of the container as optimization targets to determine an optimal design parameter combination. According to the method, the contradiction problem between the fatigue performance and the lightweight requirement in the design of the hydrogen storage container is effectively solved, the fatigue life is remarkably prolonged on the premise that the controllable quality of the container is guaranteed, meanwhile, the design period needed by a traditional trial and error method is greatly shortened, and an efficient and reliable technical scheme is provided for engineering optimization design of the III type hydrogen storage container.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of high-pressure hydrogen storage containers, in particular to an anti-fatigue and lightweight type III composite hydrogen storage container optimization design method. BACKGROUND

[0002] Hydrogen energy has become an important direction of global energy technology transformation and one of the key paths to achieve the carbon neutralization goal due to its high energy density, diverse sources, environmental friendliness and wide application. Hydrogen refueling station is a key infrastructure in the hydrogen energy industry chain, and the fixed hydrogen storage container is a key equipment for storing hydrogen in the hydrogen refueling station, which has the characteristics of large capacity, high hydrogen storage pressure and high fatigue life. At present, the most widely used fixed hydrogen storage container is type I container and type II container made of high-strength chromium-molybdenum steel and carbon fiber composite material. However, chromium-molybdenum steel is susceptible to hydrogen embrittlement, and its fatigue performance will decrease significantly in a high-pressure hydrogen environment. Moreover, with the continuous increase of hydrogen storage pressure and the increase of the inner container wall thickness, great challenges will be brought to the manufacturing and heat treatment process. In comparison, using high-nickel stainless steel with excellent hydrogen embrittlement resistance as the inner container material and combining with the full-winding carbon fiber composite layer structure can significantly improve the fatigue performance of the container. Moreover, with the continuous increase of hydrogen storage pressure and the decrease of carbon fiber material price and the maturity of winding forming process, the advantages of type III container are more and more obvious compared with type I container and type II container.

[0003] Fatigue is the main failure mode of hydrogen storage containers for hydrogen refueling stations. For type III hydrogen storage containers, relevant experiments show that the fatigue performance of the composite layer is much better than that of the metal inner container, so fatigue failure generally occurs in the inner container. There are many design parameters that affect the fatigue life of the inner container, including the outer diameter, wall thickness, head shape of the inner container, and the thickness, angle and arrangement of the composite layer. Moreover, the calculation and test of the fatigue life of the hydrogen storage container are relatively complex. The existing design methods generally consider the strength performance and fatigue performance separately, the deviation of the fatigue life calculation is large, and the influence of the composite layer design parameters is rarely considered. Therefore, the design results often cannot meet the expected requirements and need to be verified by fatigue tests. Therefore, it is of great significance to establish a type III composite hydrogen storage container optimization design method that comprehensively considers various inner container and composite layer design parameters for the optimization design and fatigue performance analysis of such hydrogen storage containers. SUMMARY

[0004] In order to overcome the defects in the prior art, the application provides an anti-fatigue and lightweight type III composite hydrogen storage container optimization design method, which comprehensively considers various inner container parameters and composite layer parameters, overcomes the shortcomings that the hydrogen storage container design can only be carried out through tests and experience, optimizes the design parameters of the composite hydrogen storage container and improves the fatigue performance of the composite hydrogen storage container.

[0005] The idea of the application is: a finite element model of a hydrogen storage container with different design parameters is established, the inner container mass and the composite layer mass are extracted, and the minimum fatigue life of the inner container under a specified cyclic load is calculated, a database is established and a deep neural network model is trained, the input of the model is the parameters corresponding to the inner container and the composite layer, the prediction result of the model is the mass of the inner container and the composite layer and the fatigue life of the inner container, the trained deep neural network model is used to optimize the design parameters of the hydrogen storage container, and the optimal design scheme is found. Compared with the traditional finite element simulation fatigue calculation which can only qualitatively obtain the influence law of some design parameters on the fatigue life of the container, the deep learning model can quantitatively analyze the corresponding relationship between the design parameters of the inner container and the composite layer and the fatigue life, and based on the prediction result, the optimization design is carried out, and the lightweight demand is considered, so that the fatigue performance of the container is improved and the cost is reduced as much as possible.

[0006] The technical scheme adopted by the application is: an anti-fatigue and lightweight type III hydrogen storage container optimization design method, which comprises the following steps:

[0007] Step 100: determining key design variables according to design targets of a type III composite material hydrogen storage container to be optimized, wherein the design targets include inner container fatigue life, inner container mass and composite layer mass, and the key design variables include inner container wall thickness, head shape and composite layer layup scheme;

[0008] Step 101: characterizing the key design variables of the hydrogen storage container, wherein the inner container wall thickness is represented by t, the head shape is represented by an ellipsoid ratio e of the head, the composite layer layup scheme is defined by a multi-dimensional parameter system, including thickness characteristics, arrangement characteristics and angle characteristics, the thickness characteristics include the number of spiral layers l α and the number of hoop layers l θ , the arrangement characteristics include the number of spiral layer segments s α and the number of hoop layer segments s θ , and the angle characteristics include the number of spiral layer angles d α and the spiral layer angle arrangement mode c;

[0009] Step 102: determining the value range and discretization accuracy of the characterization parameters of the key design variables according to design requirements and actual processing and manufacturing conditions, and performing Latin hypercube sampling on the design parameters;

[0010] Step 103: generating a random composite layer layup scheme according to the sampling results, first generating a spiral layer length list according to the number of spiral layer segments s α and the number of spiral layers l α , each element in the list represents the number of layers of the spiral layer segment, the list length is equal to the number of spiral layer segments s α , and the sum of all elements in the list is equal to the number of spiral layers l α, and the same method is used to generate the circumferential layer length list; then according to the spiral layer angle number d α , the spiral layer angle arrangement mode c and the spiral layer segment number s α Generate a spiral layer angle list, each element in the list represents the angle of the spiral layer segment, and the list length is equal to the spiral layer segment number s α , the number of elements in the list is equal to the spiral layer angle number d α , the arrangement of the elements in the list refers to the spiral layer angle arrangement mode c, which must contain the minimum winding angle θ tangent to the polar hole imin , the remaining winding angles are randomly selected from the preset angle range and distributed according to the specified arrangement mode, so that the length sequence and the angle sequence correspond to each other to represent the specific parameters of each segment of the winding layer. Finally, the circumferential layer sequence is randomly inserted into the spiral layer sequence to form a complete composite layer laying scheme, and the inner liner wall thickness t and the head ellipsoid ratio e together constitute a complete sample data;

[0011] Step 200: Based on the sample data generated by sampling and randomization, a finite element model of the type III hydrogen storage container is established in the finite element software, the performance parameters of the material are input, and the inner liner and the composite layer mass are extracted. Then, numerical simulation analysis is carried out under fatigue cycle conditions, the stress distribution of the inner liner is calculated, and the simulation data is imported into the fatigue analysis software to obtain the fatigue life distribution and the minimum fatigue life value of the inner liner;

[0012] Step 201: Collect and organize all sample data, perform numerical simulation calculation and record the corresponding inner liner mass, composite layer mass and inner liner fatigue life of each sample, and construct a structured database;

[0013] Step 202: Construct a deep neural network model, wherein the input layer includes the inner liner wall thickness of the hydrogen storage container, the head shape and the composite layer laying scheme, and the output layer corresponds to the key performance indicators: inner liner mass, composite layer mass and inner liner fatigue life. The database constructed in step 201 is used to train the deep neural network model;

[0014] Step 300: Define a multi-objective optimization problem, while ensuring that the inner liner fatigue life meets the design requirements, maximize the fatigue life and minimize the inner liner mass and the composite layer mass;

[0015] Step 301: Use the multi-objective optimization algorithm NSGA-II to solve the multi-objective optimization problem, and select the optimal design scheme from the Pareto solution set.

[0016] Further, the spiral layer angle arrangement mode c defines the arrangement rule of the spiral layer angle, including three typical modes of monotonic increasing, monotonic decreasing and random distribution.

[0017] Further, in step 102, the spiral layer number lα and the number of spiral layers l θ The value range of the number of spiral layers s α is determined according to the extracted number of spiral layers, and the number of spiral layers s θ and the number of spiral layers d α is less than or equal to the extracted number of spiral layers.

[0018] Further, in step 200, stress analysis calculation is completed in finite element software Abaqus to obtain the stress distribution result of the liner, then the generated ODB result file is imported into fatigue analysis software Fe-safe, a corresponding material model is established and S-N curve data of the material is input, Brown-Miller equation is selected as the fatigue damage calculation method, and the fatigue life distribution of the liner is obtained.

[0019] Further, the deep neural network model divides the training set and the validation set during training, and performs normalization preprocessing on the data, adopts ReLU as the activation function, and adopts mean square error loss function as the optimization target.

[0020] Further, the deep neural network model learns the complex nonlinear relationship between the design parameters P of the hydrogen storage container, the container mass C and the liner fatigue life N, establishes a mapping function f: P→(C, N), and can efficiently search the design parameter space based on the mapping function, and predict the material cost and the minimum fatigue life under different design parameter combinations.

[0021] Further, in step 300, the optimization variables include the wall thickness t of the liner, the ellipsoidal ratio e of the head, and the actual winding angle θ i of each ply of the composite layer, and the optimization objective function includes two aspects, one is to maximize the liner fatigue life N, and the other is to minimize the container material cost m l *c l +m c *c c , wherein m l and m c represent the mass of the liner and the composite layer respectively, c l and c c represent the average unit price of the material, and the constraint target is that the liner fatigue life is greater than the minimum fatigue life N design required by the design requirement, in the variable constraint aspect, the value range of t and e is uniformly mapped to maintain the coordination with the value range of θ i , wherein the value of θ i includes discrete values and continuous values, the discrete values are the minimum winding angle θ imin tangent to the polar hole or the hoop winding angle 90.0°.

[0022] Furthermore, the multi-objective optimization algorithm NSGA-II is used to solve the multi-objective optimization problem, specifically as follows: first, the design variables and their constraints are clarified and a multi-objective optimization model is established, an initial population is generated and encoded, and a pre-trained deep neural network model is used to predict the inner liner fatigue life, inner liner quality and composite layer quality corresponding to each design scheme. Then, a population hierarchical evaluation is implemented based on non-dominated sorting and crowding distance calculation, and a child population is generated through selection, crossover and mutation operations. A repair operator is used to ensure the feasibility of the solution, and the performance evaluation and non-dominated sorting process are repeated on the new generation population. The convergence is judged according to the preset convergence criterion. If the convergence condition is not reached, the iteration is continued until a stable Pareto front is obtained. Finally, the optimal design scheme is screened from the Pareto solution set through a multi-attribute decision analysis method.

[0023] Furthermore, in the process of solving the multi-objective optimization problem:

[0024] (1) Define the coding rules. All optimization variables are coded using real numbers and retain one decimal place. The inner liner wall thickness t and the head ellipsoid ratio e are placed in the first and second positions of the variable sequence respectively. The subsequent angle variable θ i represents the actual winding angle of the i-th layer in turn;

[0025] (2) Define the crossover operator. Randomly select two individuals p1 and p2 from the parent population for crossover operation. First, determine the length of the offspring individual by comparing the number of variables between the two. If the lengths of the parents are equal, the offspring after crossover will have the same length as the parents. If they are not equal, the length of the offspring will be a random value between the lengths of the two parents. For parents with excessive length, the extra variables will be discarded from the end. For parents with insufficient length, random variable values ​​within the value range will be filled from the end. Then, simulate binary crossover for all continuous variables at the same position of the two parents. Subsequently, variable segments of random lengths will be exchanged at randomly selected sites to finally generate fully mixed offspring individuals c1 and c2.

[0026] (3) Define the mutation operator and perform mutation operations in the offspring population. First, perform length mutation. For individuals with an original length of l0, the length after mutation is randomly selected in the interval [l0-4, l0+4], but it will not exceed the maximum value of the length nor be less than the minimum value of the length. Then perform numerical mutation. First, θ i The parameters are randomly converted between continuous and discrete values, with the conversion probability determined by the ratio of the two types of values ​​in the initial population. Then, a dynamic hybrid mutation strategy is implemented for all continuous variables, using a combination of polynomial mutation and uniform mutation, and the application ratio of the two is dynamically adjusted with the number of evolution generations.

[0027] (4) define repair operators, after the population is crossed and mutated, check and repair all individuals to ensure that the generated individuals have practical significance, first normalize the variable length and value, ensure that the individual length is in the preset range, all variable values are forced to be constrained in the definition domain and uniformly retain one decimal precision, then implement angle distribution correction, forced to maintain 40%-60% of the proportion of circumferential layer angles and 10%-20% of the proportion of the minimum winding angle tangent to the polar hole, when the number of specific angles is insufficient, randomly select continuous value angles for corresponding conversion, and when the proportion exceeds the standard, then the excessive number is randomly converted into the existing continuous value angle.

[0028] Further, after obtaining the Pareto solution set of the multi-objective optimization problem by the multi-objective optimization algorithm NSGA-II, first, the entropy weight method is applied to objectively allocate the weights of the two optimization targets of fatigue life and material cost, the weight coefficients are automatically calculated by analyzing the dispersion degree of each target value, the TOPSIS algorithm is used for comprehensive evaluation and sorting of the Pareto solution set, and the optimal balanced solution is finally determined by calculating the relative closeness of each solution to the ideal solution and the negative ideal solution.

[0029] The beneficial effects of the present application are embodied in:

[0030] Compared with the existing design parameter optimization method of the composite hydrogen storage container, the design space is further expanded, and multiple design parameters such as the number of composite layers, angles, arrangement modes and head shapes are increased. This comprehensive parameter consideration makes the optimization process more precise, better captures the influence of design variables on the performance of the hydrogen storage container, and can search for the global optimal solution in a larger design space, thereby significantly improving the performance of the hydrogen storage container. In addition, most of the existing methods are based on the optimization of layer angles based on the burst strength, mainly focusing on the load capacity of the container under extreme pressure, while the present method optimizes the design for the fatigue life of the liner. For fixed hydrogen storage containers, fatigue failure is a more common failure mode. During the long-term use of fixed hydrogen storage containers, they will undergo frequent charging and discharging cycles, resulting in periodic stress on the liner and causing the accumulation of fatigue damage, which is more concealed and progressive. Therefore, the present method optimizes the design parameters of the liner and the composite layer to improve the fatigue resistance of the liner, prolong the service life of the container, and ensure its safety and reliability during long-term service. The optimization method of the present application is not only suitable for theoretical research, but also can be directly applied to engineering practice. Through the combination of experimental verification and numerical simulation, a high-performance hydrogen storage container design scheme can be quickly generated, overcoming the disadvantage that the carbon fiber composite material layer can only be optimized and corrected through experiments and experience. It provides strong support for the development of hydrogen energy storage technology and has important theoretical significance and application value. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 An anti-fatigue and lightweight type III composite hydrogen storage container optimization design method implementation flowchart is provided for the present application;

[0032] Figure 2 A type III hydrogen storage container structure and size schematic diagram is provided for the present application embodiment;

[0033] Figure 3 A hydrogen storage container finite element model schematic diagram is provided for the present application embodiment;

[0034] Figure 4 A liner fatigue life cloud atlas simulation result in fatigue analysis before optimization of the hydrogen storage container model is provided for the present application embodiment;

[0035] Figure 5 A Latin hypercube sampling sampling result is provided for the present application embodiment;

[0036] Figure 6 A sample data schematic diagram randomly generated is provided for the present application embodiment;

[0037] Figure 7 A multi-objective optimization process flowchart is provided for the present application embodiment;

[0038] Figure 8 An encoding, crossover and mutation operation schematic diagram is provided for the present application embodiment;

[0039] Figure 9 A liner fatigue life cloud atlas simulation result in fatigue analysis after optimization of the hydrogen storage container model is provided for the present application embodiment. DETAILED DESCRIPTION

[0040] The present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only for the purpose of explaining the present application, and not for limiting the present application, and the drawings only show the contents related to the specific embodiments, and not the entire contents of the present application.

[0041] Figure 1 An anti-fatigue and lightweight type III composite hydrogen storage container optimization design method implementation flowchart is provided for the present application embodiment, including the following steps:

[0042] Step 100: as shown in Figure 2 , first determine the basic size of the container liner according to the design requirements and actual processing and manufacturing conditions, and calculate the required spiral layer thickness t fα and hoop layer thickness t fθ18.4mm and 20.3mm respectively, the single layer thickness of the composite layer is measured as 0.386mm, the number of spiral layers n is calculated based on the single layer thickness and rounded up to an even number α and the number of hoop layers n θ 48 and 54 respectively, the initial layup scheme is determined by empirically reaming at each bandwidth respectively as shown in Table 1, wherein the calculation formula of the grid theory is as follows:

[0043]

[0044] wherein: t fα is the spiral layer thickness; t fθ is the hoop layer thickness; R is the outer radius of the cylinder; P b is the design burst pressure; K is the spiral fiber strength utilization coefficient, generally taken as 0.6-0.8; σ b is the tensile strength limit of the fiber direction of the composite material; and a is the minimum winding angle tangent to the extreme hole;

[0045] Table 1. Initial layup scheme of the hydrogen storage container composite layer

[0046]

[0047]

[0048] Step 101: modeling the initial design scheme in the finite element software Abaqus to obtain a finite element model as shown in Figure 3 According to the finite element model and the material parameters, the mass of the liner and the composite layer is 472.420kg and 191.672kg respectively. The modeling method simulates the production process, i.e. laying, assembling, meshing and defining material properties for each layer, and finally obtaining the finite element model of the hydrogen storage container assembly body containing all structures except the valve and sealing device. The model parameters include geometric parameters, material parameters and layup parameters. The size includes the size of the liner and the size of the hydrogen storage container bottleneck structure; the liner material is 316L stainless steel and the composite layer material is T700 carbon fiber; the material parameters are shown in Tables 2 and 3; the layup parameters include the thickness, angle and arrangement of each layer; the boundary conditions are fixed at one end and free at the other end, and the internal pressure is applied on the inner surface, the self-tightening pressure is 127.5MPa, and the cyclic load is 2-102MPa. In order to save computing resources, an axisymmetric model is used for modeling, and solid elements CAX4R are used for meshing the liner and the composite layer;

[0049] Table 2. Mechanical parameters of 316L stainless steel

[0050] E / MPa μ [R p / MPa]]> [R m / / MPa]]> δ / % ρ / g cm -3 ]]> 200 0.3 279 546 77.0 7.98

[0051] Table 3. Mechanical parameters of T700 carbon fiber

[0052]

[0053]

[0054] Step 102: The stress distribution results of the inner liner are obtained by completing stress analysis calculation in Abaqus, then the generated ODB result file is imported into Fe-safe fatigue analysis software, the corresponding material model is established and the S-N curve data of the material is input, in setting the analysis parameters, according to the actual working condition, the surface roughness of the inner liner is set to 6.3 μm, and Brown-Miller equation is selected as the fatigue damage calculation method, finally the fatigue life cloud map of the inner liner is obtained as shown in Figure 4 ;

[0055] Step 103: On the basis of the related design parameters of the initial design scheme, the initial design scheme is optimized and designed, the design targets include the fatigue life of the inner liner, the mass of the inner liner and the mass of the composite layer, the design variables include the wall thickness of the inner liner, the shape of the head and the lamination scheme of the composite layer, the rest of the parameters of the inner liner are the same as those in the initial design scheme, the wall thickness of the inner liner is represented by t, the head is an ellipsoidal head, the geometric characteristics of the head are represented by the ellipsoidal ratio of the head e (the ratio of the long axis to the short axis), the lamination scheme of the composite layer is completely defined by a multi-dimensional parameter system, including thickness characteristics (the number of spiral layers l α , the number of hoop layers l θ ), arrangement characteristics (the number of spiral layer segments s α , the number of hoop layer segments s θ ) and angle characteristics (the number of spiral layer angles d α , the arrangement mode of spiral layer angles c), wherein l α and l θ represent the total number of spiral and hoop winding layers in the composite layer, the winding layers with equal and continuous arrangement angles are recorded as a segment, s α and s θ describe the segmentation of the continuous and equal angle winding layers, and d α represents the number of discrete angles of the spiral layer.

[0056] According to the design requirements and the actual processing and manufacturing conditions, the value range and the discretization accuracy of the eight representative parameters (t, e, l α , l θ , s α , s θ , d α , c) of the inner liner and the composite layer design variables are determined, as shown in Table 4, wherein the value range of the number of spiral layers l α and the number of hoop layers l θ is calculated by the grid theory, and then the left and right is shifted by 8 layers, the number of spiral layer segments s αThe value range of the spiral layer is determined according to the extracted spiral layer number, and the number of annular layer segments s θ and the spiral layer angle number d α is less than or equal to the extracted spiral layer segment number, c defines the arrangement rule of the spiral layer angle, including three typical modes of monotone increasing, monotone decreasing and random distribution, represented by 1, 2 and 3 respectively, Latin hypercube sampling is performed on the eight design parameters to generate experimental samples with good space filling and projection uniformity, ensuring that each design parameter is fully covered in its value range, a total of 100 groups are sampled, and the sampling results are shown in Figure 5 ;

[0057] Table 4. Design parameter value range and discretization accuracy

[0058] Parameter t e l α ]] l θ ]]> s α ]]> s θ ]]> d α ]]> c Range 12～22 1～2 40～56 46～62 4-1 α / 4]]> 3~s α ]] 4~s α ]] 1,2,3 Accuracy 0.5 0.05 2 2 1 1 1 1

[0059] Step 104: generating a random composite layer layup scheme according to the sampling results, first generating a random length list of spiral layers according to the spiral layer segment number s α and the spiral layer number l α , wherein each element in the list represents the number of layers of the spiral layer segment, the list length is equal to the spiral layer segment number s α , and the sum of all elements in the list is equal to the spiral layer number l α , while the difference between the maximum and minimum values in the list needs to be controlled to avoid the number of layers being too concentrated, and a random length list of annular layers is generated in the same way, then a random angle list of spiral layers is generated according to the spiral layer angle number d α , the spiral layer angle arrangement mode c and the spiral layer segment number s α , wherein each element in the list represents the angle of the spiral layer segment, the list length is equal to the spiral layer segment number s α , the number of elements in the list is equal to the spiral layer angle number d α , and the arrangement of the elements in the list refers to the spiral layer angle arrangement mode c, wherein the minimum winding angle θ imin = 19.2 must be included, and the remaining winding angles are randomly selected from the range [θ imin + 5.0, 60.0] i.e. [24.2, 60.0] and distributed according to the specified arrangement mode, so that the length sequence and the angle sequence correspond to each other to represent the specific parameters of each winding layer, finally the annular layer sequence is randomly inserted into the spiral layer sequence to form a complete composite layer layup scheme, and combined with the inner liner wall thickness t and the head ellipsoid ratio e to form a complete sample data, as shown in Figure 6 , for each sampling result, 10 random composite layer layup schemes are generated, a total of 1000 sample data can be obtained;

[0060] Step 200: Develop an automated program for Abaqus finite element software to make modeling, loading, calculation, analysis, and result export an automated process to simplify subsequent large-scale calculations. Based on 1000 sample data generated by sampling and randomization, refer to the previous calculation steps, establish a refined finite element model of type III hydrogen storage container in Abaqus, input the performance parameters of the material and extract the total mass of the liner and composite layer, then carry out numerical simulation analysis under fatigue cycle conditions, calculate the stress distribution of the liner, and import the simulation data into Fe-safe to finally obtain the fatigue life distribution of the liner and its minimum fatigue life value.

[0061] Step 201: The system collects and organizes all sample data, performs numerical simulation calculations, and records key parameters such as liner mass, composite layer mass, and liner fatigue life corresponding to each sample. Based on these multi-dimensional performance indicators, a structured database is constructed.

[0062] Step 202: Construct a deep neural network model, where the input layer contains the structural parameters of the hydrogen storage container liner and the composite layer layup scheme, and the output layer corresponds to three key performance indicators: liner mass, composite layer mass, and liner fatigue life. ReLU is used as the activation function to enhance the non-linear expression ability of the model. At the same time, the mean square error loss function is selected as the optimization target. The complete sample data collected is divided into training set t1 and test set t2 according to the preset proportion (such as 8:2), where the training set is used for model training, and the test set is used for model performance evaluation. All target parameters are normalized to avoid numerical instability and model convergence difficulties. The activation function and loss function are respectively:

[0063] ReLU(x) = max(0, x)

[0064]

[0065] where: ReLU(x) is the activation function, x is the variable of the activation function, MSELoss is the loss function, n is the total number of samples, y i is the true result of the i-th sample, is the predicted result of the i-th sample.

[0066] Step 203: After the training of the deep neural network model is completed, the trained deep neural network model is evaluated for performance using an independent test set to verify the accuracy of the prediction results and the generalization ability. When the training set and the test set both achieve satisfactory prediction results under the same deep neural network model, the training is completed, and a trained deep neural network model is obtained. The deep neural network model learns the complex nonlinear relationship between the hydrogen storage container design parameters P and the container mass C and the liner fatigue life N, establishes a mapping function f: P→(C, N), and based on this mapping relationship, the deep neural network model can efficiently search the design parameter space and predict the material cost and the minimum fatigue life under different design parameter combinations, providing data support for subsequent optimization;

[0067] Step 300: Define a multi-objective optimization problem, the core of which is to maximize the fatigue life of the liner and minimize the mass of the liner and the composite layer while ensuring that the fatigue life of the liner meets the design requirements. The optimization variables include the liner wall thickness t, the ellipsoid ratio of the head e, and the composite layer lamination scheme θ i (where subscript i represents the lamination order, θ i represents the actual winding angle of each lamination), the number of parameters is variable, depending on the value range of the number of hoop layers and the number of spiral layers during sampling, which is 45-61 here. The optimization objective function includes two aspects: maximizing the fatigue life N of the liner and minimizing the material cost m l *c l +m c *c c , where m l and m c represent the mass of the liner and the composite layer, c l and c c represent the average unit price of the material, i.e., the weight coefficient of the mass term. The constraint target is that the fatigue life of the liner is greater than the minimum fatigue life N design required by the design requirements, which is 25000 times here, but considering a certain safety margin, the actual optimization takes 28000 times. In terms of variable constraints, the value range of t and e is uniformly mapped to the interval [20.0, 70.0] to maintain coordination with the value range of θ i , where the value of θ i includes discrete values (the minimum winding angle θ imin = 19.2° tangent to the extreme hole or the hoop winding angle 90.0°) and continuous values (the hole expansion winding angle range [24.2, 60.0]), and the parameter value precision is 0.1, which is mathematically represented as:

[0068] Variable t, e, θ1, θ2, θ3, θ4, θ5… θ i

[0069] N_variable 45~61

[0070] maximum N

[0071] Minimum m l ·c l +m c ·c c

[0072] Subject to N≥28000

[0073] 20.0≤t,e≤70.0

[0074] 24.2≤θ i ≤60.0 orθ imin or 90.0

[0075] Step 301: the multi-objective optimization problem is solved using the multi-objective optimization algorithm NSGA-II, and the complete process includes: first, the design variables and their constraints are determined and a multi-objective optimization model is established, an initial population is generated and coded, a pre-trained deep neural network model is used to predict the key performance indicators of each design scheme, such as the liner fatigue life, the liner mass and the composite layer mass, then the non-dominated sorting and crowding distance calculation are implemented to evaluate the population, the offspring population is generated through genetic operations such as selection, crossover and mutation, and the repair operator is used to ensure the feasibility of the solution, the performance evaluation and non-dominated sorting process are repeated for the new population, the convergence criterion is determined according to the pre-set convergence criterion, and the algorithm converges until a stable Pareto front is obtained, and finally the optimal design scheme is selected from the Pareto solution set through multi-attribute decision analysis method, as shown in Figure 7 ;

[0076] Step 302: define the coding rule, all optimization variables are coded with real numbers and one decimal place is reserved, the liner wall thickness t and the head ellipsoid ratio e are placed in the first and second positions of the variable sequence, respectively, and the two parameters take continuous real values in the interval [20.0, 70.0], and the subsequent angle variables θ i represent the actual winding angle of the i-th layer of the layer in turn, which takes discrete values in {19.2, 90.0} or continuous values in [24.2, 60.0], as shown in Figure 8 ;

[0077] Step 303: Define the crossover operator, randomly select two individuals p1 and p2 from the parent population for crossover operation, first determine the length of the offspring individual by comparing the number of variables between the two, if the lengths of the parents are equal, the offspring after crossover will maintain the same length as the parents, if they are not equal, the length of the offspring will be a random value between the lengths of the two parents, for parents with excessive lengths, the redundant variables will be discarded from the end, for parents with insufficient lengths, random variable values ​​within the value range will be filled from the end, then simulate binary crossover for all continuous variables at the same position of the two parents, then exchange variable segments of random length at randomly selected sites, and finally generate fully mixed offspring individuals c1 and c2, as shown in Figure 8 As shown;

[0078] Step 304: Define the mutation operator and perform mutation operations in the offspring population. First, perform length mutation. For individuals with an original length of l0, the length after mutation is randomly selected in the interval [l0-4, l0+4], but it will not exceed the maximum length (61) nor be less than the minimum length (45). Then perform numerical mutation. First, θ i The parameter randomly switches between continuous and discrete values, and its conversion probability is determined by the ratio of the two types of values ​​in the initial population. Here, the probability of taking the value of 19.2 is 15%, the probability of taking the value of 90.0 is 50%, and the probability of taking the value of the continuous angle value is 35%. Then, a dynamic hybrid mutation strategy is implemented for all continuous variables, using a combination of polynomial mutation and uniform mutation. The application ratio of the two is dynamically adjusted with the evolutionary generation number, which is expressed as:

[0079] ratio=max(0,1.0-(Gen / Total_Gen) 0.5 )

[0080] In the formula, ratio is the proportion of uniform mutation, Gen is the current number of iterations, and Total_Gen is the total number of iterations. In the early stage, uniform mutation is mainly used to ensure global search capability, and in the later stage, polynomial mutation is mainly used to enhance local development capability to ensure convergence. For example, Figure 8 As shown;

[0081] Step 305: defining a repair operator, after the population is subjected to crossover and mutation operations, a check and repair operation is performed on all individuals to ensure that the generated individuals have practical significance, first, the variable length and value are normalized to ensure that the individual length is within the preset range (if insufficient, the tail is randomly completed with valid variables, and if excessive, the excess part is truncated), all variable values are forced to be within the defined domain (if out of range, take the boundary value) and uniformly retain one decimal place of accuracy, then implement angle distribution correction to ensure that the ring layer angle (90.0°) maintains a proportion of 40%-60%, and the minimum winding angle (19.2°) tangent to the polar hole maintains a proportion of 10%-20%, when the number of specific angles is insufficient, a continuous value angle is randomly selected for corresponding conversion, and when the proportion exceeds the standard, the excess number is randomly converted to an existing continuous value angle;

[0082] Step 306: find the Pareto optimal solution set through the multi-objective optimization algorithm NSGA-II, which represents the design scheme that achieves the best balance between container mass and inner liner minimum fatigue life, i.e., minimizes cost under the premise of ensuring a certain fatigue life, or maximizes fatigue life under the given cost constraint, this step provides the designer with multiple feasible optimization design schemes for selection according to actual engineering requirements, in this embodiment, according to actual design requirements, the container mass and inner liner minimum fatigue life of any layer angle within the search range can be predicted, for example, for a type III hydrogen storage container with a design pressure of 102 MPa, testing a single layer scheme using the experimental method requires about a month, using finite element simulation to model and calculate a single axisymmetric solid model takes about 20 minutes, establishing 1000 groups of data takes about 20 days, and using the trained deep neural network model, about 10 million layer schemes can be calculated and evaluated in the entire design parameter space within 1 hour. For the obtained Pareto solution set, first apply the entropy weight method to objectively allocate weights to the two optimization targets of fatigue life and material cost, automatically calculate the weight coefficients as 0.477 and 0.523 respectively by analyzing the dispersion of each target value, thereby avoiding decision-making bias caused by subjective assignment, and then use the TOPSIS algorithm to comprehensively evaluate and sort the Pareto solution set, by calculating the relative closeness of each solution to the ideal solution and the negative ideal solution, the optimal balanced solution that takes into account fatigue performance and economic benefits is finally determined;

[0083] Step 307: Establish the corresponding optimized finite element model for verification, accurately evaluate the minimum fatigue life of the inner container of the container through simulation calculation, and estimate the actual container quality combined with the manufacturing process. Compare the finite element analysis results with the prediction results of the deep neural network model. If the two results are basically consistent within the allowable error range, and the fatigue life or cost of the design scheme after optimization is significantly improved compared with the design scheme before optimization, the effectiveness and engineering practicability of the method are verified. The total mass of the inner container after optimization is reduced from 472.420 kg to 461.512 kg, which is reduced by 10.908 kg, about 2.3%, and the total mass of the composite layer is reduced from 191.672 kg to 165.776, which is reduced by 25.896 kg, about 13.5%. The minimum fatigue life of the inner container is increased from 28115 to 28208, which is increased by 93 times. The wall thickness of the inner container of the design scheme is 19.6 mm, the ellipsoid ratio of the head is 1.05, and the composite layer stacking scheme is shown in Table 5. The simulation results of the fatigue life cloud map of the optimized inner container are shown in Figure 9

[0084] Table 5. Stacking scheme of the composite layer of the hydrogen storage container after optimization

[0085]

[0086]

[0087] The above is a further detailed description of the preferred embodiments of the present application, which is not a limitation of the present application. It should be noted that for ordinary skilled persons in the technical field to which the present application belongs, any simple deduction and optimization of the present application based on the core idea of the present application should be considered within the protection scope of the present application.​

Claims

1. A fatigue-resistant and lightweight type III hydrogen storage container optimization design method, characterized in that: The following steps are involved: Step 100: Determine key design variables based on the design objectives of the Type III composite hydrogen storage container to be optimized, wherein the design objectives include liner fatigue life, liner quality, and composite layer quality, and the key design variables include liner wall thickness, head shape, and composite layer layup scheme; Step 101: Characterize the key design variables of the hydrogen storage container. The inner wall thickness is characterized by t, and the head shape is characterized by the head ellipsoid ratio e. The composite layer layup scheme is defined using a multi-dimensional parameter system, including thickness characteristics, arrangement characteristics, and angle characteristics. The thickness characteristics include the number of spiral layers l α and the number of circumferential layers l θ , the arrangement characteristics include the number of spiral layers s α and the number of circumferential segments s θ , the angle characteristics include the helical layer angle d α and the spiral layer angle arrangement c; Step 102: Determine the value range and discretization accuracy of the key design variable characterization parameters based on the design requirements and actual processing and manufacturing conditions, and perform Latin hypercube sampling on the design parameters; Step 103: Generate a random composite layer layup plan based on the sampling results. First, according to the number of spiral layer segments s α and the number of helical layers l α Generate a list of spiral layer lengths, each element in the list represents the number of layers in the spiral layer, and the length of the list is equal to the number of spiral layer segments s α , the sum of all elements in the list is equal to the number of spiral layers l α , the same method is used to generate the circumferential layer length list; then according to the spiral layer angle d α , spiral layer angle arrangement c and spiral layer segment number s α Generate a list of spiral layer angles. Each element in the list represents the angle of the spiral layer in that segment. The length of the list is equal to the number of spiral layer segments s. α , the number of elements in the list is equal to the number of spiral layer angles d α The arrangement of the elements in the list refers to the angle arrangement of the helical layer c, which must include the minimum winding angle θ tangent to the polar hole imin The remaining winding angles are randomly selected from the preset angle range and distributed according to the specified arrangement, so that the length sequence and the angle sequence correspond one-to-one to represent the specific parameters of each winding layer. Finally, the circumferential layer sequence is randomly inserted into the helical layer sequence to form a complete composite layer layup scheme, and combined with the liner wall thickness t and the head ellipsoid ratio e to form a complete sample data; Step 200: Based on the sampled and randomly generated sample data, a finite element model of the Type III hydrogen storage container is established in the finite element software. The material performance parameters are input and the mass of the liner and the composite layer is extracted. Then, a numerical simulation analysis is performed under fatigue cycle conditions to calculate the stress distribution of the liner. The simulation data is imported into the fatigue analysis software to obtain the fatigue life distribution and minimum fatigue life value of the liner. Step 201: Collect and organize all sample data, perform numerical simulation calculations, and record the liner mass, composite layer mass, and liner fatigue life corresponding to each sample to build a structured database; Step 202: Construct a deep neural network model, wherein the input layer includes the inner liner wall thickness, head shape, and composite layer layup scheme of the hydrogen storage container, and the output layer corresponds to the key performance indicators: inner liner quality, composite layer quality, and inner liner fatigue life. The deep neural network model is trained using the database constructed in step 201; Step 300: Define a multi-objective optimization problem to maximize the fatigue life and minimize the mass of the liner and the composite layer while ensuring that the fatigue life of the liner meets the design requirements; Step 301: Use the multi-objective optimization algorithm NSGA-II to solve the multi-objective optimization problem, and select and determine the optimal design solution from the Pareto solution set.

2. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: The spiral layer angle arrangement mode c defines the arrangement rule of the spiral layer angle, including three typical modes: monotonically increasing, monotonically decreasing and random distribution.

3. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: In step 102, the number of spiral layers is l α and the number of circumferential layers l θ The value range of is obtained by calculating the reference value through grid theory and then offsetting the setting range left and right. The number of spiral segments s α The value range of is determined by the number of spiral layers extracted, and the number of circumferential layers s θ and the helical layer angle d α Must be less than or equal to the number of spiral segments extracted.

4. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: In step 200, stress analysis calculations are completed in the finite element software Abaqus to obtain the stress distribution results of the inner liner, and then the generated ODB result file is imported into the fatigue analysis software Fe-safe, the corresponding material model is established and the SN curve data of the material is input, and the Brown-Miller equation is selected as the fatigue damage calculation method to obtain the fatigue life distribution of the inner liner.

5. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: The deep neural network model is divided into a training set and a validation set during training, and the data is normalized and preprocessed. ReLU is used as the activation function and the mean square error loss function is used as the optimization target.

6. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: The deep neural network model learns the complex nonlinear relationship between the design parameters P of the hydrogen storage container and the container mass C and the fatigue life N of the liner, and establishes a mapping function f:P→(C,N). Based on the mapping function, the design parameter space can be efficiently searched and the material cost and minimum fatigue life under different design parameter combinations can be predicted.

7. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: In step 300, the optimization variables include the liner wall thickness t, the head ellipsoid ratio e, and the actual winding angle θ of each composite layer. i ,The optimization objective function includes two aspects, one is to maximize the fatigue life of the liner N, and the other is to minimize the container material cost m l *c l +m c *c c , where m l and m c Represent the mass of the liner and the composite layer respectively, c l and c c represents the average unit price of the material. The constraint target is that the fatigue life of the liner should be greater than the minimum fatigue life value N specified in the design requirements. design In terms of variable constraints, the value ranges of t and e are uniformly mapped to maintain the same i The coordination of the value range, where θ i The value of includes discrete value and continuous value. The discrete value is the minimum winding angle θ tangent to the pole hole. imin Or the hoop winding angle is 90.0°.

8. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 1 is characterized in that: The multi-objective optimization algorithm NSGA-II is used to solve the multi-objective optimization problem. Specifically, the design variables and their constraints are first clarified and a multi-objective optimization model is established. An initial population is generated and encoded. A pre-trained deep neural network model is used to predict the inner liner fatigue life, inner liner quality, and composite layer quality corresponding to each design scheme. Then, a population hierarchical evaluation is performed based on non-dominated sorting and crowding distance calculation. A child population is generated through selection, crossover, and mutation operations, and a repair operator is used to ensure the feasibility of the solution. The performance evaluation and non-dominated sorting processes are repeated on the new generation population. Convergence is judged according to a preset convergence criterion. If the convergence condition is not met, iteration is continued until a stable Pareto front is obtained. Finally, the optimal design scheme is screened from the Pareto solution set through a multi-attribute decision analysis method.

9. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 8, characterized in that: In the process of solving the multi-objective optimization problem: (1) Define the coding rules. All optimization variables are coded using real numbers and retain one decimal place. The inner liner wall thickness t and the head ellipsoid ratio e are placed in the first and second positions of the variable sequence respectively. The subsequent angle variable θ i represents the actual winding angle of the i-th layer in turn; (2) Define the crossover operator. Randomly select two individuals p1 and p2 from the parent population for crossover operation. First, determine the length of the offspring individual by comparing the number of variables between the two. If the lengths of the parents are equal, the offspring after crossover will have the same length as the parents. If they are not equal, the length of the offspring will be a random value between the lengths of the two parents. For parents with excessive length, the extra variables will be discarded from the end. For parents with insufficient length, random variable values ​​within the value range will be filled from the end. Then, simulate binary crossover for all continuous variables at the same position of the two parents. Subsequently, variable segments of random lengths will be exchanged at randomly selected sites to finally generate fully mixed offspring individuals c1 and c2. (3) Define the mutation operator and perform mutation operations in the offspring population. First, perform length mutation. For individuals with an original length of l0, the length after mutation is randomly selected in the interval [l0-4, l0+4], but it will not exceed the maximum value of the length nor be less than the minimum value of the length. Then perform numerical mutation. First, θ i The parameters are randomly converted between continuous and discrete values, with the conversion probability determined by the ratio of the two types of values ​​in the initial population. Then, a dynamic hybrid mutation strategy is implemented for all continuous variables, using a combination of polynomial mutation and uniform mutation, and the application ratio of the two is dynamically adjusted with the number of evolution generations. (4) Define the repair operator. After the population undergoes crossover and mutation operations, all individuals are checked and repaired to ensure that the individuals generated are meaningful. First, the variable lengths and values ​​are normalized to ensure that the individual lengths are within the preset range. All variable values ​​are constrained within the definition domain and uniformly retain one decimal place of precision. Then, the angle distribution correction is implemented to ensure that the proportion of the circumferential layer angle is maintained at 40%-60% and the proportion of the minimum winding angle tangent to the polar hole is maintained at 10%-20%. When the number of specific angles is insufficient, the continuous value angles are randomly selected for corresponding conversion. When the proportion exceeds the standard, the excess number is randomly converted to the existing continuous value angles.

10. The fatigue-resistant and lightweight type III hydrogen storage container optimization design method according to claim 8, characterized in that: After obtaining the Pareto solution set of the multi-objective optimization problem through the multi-objective optimization algorithm NSGA-II, the entropy weight method is first applied to objectively assign weights to the two optimization objectives of fatigue life and material cost. The weight coefficient is automatically calculated by analyzing the discrete degree of each objective value. The TOPSIS algorithm is used to comprehensively evaluate and rank the Pareto solution set. The optimal equilibrium solution is finally determined by calculating the relative proximity of each solution to the ideal solution and the negative ideal solution.

Images