A real-time prediction method for hydrogen storage cylinder fast charging temperature field

Abstract

The present application belongs to the field of hydrogen energy infrastructure safety, and discloses a real-time prediction method for the temperature field of hydrogen storage cylinder fast charging. The method first constructs a feature extractor to reduce the dimension and reconstruct the temperature field data obtained by the numerical simulation method, and extracts a low-dimensional feature vector; then a prediction module is constructed to realize accurate mapping from the hydrogen charging working condition parameters to the low-dimensional feature vector. In order to enhance the generalization ability of the model to random working conditions, an uncertainty quantification module is built to quantize the uncertainty of the prediction results of the feature vector, the posterior distribution of the output feature vector is jointly sampled to obtain the feature vector, the temperature field is reconstructed by the reconstruction module, and finally the temperature field prediction results with confidence interval are output in real time, providing quantitative decision basis for safety evaluation and optimization of hydrogen charging strategy.

Description

Technical Field

[0001] This invention belongs to the field of hydrogen energy infrastructure safety and relates to a real-time prediction method for the temperature field of fast charging hydrogen storage cylinders. Background Technology

[0002] Carbon fiber wound composite gas cylinders, due to their high strength and lightweight properties, have become a key component of hydrogen storage systems in new energy storage and transportation fields such as hydrogen fuel cell vehicles. However, during rapid filling, a dramatic temperature rise occurs inside the cylinder due to gas compression and other factors. When the temperature rise exceeds a certain limit, it affects the mechanical properties of the composite material, thereby jeopardizing the cylinder's load-bearing capacity and service life. Therefore, effectively controlling the temperature rise of hydrogen storage cylinders during fast charging has become one of the core issues for ensuring the safety of hydrogen energy storage and transportation.

[0003] Currently, research on the temperature rise during fast charging of hydrogen storage cylinders mainly relies on experimental measurements and numerical simulations. While both methods can provide high-precision temperature data, they are limited by experimental costs and computation time when facing diverse hydrogen charging conditions, especially the initial pressure of randomly remaining hydrogen in the cylinder. This makes it difficult to meet the requirements for real-time temperature field prediction and rapid evaluation of hydrogen charging strategies. Deep learning-based techniques can efficiently learn the mapping relationship between input and output from a limited number of data samples, enabling rapid inference on new data inputs based on good generalization ability. Gaussian processes, as a Bayesian nonparametric model, can provide an effective solution for quantifying the uncertainty of the aforementioned point estimation results. Therefore, combining deep learning with numerical simulation techniques to propose a method for real-time temperature field prediction and reliability assessment during the fast charging process of hydrogen storage cylinders is of great significance for the safe use of cylinders. Summary of the Invention

[0004] This invention addresses the need for rapid temperature field prediction in composite material gas cylinders under diverse fast-charging conditions by proposing a real-time prediction method for the temperature field of hydrogen storage cylinders during fast charging. Based on high-fidelity data generated through numerical simulation, this method first extracts low-dimensional feature vectors of the temperature field using a deep learning-based autoencoder. Then, a deep neural network is constructed to establish a nonlinear mapping relationship between operating parameters and the low-dimensional feature vectors. Furthermore, a multi-task Gaussian process is introduced to quantify the prediction uncertainty of the feature vectors, and a decoder transmits the uncertainty in the feature space to the physical space, ultimately achieving end-to-end real-time prediction and uncertainty quantification of the hydrogen storage cylinder temperature field.

[0005] The technical solution of the present invention:

[0006] A method for real-time prediction of the temperature field during fast charging of hydrogen storage cylinders, comprising the following steps:

[0007] Step 1: Establish a geometric model of the hydrogen storage cylinder using finite element software and discretize the model to obtain a discretized model;

[0008] Step 2: The discretized model obtained in Step 1 is imported into the numerical simulation software, and a numerical calculation model for fast charging of hydrogen storage cylinders is set. Different operating parameters are set as inlet boundary conditions and initial conditions, and multiple sets of high-fidelity temperature field data are obtained by solving.

[0009] Step 3: Process multiple sets of high-fidelity temperature field data to construct the input feature dataset and the temperature field dataset;

[0010] Step 4: To address the high-dimensionality and nonlinearity of the temperature field data, a feature extractor is constructed for dimensionality reduction and reconstruction. The feature extractor comprises an encoder and a decoder, built using an artificial neural network. The encoder consists of alternating layers of linear layers and ReLU activation function layers, with the number of neurons decreasing layer by layer at a fixed ratio, and the input layer dimension being the same as the number of nodes. The decoder adopts a structure symmetrical to the encoder. During training, the temperature field dataset described in Step 3 is used for unsupervised learning of the feature extractor. The trained feature extractor is used to extract low-dimensional feature vectors from all temperature field data, constructing an output feature dataset corresponding to the input feature dataset described in Step 3. The decoder with fixed parameters serves as the reconstruction module for reconstructing the temperature field from the low-dimensional feature vectors.

[0011] Step 5: Construct a prediction module consisting of one input layer, multiple hidden layers, and one output layer. Except for the output layer, which uses a linear layer, the input and hidden layers undergo nonlinear transformation using the ReLU function. The prediction module accepts operating condition parameters as input and maps them to the low-dimensional feature vector described in Step 4. After training, the prediction module quickly maps any new operating condition parameters to the corresponding low-dimensional feature vector, providing point estimation input for the uncertainty quantification module in Step 6.

[0012] Step 6: Construct an uncertainty quantification module, taking the output of the prediction module described in Step 5 as input, and output the posterior prediction mean and variance of the feature vector;

[0013] Step 7: Sample m feature vector samples from the posterior predicted mean and variance of the feature vectors output by the uncertainty quantization module described in Step 6, and reconstruct m temperature fields through the decoder of the feature extractor described in Step 4. Propagate the uncertainty in the feature space to the temperature field space, and perform statistics on each position in the temperature field space to obtain the corresponding mean and variance.

[0014] Step 8: Combining steps 4 to 7, integrate the trained prediction module, uncertainty quantification module, and reconstruction module. The data flow is as follows: the operating parameters are input into the prediction module and the point-estimated feature vector is output; the point-estimated feature vector is input into the uncertainty quantification module and the posterior predicted mean and variance of the feature vector are output; the feature vector is sampled in the feature space; the feature vector is input into the reconstruction module for temperature field reconstruction; finally, an end-to-end integrated model of real-time prediction and uncertainty quantification is formed.

[0015] Furthermore, in step 1, the geometric model includes a two-dimensional model and a three-dimensional model.

[0016] Furthermore, in step 2, the operating parameters include inflation rate, intake temperature, and initial pressure.

[0017] Furthermore, in step 3, the processing of multiple sets of high-fidelity temperature field data involves extracting the temperature values ​​on the grid nodes from the high-fidelity temperature field data, arranging them into a high-dimensional vector according to the node number, and standardizing or normalizing all temperature values ​​and operating parameters.

[0018] Furthermore, in step 4, the feature extractor is a variational autoencoder built based on a fully connected network.

[0019] Furthermore, the feature extractor includes an encoder and a decoder; the encoder is used for dimensionality reduction of the high-dimensional temperature field and outputs a feature vector; the decoder is used for reconstructing the temperature field from the feature vector.

[0020] Furthermore, in step 6, the uncertainty quantification module is implemented through a multi-task Gaussian process model. It takes the feature vector output by the prediction module in step 5 as input and outputs the mean and variance of the feature vectors in each dimension to characterize the prediction uncertainty in the feature space.

[0021] Furthermore, in step 7, the sampling of feature vectors in the feature space is performed using a joint sampling method.

[0022] The beneficial effects of this invention are as follows: When facing random hydrogen filling conditions, traditional numerical simulation methods often struggle to quickly obtain the temperature field response, resulting in a lack of real-time and reliable judgment for safety assessment and filling strategy optimization. This invention constructs a feature extractor to extract features from the high-dimensional temperature field, establishes a mapping between operating parameters and low-dimensional features through a prediction module, builds an uncertainty quantification module in the feature space, and reconstructs the temperature field using a reconstruction module. This enables real-time prediction and uncertainty quantification of the temperature field for fast-charging hydrogen storage cylinders under random operating conditions. Compared to traditional numerical simulation methods, this invention significantly reduces computational costs while maintaining computational accuracy; it can quantify the prediction uncertainty caused by data sparsity or random operating conditions, outputting a temperature field with confidence intervals, providing a more reliable decision-making basis for cylinder safety assessment; and it allows for rapid optimization of hydrogen filling strategies based on the prediction results, effectively ensuring cylinder safety. Furthermore, the constructed model has excellent data-driven update capabilities, can be continuously optimized with sample accumulation, and is adaptable to a wider range of engineering application scenarios. Attached Figure Description

[0023] Figure 1 This is a schematic diagram illustrating the implementation framework of the present invention;

[0024] Figure 2 This is a schematic diagram of the structure of a hydrogen storage cylinder in one embodiment;

[0025] Figure 3 This is a schematic diagram of the feature extractor in one implementation method;

[0026] Figure 4 This is a schematic diagram of the prediction module in one implementation method;

[0027] Figure 5 The following is a temperature field prediction result and error distribution diagram in one embodiment; wherein, (a) is the high-fidelity numerical simulation value, (b) is the mean value predicted by the integrated model, and (c) is the absolute error distribution diagram;

[0028] Figure 6 The following is a representation of the temperature rise prediction and uncertainty quantification results at different spatial locations in one embodiment; wherein, (a) is a schematic diagram of the prediction results for the bottle shoulder region, (b) is a schematic diagram of the prediction results for the cylinder region, and (c) is a schematic diagram of the prediction results for the bottle bottom region. Detailed Implementation

[0029] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0030] A real-time prediction method for the temperature field of a hydrogen storage cylinder during fast charging, such as... Figure 1 As shown, it includes the following steps:

[0031] Step 1: Establish a geometric model of the hydrogen storage cylinder using finite element software and discretize the model. A 150L Type III hydrogen storage cylinder is selected as the example for implementation. Figure 2 As shown, the gas cylinder structure consists of a 3mm thick aluminum alloy liner, an 8mm thick carbon fiber winding layer, and a 1mm thick glass fiber layer. The overall length of the gas cylinder is 1652mm, the inner diameter is 376mm, and the inlet diameter is 50mm. A two-dimensional axisymmetric model of the hydrogen storage cylinder is established, and the model is discretized using a hybrid structured and unstructured mesh: an unstructured mesh is used for the gas domain; a structured mesh is used for the solid domain; and the total number of meshes satisfies mesh convergence.

[0032] Step 2: The discretized model obtained in Step 1 is imported into the numerical simulation software to set up the numerical calculation model for fast charging of hydrogen storage cylinders. The charging rate is represented by the mass flow rate. By setting multiple sets of different operating parameters (mass flow rate, inlet temperature and initial pressure) as inlet boundary conditions and initial conditions, the numerical simulation of fast charging of hydrogen storage cylinders is carried out to obtain multiple sets of high-fidelity temperature field data.

[0033] Step 3: Process the sample data, that is, extract the temperature values ​​at the grid nodes in the high-fidelity temperature field data, arrange them into a high-dimensional vector according to the node number, and standardize the temperature values ​​and operating parameters of all samples. The temperature values ​​at each node are standardized independently, using the following formula:

[0034]

[0035] In the formula, T represents the standardized value of the i-th node; i This represents the temperature value of the i-th node. This represents the average temperature of the i-th node in the training set over the time scale. This represents the temperature standard deviation of the i-th node in the training set over the time scale.

[0036] Step 4: To address the high dimensionality and nonlinearity of the temperature field data, a variational autoencoder (VAE) based on a fully connected neural network is constructed as a feature extractor for dimensionality reduction and reconstruction. The encoder consists of alternating layers of linear layers and ReLU activation function layers, with the number of neurons decreasing progressively in a fixed ratio. The input layer dimension is the same as the number of nodes in the temperature field, and the final layer outputs a low-dimensional feature vector and a log-variance vector. The decoder adopts a structure symmetrical to the encoder, such as... Figure 3As shown; during training, a weighted sum of mean squared error and KL divergence is used as the loss function to constrain the distribution of the latent space. The temperature field dataset described in step 3 is used for unsupervised learning of the feature extractor; the trained feature extractor is used to extract low-dimensional feature vectors from all temperature field data, constructing an output feature dataset corresponding to the input feature dataset described in step 3; its decoder part then serves as... Figure 1 The refactoring module shown is used.

[0037] Step 5: For example Figure 1 The prediction module shown is constructed, specifically, as shown in the diagram. Figure 4 The multilayer fully connected neural network (FCNN) shown consists of one input layer, five hidden layers, and one output layer. Except for the output layer, which is linear, all other layers undergo nonlinear transformation using the ReLU function. The network accepts four-dimensional features as input, which are mapped to the low-dimensional feature vector of the temperature field described in step 4. During training, the mean squared error is used as the loss function, the input feature dataset described in step 3 is used as input, and the output feature dataset described in step 4 is used as output for supervised learning. After training, the FCNN network can quickly map any new operating condition parameters to the corresponding low-dimensional feature vector of the temperature field, providing point estimation input for the uncertainty quantification in step 6.

[0038] Step 6: For example Figure 1 The uncertainty quantification module shown is constructed using a multi-task Gaussian process regression model. This model takes the output of the FCNN network described in step 5 as input and outputs the posterior predicted mean and variance of the feature vector. Each feature dimension is treated as a Gaussian task, and the correlation between dimensions is captured by sharing a covariance module with radial basis function (RBF) as the kernel function. During training, the output feature vector of the FCNN network is used as input, and the output feature dataset described in step 4 is used as the label. The Adam optimizer is used to update the parameters based on minimizing the negative log-marginal likelihood function.

[0039] Step 7: Sample 100 feature vector samples from the posterior distribution of the feature vector output of the multi-task Gaussian process described in Step 6. That is, perform joint sampling on the multivariate Gaussian distribution output of the multi-task Gaussian process in the feature space. Based on the mean vector and covariance matrix of the distribution, generate multiple sets of relevant feature vector samples that conform to the distribution through methods such as Cholesky decomposition. Then, reconstruct the temperature field through the decoder, thereby propagating the uncertainty in the feature space to the temperature field space, and performing statistical calculations in the temperature field space to quantify the uncertainty.

[0040] Step 8: Combining steps 4 to 7, integrate the trained FCNN network, MTGP model, and VAE model decoder. The data flow is as follows: operating parameters are input into the FCNN network, which outputs point-estimated feature vectors; the point-estimated feature vectors are input into the MTGP model, which outputs the posterior predicted mean and variance of the feature vectors; feature vectors are sampled in the feature space; the feature vector samples are input into the VAE decoder to reconstruct the temperature field; finally, the following is formed: Figure 1 The end-to-end integrated model for real-time prediction and uncertainty quantification is shown.

[0041] Finally, in practical applications, the integrated model, by inputting the charging rate, inlet temperature, and initial state of the cylinder, can achieve real-time prediction of the temperature field during the fast charging process of hydrogen storage cylinders, and has high accuracy. Figure 5 As shown; simultaneously, the uncertainty of temperature rise prediction results at different locations in space can be quantified, such as... Figure 6 As shown, this can provide real-time decision-making support for the safety assessment and optimization of hydrogen charging strategies.

[0042] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, appropriate improvements and optimizations can be made to the implementation steps based on the basic principles of the present invention, but these improvements and optimizations should also be considered to be within the protection scope of the present invention.

Claims

1. A real-time prediction method of hydrogen storage cylinder fast charging temperature field, characterized in that, Includes the following steps: Step 1: Establish a geometric model of the hydrogen storage cylinder using finite element software and discretize the model to obtain a discretized model; Step 2: The discretized model obtained in Step 1 is imported into the numerical simulation software, and a numerical calculation model for fast charging of hydrogen storage cylinders is set. Different operating parameters are set as inlet boundary conditions and initial conditions, and multiple sets of high-fidelity temperature field data are obtained by solving. Step 3: Process multiple sets of high-fidelity temperature field data to construct the input feature dataset and the temperature field dataset; Step 4: To address the high-dimensionality and nonlinearity of the temperature field data, a feature extractor is constructed for dimensionality reduction and reconstruction. The feature extractor comprises an encoder and a decoder, built using an artificial neural network. The encoder consists of alternating layers of linear layers and ReLU activation function layers, with the number of neurons decreasing layer by layer at a fixed ratio, and the input layer dimension being the same as the number of nodes. The decoder adopts a structure symmetrical to the encoder. During training, the temperature field dataset described in Step 3 is used for unsupervised learning of the feature extractor. The trained feature extractor is used to extract low-dimensional feature vectors from all temperature field data, constructing an output feature dataset corresponding to the input feature dataset described in Step 3. The decoder with fixed parameters serves as the reconstruction module for reconstructing the temperature field from the low-dimensional feature vectors. Step 5: Construct a prediction module consisting of one input layer, multiple hidden layers, and one output layer. Except for the output layer, which uses a linear layer, the input and hidden layers undergo nonlinear transformation using the ReLU function. The prediction module accepts operating condition parameters as input and maps them to the low-dimensional feature vector described in Step 4. After training, the prediction module quickly maps any new operating condition parameters to the corresponding low-dimensional feature vector, providing point estimation input for the uncertainty quantification module in Step 6. Step 6: Construct an uncertainty quantification module, taking the output of the prediction module described in Step 5 as input, and output the posterior prediction mean and variance of the feature vector; Step 7: Sample m feature vector samples from the posterior predicted mean and variance of the feature vectors output by the uncertainty quantization module described in Step 6, and reconstruct m temperature fields through the decoder of the feature extractor described in Step 4. Propagate the uncertainty in the feature space to the temperature field space, and perform statistics on each position in the temperature field space to obtain the corresponding mean and variance. Step 8: Combining steps 4 to 7, integrate the trained prediction module, uncertainty quantification module, and reconstruction module. The data flow is as follows: the operating parameters are input into the prediction module and the point-estimated feature vector is output; the point-estimated feature vector is input into the uncertainty quantification module and the posterior predicted mean and variance of the feature vector are output; the feature vector is sampled in the feature space; the feature vector is input into the reconstruction module for temperature field reconstruction; finally, an end-to-end integrated model of real-time prediction and uncertainty quantification is formed.

2. The method of claim 1, wherein, In step 1, the geometric model includes a two-dimensional model and a three-dimensional model.

3. The method of claim 1, wherein, In step 2, the operating parameters include inflation rate, intake temperature, and initial pressure.

4. The method of claim 1, wherein, In step 3, the processing of multiple sets of high-fidelity temperature field data involves extracting the temperature values ​​on the grid nodes from the high-fidelity temperature field data, arranging them into a high-dimensional vector according to the node number, and standardizing or normalizing all temperature values ​​and operating parameters.

5. The method of claim 1, wherein, In step 4, the feature extractor is a variational autoencoder built on a fully connected network.

6. The real-time prediction method for the temperature field of a hydrogen storage cylinder during fast charging according to claim 1 or 5, characterized in that, The feature extractor includes an encoder and a decoder; the encoder is used for dimensionality reduction of the high-dimensional temperature field and outputs a feature vector; the decoder is used for reconstructing the temperature field from the feature vector.

7. The real-time prediction method for the temperature field of a hydrogen storage cylinder during fast charging according to claim 1, characterized in that, In step 6, the uncertainty quantification module is implemented through a multi-task Gaussian process model. It takes the feature vector output by the prediction module in step 5 as input and outputs the mean and variance of the feature vectors in each dimension to characterize the prediction uncertainty in the feature space.

8. The real-time prediction method for the temperature field of a hydrogen storage cylinder during fast charging according to claim 1, characterized in that, In step 7, the feature vectors are sampled in the feature space, that is, the joint sampling method is used for sampling.

Images