A data-driven multiphase flow prediction method for coal supercritical water hydrogen production reactor

Abstract

The present invention provides a data-driven method for predicting multiphase flow in a coal supercritical water hydrogen production reactor, comprising the following steps: constructing a mathematical simulation model of a coal supercritical water fluidized bed reactor, conducting simulation experiments based on a preset reactor wall temperature, generating a sample data set required for constructing a deep spatiotemporal sequence model based on the experimental results, and preprocessing the sample data; constructing and training a deep spatiotemporal sequence model based on the sample data set, wherein the deep spatiotemporal sequence model adopts a U-Net structure in an encoding-decoding form, coupled with a convolution module and a bidirectional convolutional long short-term memory network module, for extracting the multiscale spatiotemporal evolution characteristics of the multiphase flow field in the reactor; and using the trained deep spatiotemporal sequence model for rapid prediction of the multiphase flow field in the reactor. The present invention can accurately and efficiently predict the multiphase flow field in the reactor, which is beneficial to improving the efficiency and quality of downstream tasks such as reactor design and optimization.

Description

Technical Field

[0001] The present invention relates to the field of coal supercritical water hydrogen production reactor design, and in particular to a data-driven coal supercritical water hydrogen production reactor multiphase flow prediction method. Background Art

[0002] Supercritical water fluidized bed (SCWFB) reactors enable clean and efficient coal conversion and hydrogen generation. However, studying the complex multiphase flow field and heat and mass transfer within the reactor relies on expensive and time-consuming experiments or transient numerical simulations. Traditional numerical simulations are slow and require expert prior knowledge, making them incapable of real-time prediction of the multiphase flow field within the reactor and unable to effectively support the efficient execution of downstream tasks. Summary of the Invention

[0003] In light of the shortcomings of existing technologies, the present invention provides a data-driven method for predicting multiphase flow in a coal supercritical water hydrogen production reactor. This method employs an encoding-decoding U-Net architecture, coupled with convolutional and BiConvLSTM modules for layer-by-layer mirroring. This method extracts the multiscale spatiotemporal coupled evolutionary characteristics of the multiphase flow field within the reactor, enabling rapid and accurate prediction of the multiphase flow field within the reactor, thereby effectively improving the efficiency and quality of downstream tasks.

[0004] The technical means adopted in the present invention are as follows:

[0005] A data-driven multiphase flow prediction method for a coal supercritical water hydrogen production reactor includes the following steps:

[0006] Construct a mathematical simulation model of a coal supercritical water fluidized bed reactor, conduct simulation experiments based on the preset reactor wall temperature, generate the sample data set required to construct the deep spatiotemporal sequence model based on the experimental results, and preprocess the sample data;

[0007] A deep spatiotemporal sequence model is constructed and trained based on the sample dataset. The deep spatiotemporal sequence model adopts an encoding-decoding U-Net structure, coupled with a convolutional module and a bidirectional convolutional long short-term memory network module to extract the multi-scale spatiotemporal evolution characteristics of the multiphase flow field in the reactor;

[0008] The trained deep spatiotemporal sequence model is used to quickly predict the multiphase flow field in the reactor.

[0009] Furthermore, the U-Net structure includes an encoder and a decoder; the encoder includes a three-layer combination module consisting of a convolution module and a bidirectional convolutional long short-term memory network module; the decoder is mirror-symmetrical to the encoder and also includes a three-layer combination module consisting of a deconvolution module and a bidirectional convolutional long short-term memory network module;

[0010] The hidden state of the final output of the bidirectional convolutional long short-term memory network module in the encoder is passed as the initial state to the corresponding bidirectional convolutional long short-term memory network module in the decoder through skip connections. The ReLU function and Softmax function are added to the output layer to ensure that the model prediction results meet the physical constraints of non-negativity and the sum of the mole fractions of each gasification product is 1.

[0011] Furthermore, the training of the deep spatiotemporal sequence model specifically includes:

[0012] The spatiotemporal prediction of the flow field in the reactor is defined as: in, The input flow field sequence contains W time-step historical flow field snapshots; is the flow field sequence predicted by the model, which contains H time-step future flow field snapshots; is the constructed deep spatiotemporal sequence model;

[0013] The Kaiming initialization strategy is used to initialize the weights and bias parameters in the model. The root mean square error between the true value and the predicted value of the flow field is used as the loss function. The Adam optimizer is used to iteratively optimize the parameters of the deep spatiotemporal sequence model. The minimum loss function is used as the optimization goal until the loss function on the validation set no longer decreases and the training is completed.

[0014] Furthermore, a mathematical simulation model of a coal supercritical water fluidized bed reactor was constructed. Simulation experiments were conducted based on the preset reactor wall temperature. Based on the experimental results, a sample data set required for constructing a deep spatiotemporal sequence model was generated. Specifically, the following were included:

[0015] Construct a mathematical simulation model of a coal supercritical water fluidized bed reactor. Couple the flow field control equations within the reactor with the coal supercritical water gasification reaction mechanism. Generate a computational grid for the reactor's central cross-section. Build a two-dimensional transient CFD model of the reactor. Perform simulation experiments based on preset reactor wall temperature conditions.

[0016] The temporal and spatial evolution data of the transient flow field under various temperature conditions in the reactor are extracted as the sample data set of the reactor flow field.

[0017] Furthermore, a simulation experiment is conducted based on the preset reactor wall temperature conditions, specifically including:

[0018] Extracting transient flow field spatiotemporal evolution data at regular grid nodes at the center cross section of the reactor, wherein the transient flow field spatiotemporal evolution data includes fluid temperature, fluid velocity, and mole fractions of various gasification products;

[0019] For different reactor wall temperature conditions, multiple sets of temporal and spatial evolution data of transient flow fields in the reactor are obtained through simulation;

[0020] The transient flow field spatiotemporal evolution data is stored in the form of high-dimensional tensors for training and testing deep spatiotemporal sequence models.

[0021] Furthermore, the sample data is preprocessed, specifically including:

[0022] The obtained transient flow field spatiotemporal evolution data are normalized, and the sample data set is divided and short sequences are generated for training and testing of deep spatiotemporal sequence models.

[0023] Furthermore, the obtained transient flow field spatiotemporal evolution data is normalized, and the sample data set is divided and short sequences are generated, specifically including:

[0024] The maximum and minimum normalization method is used to normalize the flow field data to the [0,1] unit space to achieve dimensionless data.

[0025] Multiple sets of reactor transient flow field spatiotemporal evolution data are divided into training set, validation set and test set according to temperature conditions and reaction time;

[0026] A sliding window with a step size of 1 is used to divide the long sequence tensor in the dataset into multiple short sequences, and the spatiotemporal evolution data of the reactor transient flow field is enhanced to improve the training quality of the deep spatiotemporal sequence model.

[0027] Furthermore, the trained deep spatiotemporal sequence model is used to quickly predict the multiphase flow field in the reactor, including:

[0028] The trained deep spatiotemporal sequence model is loaded to predict the spatiotemporal evolution of the flow field in the reactor under unknown wall temperature conditions. A supervised learning method is used to perform full-cycle prediction of the unknown flow field to test the generalization ability of the model.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] 1. This invention can simultaneously and effectively learn the spatiotemporal evolution of the multiphase flow field in a reactor under different temperature conditions, thereby achieving accurate forward multi-time-step predictions. The flow field predicted by the deep spatiotemporal sequence model is not only highly consistent with the simulation results, but also hundreds of times faster than traditional numerical simulations. This enables real-time prediction of the multiphase flow field in the reactor, effectively improving the efficiency and quality of downstream tasks such as reactor optimization and regulation. This is a feasible new method with broad application prospects.

[0031] 2. This paper uses an encoding-decoding U-Net model structure, coupled with a convolutional module and a bidirectional convolutional long-short-term memory network module to construct a deep spatiotemporal sequence model for intelligent analysis and prediction of the multiphase flow field within a coal supercritical water hydrogen production reactor. This alleviates the problem of relying on expensive and time-consuming experiments or transient numerical simulations to study the complex multiphase flow field and heat and mass transfer laws within the reactor. It achieves rapid and accurate prediction of the multiphase flow field within the reactor, effectively improving the efficiency and quality of reactor design and optimization tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0033] Figure 1 Flowchart of the data-driven multiphase flow prediction method for coal supercritical water hydrogen production reactor.

[0034] Figure 2a Schematic diagram of the geometric model of the coal supercritical water hydrogen production reactor.

[0035] Figure 2b Schematic diagram of the two-dimensional structured grid of the geometric model of the coal supercritical water hydrogen production reactor.

[0036] Figure 3 Schematic diagram of the deep spatiotemporal sequence model framework.

[0037] Figure 4a The model in the embodiment is the reactor wall temperature T w =843K, time t=60s.

[0038] Figure 4b The model in the embodiment is the reactor wall temperature T w =903K, time t=60s. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0040] like Figure 1 As shown, the present invention discloses a data-driven intelligent prediction method for multiphase flow depth in a coal supercritical water hydrogen production reactor, which mainly includes:

[0041] Step 1: Generate the sample dataset required to build the deep spatiotemporal sequence model and perform preprocessing.

[0042] Specifically include:

[0043] Step 1.1: Construct a mathematical simulation model of the coal supercritical water fluidized bed reactor, coupling the flow field control equations within the reactor and the complex coal supercritical water gasification reaction mechanism; generate a computational grid for the central cross-section of the reactor, establish a two-dimensional transient CFD model of the reactor, and set a series of reactor wall temperature conditions for simulation experiments.

[0044] Specifically, multiphase flow field data are extracted at the regular grid nodes at the central cross-section of the reactor, including transient flow field spatiotemporal evolution data such as fluid temperature, fluid velocity, and mole fraction of each gasification product. For different reactor wall temperature conditions, multiple sets of transient flow field spatiotemporal evolution data sets in the reactor are simulated and stored in the form of high-dimensional tensors for the training and testing of deep spatiotemporal sequence models.

[0045] Step 1.2: Extract the transient spatiotemporal evolution data of the flow field under various temperature conditions within the reactor as a sample dataset of the reactor flow field. Furthermore, this step also includes normalizing the obtained transient spatiotemporal evolution data, partitioning the sample dataset, and generating short sequences for training and testing the deep spatiotemporal sequence model.

[0046] Specifically, first, the maximum and minimum normalization method is used to normalize the flow field data to the [0,1] unit space to achieve dimensionless data; then, multiple sets of reactor transient flow field spatiotemporal evolution data are divided into training set, validation set and test set according to temperature conditions and reaction time; finally, a sliding window with a step size of 1 is used to divide the long sequence tensor in the dataset into multiple short sequences, and the data is enhanced to improve the training quality of the deep spatiotemporal sequence model.

[0047] This embodiment uses Figure 2a The flow field at the center cross section of the coal supercritical water hydrogen production reactor shown in the figure is taken as the research object. Considering that the reactor wall temperature has a great influence on the flow morphology and spatial distribution of the multiphase transient flow field in the reactor. Therefore, in the applicable temperature range T of the coal supercritical water gasification reaction mechanism model w A set of simulation experiments is set up every ΔT=10K within ∈[823K,923K], and the simulation time is t=70s.

[0048] First, the ICEM CFD software was used to generate a two-dimensional structured grid at the reactor cross section with a total number of 32 × 512 grid nodes for flow field transient simulation calculations, such as Figure 2b shown.

[0049] Secondly, a numerical simulation model of the reactor was established based on the Euler-Lagrangian method and coupled with the coal supercritical water gasification kinetic model. The RNG k-ε turbulence model was used to solve the continuous phase flow field, the discrete coordinate method (DO) was used to calculate the thermal radiation, and the finite rate model was used to obtain the chemical reaction rate. The flow control equations in the reactor were discretized using the first-order upwind scheme and solved using the Phase Coupled SIMPLE algorithm in the commercial software Fluent solver. The time step of the transient simulation was set to 0.005s, and the flow field calculation results were saved every 8 iterative time steps (sampling time interval Δt = 0.04s). The transient simulation time was set to t = 70s. The residuals of the energy and radiation equations are required to be less than 10 -6 , the residual of the energy equation should be less than 10 -3 .

[0050] Next, flow field data were extracted from a regular grid of nodes in the reactor's central cross-section. This data includes transient spatiotemporal evolution data of the flow field, including fluid temperature, velocity, and mole fractions of various gasification products (H₂, CO₂, CO, CH₄, and Vol). For different reactor wall temperature conditions, 11 sets of transient spatiotemporal evolution data within the reactor were generated. Each set of transient spatiotemporal evolution data was processed into a high-dimensional tensor of 1750 × 7 × 32 × 512 dimensions for training and testing the deep spatiotemporal sequence model. Table 1 describes the 11 generated reactor flow field datasets under different temperature conditions, which are divided into training, validation, and test sets.

[0051] Table 1 11 groups of reactor simulation data sets and data set divisions

[0052]

[0053] Finally, after simulating the flow field data, the maximum and minimum normalization method is used to make the original flow field data dimensionless. Then, a sliding window with a step size of 1 is used to divide the long sequence tensor in the dataset into multiple short sequences. The data is enhanced to improve the training quality of the deep spatiotemporal sequence model.

[0054] Step 2: Build and train a deep spatiotemporal sequence model.

[0055] Specifically, in this embodiment, the entire deep spatiotemporal sequence model adopts a U-Net structure in the form of encoding and decoding, coupling a convolutional module (Conv) and a bidirectional convolutional long short-term memory network module (BiConvLSTM). The encoder includes a three-layer Conv+BiConvLSTM combination module to achieve layer-by-layer downsampling of the flow field, and the number of feature channels gradually increases from 7 to 128 to better extract the spatiotemporal coupling characteristics of the flow field; the decoder and the encoder are mirror-symmetrical and also include a three-layer DeConv+BiConvLSTM combination module, in which the deconvolution layer DeConv restores the high-resolution flow field through an upsampling operation and predicts the flow field evolution at future moments through the corresponding BiConvLSTM layer; in addition, the hidden state (H) of the final output of the BiConvLSTM in the encoder is connected through a jump connection. t ,C t ) is passed as the initial state to the corresponding BiConvLSTM layer in the decoder; finally, the ReLU function and Softmax function are added to the output layer to ensure that the model prediction results meet the physical constraints of non-negativity and the sum of the mole fractions of each gasification product is 1. The model structure diagram is shown in Figure 3 shown.

[0056] The training process of the deep spatiotemporal sequence model is mainly based on the Pytorch deep learning framework for model construction and training. First, the kaiming initialization strategy is used to initialize the weights and bias parameters in the model. Then, the root mean square error (RMSE) between the true value and the predicted value of the flow field is used as the loss function, and the Adam optimizer is used to iteratively optimize the parameters of the deep spatiotemporal sequence model. The minimum loss function is used as the optimization goal, and the initial learning rate is set to 1×10 -4 , until the loss function on the validation set no longer decreases and the training is completed.

[0057] Step 3: Rapid prediction of the multiphase flow field in the reactor, specifically including loading the trained deep spatiotemporal sequence model to predict the spatiotemporal evolution of the flow field in the reactor under unknown wall temperature conditions, and using supervised learning to perform full-cycle prediction of the unknown flow field to test the generalization ability of the model.

[0058] Specifically, in this embodiment, the trained deep spatiotemporal sequence model is loaded to predict the unknown temperature T w The spatiotemporal evolution of the multiphase flow field in the reactor under the following conditions is investigated. The following test scenarios are set to test the predictive ability of the constructed model: w =843K and T w =903K, the unknown flow field with reaction time t∈[6s,60s] is predicted in full cycle using supervised learning. The model prediction results are shown in Figure 4. Figure 4a is the reactor wall temperature T w =843K, time t=60s flow field prediction results and comparison with CFD simulation results; Figure 4b is the reactor wall temperature T w =903K, time t = 60s, and comparison with CFD simulation results. These test results demonstrate that the constructed deep spatiotemporal sequence model can simultaneously learn the spatiotemporal evolution of the multiphase flow field within the reactor under different temperature conditions, thereby achieving accurate multi-step predictions and excellent generalization capabilities. The model-predicted flow field is not only highly consistent with the simulation results, but also hundreds of times faster than traditional numerical simulation methods.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven multiphase flow prediction method for a coal supercritical water hydrogen production reactor, characterized in that: The following steps are involved: Constructing a mathematical simulation model of a coal supercritical water fluidized bed reactor, conducting simulation experiments based on a preset reactor wall temperature, generating a sample data set required for constructing a deep space-time sequence model based on the experimental results, and preprocessing the sample data; wherein, constructing a mathematical simulation model of a coal supercritical water fluidized bed reactor, conducting simulation experiments based on a preset reactor wall temperature, and generating a sample data set required for constructing a deep space-time sequence model based on the experimental results, specifically includes: constructing a mathematical simulation model of a coal supercritical water fluidized bed reactor, coupling the flow field control equation in the reactor and the coal supercritical water gasification reaction mechanism, generating a computational grid for the central cross-section of the reactor, establishing a two-dimensional transient CFD model of the reactor, conducting simulation experiments based on a preset reactor wall temperature condition, and extracting the transient flow field spatiotemporal evolution data under various temperature conditions in the reactor as a sample data set of the reactor flow field; A deep spatiotemporal sequence model is constructed and trained based on a sample data set. The deep spatiotemporal sequence model adopts a U-Net structure in the form of encoding and decoding, coupled with a convolution module and a bidirectional convolutional long short-term memory network module, to extract the multi-scale spatiotemporal evolution characteristics of the multiphase flow field in the reactor. The training of the deep spatiotemporal sequence model specifically includes: defining the spatiotemporal prediction of the flow field in the reactor as: in, The input flow field sequence contains W time-step historical flow field snapshots; is the flow field sequence predicted by the model, which contains H time-step future flow field snapshots; The constructed deep spatiotemporal sequence model is initialized using the Kaiming initialization strategy to initialize the weights and bias parameters in the model. The root mean square error between the true value and the predicted value of the flow field is used as the loss function. The Adam optimizer is used to iteratively optimize the parameters of the deep spatiotemporal sequence model, with the minimum loss function as the optimization goal. The training is completed when the loss function on the validation set no longer decreases. The trained deep spatiotemporal sequence model is used to quickly predict the multiphase flow field in the reactor.

2. A data-driven coal supercritical water hydrogen production reactor multiphase flow prediction method according to claim 1, characterized in that: The U-Net structure includes an encoder and a decoder; the encoder includes a three-layer combination module consisting of a convolution module and a bidirectional convolutional long short-term memory network module; the decoder is mirror-symmetrical to the encoder and also includes a three-layer combination module consisting of a deconvolution module and a bidirectional convolutional long short-term memory network module; The hidden state of the final output of the bidirectional convolutional long short-term memory network module in the encoder is passed as the initial state to the corresponding bidirectional convolutional long short-term memory network module in the decoder through skip connections. The ReLU function and Softmax function are added to the output layer to ensure that the model prediction results meet the physical constraints of non-negativity and the sum of the mole fractions of each gasification product is 1.

3. The data-driven multiphase flow prediction method for a coal supercritical water hydrogen production reactor according to claim 1 is characterized in that: The simulation experiment is carried out based on the preset reactor wall temperature conditions, including: Extracting transient flow field spatiotemporal evolution data at regular grid nodes at the center cross section of the reactor, wherein the transient flow field spatiotemporal evolution data includes fluid temperature, fluid velocity, and mole fractions of various gasification products; For different reactor wall temperature conditions, multiple sets of temporal and spatial evolution data of transient flow fields in the reactor are obtained through simulation; The transient flow field spatiotemporal evolution data is stored in the form of high-dimensional tensors for training and testing deep spatiotemporal sequence models.

4. A data-driven multiphase flow prediction method for a coal supercritical water hydrogen production reactor according to claim 1, characterized in that: Preprocess the sample data, including: The obtained transient flow field spatiotemporal evolution data are normalized, and the sample data set is divided and short sequences are generated for training and testing of deep spatiotemporal sequence models.

5. A data-driven multiphase flow prediction method for a coal supercritical water hydrogen production reactor according to claim 4, characterized in that: The obtained transient flow field spatiotemporal evolution data are normalized, and the sample data set is divided and short sequences are generated, including: The maximum and minimum normalization method is used to normalize the flow field data to the [0,1] unit space to achieve dimensionless data. Multiple sets of reactor transient flow field spatiotemporal evolution data are divided into training set, validation set and test set according to temperature conditions and reaction time; A sliding window with a step size of 1 is used to divide the long sequence tensor in the dataset into multiple short sequences, and the spatiotemporal evolution data of the reactor transient flow field is enhanced to improve the training quality of the deep spatiotemporal sequence model.

6. A data-driven multiphase flow prediction method for a coal supercritical water hydrogen production reactor according to claim 1, characterized in that: The trained deep spatiotemporal sequence model is used for rapid prediction of the multiphase flow field in the reactor, including: The trained deep spatiotemporal sequence model is loaded to predict the spatiotemporal evolution of the flow field in the reactor under unknown wall temperature conditions. A supervised learning method is used to perform full-cycle prediction of the unknown flow field to test the generalization ability of the model.

Images