Gas explosion venting overpressure partition calculation method combined with physical information neural network

Abstract

The invention discloses a gas explosion venting overpressure zoning calculation method combined with a physical information neural network, which comprises the following steps of: constructing a three-zone-one-state balance state control equation through a conservation equation, an explosion mechanics theory and a zoning calculation thought, and dividing required input parameters into static characteristics and dynamic characteristics as input characteristics of a TFT (Thin Film Transistor) algorithm, a theoretical model is introduced into a loss function to construct constraint terms of explosion physical information, a PINN-TFT algorithm coupled with the physical information is constructed, a data set is constructed through gas explosion venting test data, PINN-TFT algorithm training is carried out, and a gas explosion venting overpressure partition calculation method combined with a physical information neural network is formed. The method has excellent interpretability and generalization performance, and can be used for explosion venting overpressure prediction of any gas and any ignition position in a limited space.

Description

Technical Field

[0001] This invention belongs to the field of process safety and relates to the theory of hazardous material explosion prediction, and in particular to a method for calculating gas explosion overpressure zoning by combining physical information neural networks. Background Technology

[0002] Combustible gases such as methane, ethane, hydrogen, ammonia, and natural gas are increasingly used in industrial production and daily life due to their wide availability, low pollution levels, and ease of use. However, this also increases the risk of explosions. Explosion venting, as an economical and effective explosion protection measure, has been widely applied in the design of various buildings and structures.

[0003] Currently, methods for predicting explosion overpressure are mainly divided into two categories: numerical simulation prediction and formula prediction. However, existing empirical / semi-empirical / simplified calculation formulas generally have limitations in applicability; while numerical simulation methods require high computational resources and have long cycles, making it difficult to meet the needs of rapid evaluation in engineering design. Therefore, there is an urgent need to develop simplified prediction models with high computational efficiency and reliable accuracy.

[0004] Sun Song et al. (Sun Song, Gao Kanghua, Qiu Yanyu, et al. Stepwise calculation model of gas explosion pressure and its turbulence correction [J]. Explosion and Shock, 2019, 39(05): 92-100.) decomposed the gas explosion process in a confined space into three independent stages: combustion, release, and pressure balance for stepwise calculation. This model sets the maximum explosion overpressure as a constant and simplifies the flame propagation velocity to a corrected value of the laminar flame velocity. The macroscopic flame area is approximated as the spherical part that does not contact the wall. This model has significant limitations: the constant maximum explosion overpressure does not match the dynamic release process, and the simplified flame velocity and area model fails to fully consider the dynamic deformation characteristics and nonlinear acceleration mechanism during the flame propagation process, thus limiting the prediction accuracy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention constructs theoretical models of the burned zone, unburned zone, and venting zone based on existing theoretical research, thereby forming a calculation model for explosion overpressure zoning. The input parameters are divided into static and dynamic features, which serve as input features for the Temporal Fusion Transformer (TFT) algorithm. A constraint term is introduced into the loss function to construct a PINN-TFT (Physics-informed neural networks-Temporal Fusion Transformer) algorithm model that couples physical information. The method of this invention is interpretable and possesses excellent prediction performance and generalization ability, enabling the prediction of the entire time-series evolution of explosion overpressure, thus overcoming the deficiencies of the prior art.

[0006] The present invention is achieved by at least one of the following technical solutions.

[0007] A gas explosion overpressure zoning calculation method combining physical information neural networks includes the following steps: inputting the working condition parameters of the structure explosion scenario into the trained PINN-TFT algorithm model, determining whether an explosion has occurred, and if the predicted explosion overpressure is greater than the static explosion overpressure, then an explosion has occurred.

[0008] Furthermore, the PINN-TFT algorithm model includes a venting overpressure partitioning calculation model and a TFT algorithm. The input parameters are divided into static features and dynamic features, which are used as input features for the TFT algorithm. Constraint terms are introduced into the loss function to construct a PINN-TFT algorithm model that couples physical information.

[0009] Furthermore, static characteristics include the length, width, and height of the explosion space, as well as the coordinates of the explosion point, the adiabatic explosion overpressure, the laminar flame velocity, the gas adiabatic index, the density of unburned gas, the density of burned gas, the area of ​​the explosion vent, and the throttling coefficient of the explosion vent.

[0010] Dynamic characteristics include ignition time and explosion overpressure rise rate.

[0011] Furthermore, the explosion overpressure zoning calculation model includes a mathematical model for calculating the parameters of the burned zone, the unburned zone, and the venting zone.

[0012] Furthermore, the mathematical model of the already burned zone solves for the following parameters: overpressure increment of the already burned zone, density of the already burned zone, volume of the already burned zone, and flame propagation distance.

[0013] Furthermore, the mathematical model of the unburned zone solves for the following parameters: unburned zone overpressure, unburned zone volume, unburned zone mass, and unburned zone density.

[0014] Furthermore, the mathematical model of the venting zone solves for the following parameters: overpressure at the vent, flow velocity at the vent, and flow rate at the vent; the equilibrium pressure within the confined space after venting is calculated using the following mathematical model of equilibrium state:

[0015]

[0016] Where, m bi Let ρ be the volume of the combusted gas at time i. bi P is the density of the burned zone. bi For the overpressure in the combustible zone at time i, P i-net γ is the predicted value of the explosion overpressure. b Indicates the adiabatic index of unburned gas, m ui ρ is the total mass of unburned gas. ui Let P be the density of the unburned gas at time i. ui For the overpressure in the unburned zone at time i, γ uV represents the adiabatic index of the unburned gas, and V is the volume of the confined space.

[0017] Furthermore, the training of the PINN-TFT algorithm model includes the following steps:

[0018] 1. Input a set of data from the dataset and the parameters of the burned zone, unburned zone, venting zone and equilibrium state output by the explosion overpressure zoning calculation model into the TFT algorithm, and output the turbulent flame velocity, unburned gas velocity and macroscopic flame area of ​​the next time step;

[0019] 2. Substitute the predicted turbulent flame velocity, unburned gas velocity and macroscopic flame area from step 1 into the explosion venting overpressure zoning calculation model, and calculate the parameters of the burned zone, unburned zone, venting zone and equilibrium state respectively.

[0020] 3. The predicted overpressure value P for the explosion relief i-net The explosion overpressure P obtained from the experiment i-exp Compare, through The loss value is calculated and used for backward gradient propagation in the TFT algorithm;

[0021] 4. Based on the predicted overpressure value P of the explosion relief i-net and static explosion overpressure P stat The relative magnitude is used to determine the current predicted value P of the explosion overpressure. i-net Check if the static explosion overpressure has been reached. If not, repeat steps 1 to 3 above.

[0022] 5. Train on the next set of data, repeating steps 1 to 4 above, until iterative training on all data is completed.

[0023] A computer device according to the present invention includes a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, which, when executed by the processor, causes the processor to implement the method described herein.

[0024] The present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor implements the method described herein.

[0025] Compared with existing technologies, the beneficial effects of the present invention are as follows:

[0026] 1. This invention proposes a dynamic calculation method that calculates the adiabatic explosion overpressure in real time based on the state of the previous time step. By integrating physical information and neural network technology, it can dynamically solve key parameters (flame area, turbulent flame velocity, and unburned gas velocity), thereby significantly improving the accuracy and reliability of explosion overpressure prediction. Starting from theoretical equations, the control equations are constructed based on the gas explosion propagation theory.

[0027] 2. This invention takes into account the difference in the impact of static features and dynamic features on the results through the TFT algorithm.

[0028] 3. This invention incorporates physical information into the loss function of the TFT algorithm, ensuring the accuracy, generalizability, and interpretability of the calculation results. Attached Figure Description

[0029] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0030] Figure 1 This is a schematic diagram illustrating the partitioning calculation principle of an embodiment;

[0031] Figure 2 This is a flowchart of an embodiment of a method for calculating explosion overpressure based on a physical information neural network;

[0032] Figure 3 This is a diagram showing the results of simulating explosion overpressure using the method of the present invention in an embodiment. Detailed Implementation

[0033] Exemplary embodiments will now be described in detail. When the following description relates to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of scenarios and methods consistent with some aspects of the invention as detailed in the appended claims.

[0034] This invention discloses a gas explosion overpressure zoning calculation method combining physical information neural networks. It constructs a gas explosion overpressure prediction model (PINN-TFT algorithm model) by integrating physical information with the Temporal Fusion Transformer (TFT) algorithm. Specifically, the PINN-TFT algorithm model includes an explosion overpressure zoning calculation model and a TFT algorithm. The explosion overpressure zoning calculation model serves as the physical information component, while the TFT algorithm serves as the intelligent regression algorithm component. The model's process includes dividing the input parameters into static and dynamic features, which are used as input features for the TFT algorithm. A constraint term based on explosion physical information is introduced into the loss function to construct the PINN-TFT algorithm coupled with physical information. A dataset is constructed using gas explosion test data, and the PINN-TFT algorithm is trained to form a gas explosion overpressure zoning calculation method combining physical information neural networks. The explosion overpressure zoning calculation model is a three-zone (burned zone, unburned zone, and venting zone) state (equilibrium state) control equation constructed using conservation equations, explosion mechanics theory, and the concept of zoning calculation.

[0035] This embodiment of the system for implementing the gas explosion venting overpressure zoning calculation method combining physical information neural networks includes the following three parts. The first part is to construct a mathematical model, specifically, to establish an explosion venting overpressure zoning calculation model according to the zoning calculation idea, such as... Figure 1 As shown; the second part is data acquisition and processing: the acquired experimental data is processed into the parameters required by the model and feature processing is performed; the third part is the construction of the PINN-TFT algorithm. Based on the feature type, the PINN-TFT algorithm is constructed with static features, time-varying features (time-varying features observed in the past and time-varying features known in the prior a priori) as inputs and multi-objective outputs. The loss value is calculated each time based on the predicted explosion overpressure value and the explosion overpressure obtained from the experiment through the loss function.

[0036] The following section selects a hydrogen pipeline explosion scenario and, using the obtained experimental data, presents the specific steps of a gas explosion overpressure zoning calculation method that combines physical information neural networks in this embodiment.

[0037] like Figure 2 As shown, a method for calculating gas explosion overpressure zones using a physical information neural network includes the following steps:

[0038] Input the working condition parameters under the structure explosion scenario into the trained gas explosion overpressure prediction model (PINN-TFT algorithm model) as initialization parameters, and input the time step as the step size of each iteration step; the working condition parameters include the length, width, and height of the confined space, as well as the laminar combustion velocity of the premixed combustible gas, the ambient pressure, the temperature of the premixed combustible gas, the adiabatic index and density of the unburned gas, and the adiabatic index of the burned gas, etc.

[0039] The trained gas explosion overpressure prediction model (PINN-TFT algorithm model) determines whether an explosion has occurred. If no explosion has occurred, the released gas flow rate is 0, and only the parameters of the combusted and uncombustible zones are calculated. If an explosion has occurred, the calculation method for the parameters of the combusted and uncombustible zones is the same as under the no-explosion condition. In this case, the parameters of the venting zone need to be calculated based on the released gas flow rate, and then the current explosion overpressure prediction value P is output. i-net ;

[0040] Whether a deflation has occurred is determined by the explosion pressure within the confined space, i.e., the predicted overpressure value P. i-net With static explosion overpressure P stat The relative magnitude is determined. If the explosion pressure within the confined space exceeds the static explosion relief overpressure (i.e., the explosion relief design pressure), the explosion relief plate fails, and an explosion occurs.

[0041] Furthermore, if the flame propagation distance equals the length of the confined space, the reaction ends; otherwise, the equilibrium parameters (pressure P) will be affected. i Density of combusted gas ρcbi and the density of unburned gas ρ cui The condition parameter is used for calculation in step t+1, and the calculation stops when the reaction ends.

[0042] The training of the gas explosion overpressure prediction model (PINN-TFT algorithm model) in this embodiment includes the following steps:

[0043] S1. Conduct hydrogen explosion tests under the pressure relief pipeline to obtain data on the change of explosion overpressure over time. As an example, the sampling frequency is 100kHz. The dimensions of the test pipeline, the size of the explosion relief port, and the explosion pressure are varied, and multiple sets (50 sets) of tests are conducted for each variable to obtain parameters including the length, width, and height of the explosion space, the coordinates of the explosion point, the adiabatic explosion overpressure, the laminar flame velocity, the adiabatic index of the unburned gas, the density of the unburned gas, the adiabatic index of the burned gas, the density of the burned gas, the area of ​​the explosion relief port, the throttling coefficient of the explosion relief port, the time after ignition, the rate of explosion overpressure rise, and the change of explosion overpressure over time (rate of explosion overpressure rise).

[0044] S2. Process the data to form a dataset. The explosion space length, width, and height, explosion point coordinates, adiabatic explosion overpressure, laminar flame velocity, gas adiabatic index, unburned gas density, burned gas density, vent area, and vent throttling coefficient are static variables. Ignition time and explosion overpressure rise rate are time-varying variables. Explosion overpressure is the label. The static variables need feature processing to expand them to match the dimensions of the time-varying variables. The explosion overpressure rise rate among the time-varying variables needs feature processing, shifting all overpressures to the next time step, initially filling the overpressure with 0.

[0045] S3. Discretize the computational domain into three parts: the burned zone, the unburned zone, and the venting zone. Specifically, the confined space is divided into the burned zone and the unburned zone by the flame front, and the area outside the confined space is the venting zone. As an example, it is assumed that the physical parameters within the same region are consistent, that the gas volume change caused by the pressure change is isentropic compression, and that the unburned and burned gases satisfy the multi-equation of state for gases. A mathematical model is then constructed to calculate the parameters of the burned zone, the unburned zone, and the venting zone respectively.

[0046] The parameters of the burned zone are calculated using the following mathematical model:

[0047] Overpressure increment in the already burned zone:

[0048] Quality of the burned area:

[0049] Density of the burned zone:

[0050] Volume of the burned area:

[0051] Flame propagation distance: Δx i =(S ti +S ui )·Δt

[0052] Where γ u Indicates the adiabatic index of unburned gas, γ b Indicates the adiabatic index of unburned gas, S t A represents the propagation speed of turbulent flame. l p represents the macroscopic flame area. bi Let be the adiabatic explosion overpressure at time i (explosion pressure within the confined space), Θ be the premixed gas reaction expansion rate, V be the volume of the confined space (V=L×D×H), where L, D, and H are the length, width, and height of the container, respectively, p0 be the atmospheric pressure, and A be the pressure at time i. li Let S be the macroscopic flame area at time i. ti Let m be the turbulent flame velocity at time i. bi Let S be the volume of the combusted gas at time i. ui Let ρ be the velocity of the unburned gas at time i. u(i-1) Let ρ be the density of the unburned gas at time i-1. b0 V is the density of the combusted gas at atmospheric pressure. bi Let x be the volume of the combusted gas at time i. i ρ is the flame propagation distance at time i, Δt is the time step size; bi The density of the burned zone, Δx i This represents the distance the flame travels.

[0053] The parameters of the unburned zone are calculated using the following mathematical model:

[0054] Overpressure in unburned zone: P ui =P bi

[0055] Unburned zone volume: V ui =VV bi

[0056] Unburned zone mass: m ui =m0-m ei -m bi

[0057] Density of unburned zone:

[0058] Where P ui For the overpressure in the unburned zone at time i, P bi For the overpressure in the combustible zone at time i, V ui Let m be the volume of the unburned zone at time i, m0 be the total gas mass at the initial time, and m ei For the total mass of the gas released, m bi m is the total mass of the combusted gas.ui ρ is the total mass of unburned gas. ui Let be the density of unburned gas at time i.

[0059] The parameters of the vent area are calculated using the following mathematical model:

[0060] If the explosion overpressure P obtained in the experiment i-exp Greater than static explosion overpressure P stat If this happens, an explosion will occur.

[0061] Overpressure at the vent:

[0062]

[0063] P vi For overpressure in the discharge zone at time i, P ei At time i, the vent is overpressured. This venting process is considered to be in a steady state. The flow velocity at the vent is:

[0064]

[0065] Release gas flow rate:

[0066]

[0067] S vi Let P0 be the velocity of the flow field at time i, P0 be the initial pressure of the flow field, ρ0 be the initial density of the flow field, and C be the initial density of the flow field. d A is the exhaust index of the explosion vent. v The area of ​​the vent.

[0068] The equilibrium pressure within the confined space after venting is calculated using the following mathematical model based on the following equilibrium state:

[0069]

[0070] Where P i-net The predicted overpressure value for explosion relief is the pressure after equilibrium.

[0071] S4. Using the turbulent flame velocity, unburned gas velocity, and macroscopic flame area predicted by the TFT algorithm as input values ​​for the explosion relief overpressure zoning calculation model, the explosion relief overpressure zoning calculation model calculates the predicted explosion relief overpressure value P at the current time step. i-net P i-net The explosion overpressure P obtained from the experiment i-exp The loss value is used to calculate the loss value, which is then used for iterative correction of the TFT algorithm parameters. The training process forms the PINN-TFT algorithm, which includes the following steps:

[0072] 1. Input a set of data from the dataset along with the parameters of the burned zone, unburned zone, venting zone, and equilibrium state output from the explosion overpressure zoning calculation model into the TFT algorithm, and output the turbulent flame velocity S for the next time step. t Unburned gas flow velocity S u and macroscopic flame area A l ;

[0073] 2. Substitute the predicted turbulent flame velocity, unburned gas velocity and macroscopic flame area from step 1 into the explosion venting overpressure zoning calculation model, and calculate the parameters of the burned zone, unburned zone, venting zone and equilibrium state respectively.

[0074] 3. The predicted overpressure value P for the explosion relief i-net The explosion overpressure P obtained from the experiment i-exp Compare, through The loss value is calculated and used for backward gradient propagation in the TFT algorithm, where P i-net It is calculated through step 2.

[0075] 4. Based on the predicted overpressure value P of the explosion relief i-net and static explosion overpressure P stat The relative magnitude is used to determine the current predicted value P of the explosion overpressure. i-net If the static explosion relief overpressure has not been reached, repeat steps 1 to 3 above until the calculation is completed (complete reaction).

[0076] 5. Train on the next set of data, and repeat this process until iterative training on all data is completed;

[0077] S5. Optimize the hyperparameters in the PINN-TFT algorithm using a genetic algorithm to form a gas explosion overpressure prediction model. Use the trained PINN-TFT algorithm to predict explosion overpressure.

[0078] It should be noted that the classical formulas already existing in fluid dynamics are not described in detail in this invention.

[0079] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention.

[0080] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A gas overpressure zoning calculation method combined with a physical information neural network, characterized in that, The method comprises the following steps: inputting the working condition parameters of the construction explosion scene into the trained PINN-TFT algorithm model, judging whether the explosion relief occurs, and if the predicted value of the explosion relief overpressure is greater than the static explosion relief overpressure, the explosion relief occurs.

2. The gas overpressure zoning calculation method combined with physical information neural network according to claim 1, characterized in that, The PINN-TFT algorithm model comprises an explosion relief overpressure partition calculation model and a TFT algorithm, divides the input parameters into static characteristics and dynamic characteristics as the input characteristics of the TFT algorithm, introduces a constraint term in the loss function, and constructs the PINN-TFT algorithm model coupled with physical information.

3. The gas overpressure zoning calculation method combined with physical information neural network according to claim 2, characterized in that, The static characteristics comprise explosion space length, width and height, explosion point coordinates, adiabatic explosion overpressure, laminar flame speed, gas adiabatic index, unburned gas density, burned gas density, explosion relief port area and explosion relief port throttling coefficient. The dynamic characteristics comprise time after ignition and explosion overpressure rising rate.

4. The gas overpressure zoning calculation method combined with physical information neural network according to claim 2, characterized in that, The explosion relief overpressure partition calculation model comprises mathematical models for calculating parameters of the burned zone, the unburned zone and the explosion relief zone.

5. The gas overpressure zoning calculation method combined with physical information neural network according to claim 4, characterized in that, The mathematical model of the burned zone solves the following parameters: burned zone overpressure increment, burned zone density, burned zone volume and flame propagation distance.

6. The gas overpressure zoning calculation method combined with physical information neural network according to claim 4, characterized in that, The mathematical model of the unburned zone solves the following parameters: unburned zone overpressure, unburned zone volume, unburned zone mass and unburned zone density.

7. The gas overpressure zoning calculation method combined with physical information neural network according to claim 4, characterized in that, The mathematical model of the explosion relief zone solves the following parameters: explosion relief port overpressure, explosion relief port flow field speed and explosion relief port flow rate; and the equilibrium pressure in the confined space after the explosion relief is calculated by the following mathematical model of the equilibrium state: where m bi is the volume of burned gas at time i, p bi is the density of the burned zone, P bi is the overpressure of the burned zone at time i, P i-mry is the predicted overpressure of the vented explosion, γ n represents the adiabatic index of unburned gas, m ui is the total mass of unburned gas, p ui is the density of the unburned gas at time i, P ui is the overpressure of the unburned zone at time i, γ u represents the adiabatic index of unburned gas, V is the volume of the confined space.

8. The gas overpressure zoning calculation method combined with physical information neural network according to claim 1, characterized in that, The training of the PINN-TFT algorithm model comprises the following steps: (1) inputting a group of data of the data set and the burned zone parameters, the unburned zone parameters, the explosion relief zone parameters and the equilibrium state parameters output by the explosion relief overpressure partition calculation model into the TFT algorithm to output the turbulent flame speed, the unburned gas flow rate and the macroscopic flame area at the next time step; (2) substituting the predicted turbulent flame speed, the unburned gas flow rate and the macroscopic flame area in step 1 into the explosion relief overpressure partition calculation model to respectively calculate the burned zone parameters, the unburned zone parameters, the explosion relief zone parameters and the equilibrium state parameters; (3), the predicted value P of the explosion relief overpressure i-net The experimental obtained explosion relief overpressure P i-exp Comparison, through Loss value calculation is carried out, and is used for reverse gradient promotion of TFT algorithm; (4) judging whether the current overpressure prediction value P i-net and the relative size of the static overpressure P stat , whether the current overpressure prediction value P i-net reaches the static overpressure, and repeating steps 1-3 if it does not reach the static overpressure. (5) performing the next group of data training, repeating the above steps 1-4 until the iterative training of all data is completed.

9. A computer device comprising a memory and a processor, the memory and the processor being electrically connected, the memory storing a computer program, characterized in that: The computer program is executed by the processor to realize the method of any one of claims 1-8.

10. A computer-readable storage medium, the computer-readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to realize the method of any one of claims 1-8.

Images