A large underground garage hydrogen leakage prediction and risk assessment method based on an informer model

Abstract

The application belongs to the technical field of hydrogen diffusion prediction, and proposes a large underground garage hydrogen leakage prediction and risk assessment method based on an Informer model. The method establishes a simulation model of the underground garage through Fluent; numerical simulation is carried out on the hydrogen random leakage scene of the automobile when parking in the underground parking lot; the Informer prediction model is obtained through training, the previous time sequence hydrogen concentration is taken as the input, and the next time sequence hydrogen concentration cloud picture is predicted, solving the problem that the hydrogen leakage diffusion range of the large underground garage using induced ventilation cannot be accurately predicted and risk assessed. The hydrogen leakage prediction and risk assessment method constructed by the application can predict long time sequences with high precision and high efficiency, realize the spatio-temporal evolution process of hydrogen leakage, help the emergency handling of hydrogen leakage accidents, and provide certain guidance for the layout of induced fans and other facilities in large underground garages.

Description

Technical Field

[0001] This invention relates to the field of hydrogen diffusion prediction technology, specifically a method for predicting and assessing hydrogen leakage in large underground parking garages based on the Informer model. Background Technology

[0002] Underground parking garages are an important component of modern urban transportation systems, serving to meet parking needs, improve land utilization, and reduce traffic congestion. However, with the rapid development and widespread application of hydrogen energy technology, the number of hydrogen fuel cell vehicles in underground parking garages is constantly increasing. Although hydrogen fuel is efficient, clean, and renewable, hydrogen leaks can pose a higher safety risk compared to traditional fuels. Induced ventilation systems reduce the risk of hydrogen leaks by controlling airflow and emissions. They can effectively remove accumulated gas in the hydrogen leak area, providing a good ventilation environment, thereby reducing the potential risk of fire and explosion, and are commonly used in underground parking garages. Hydrogen has a low density and high flammability; once leaked, it can spread rapidly and form a flammable mixture. In the enclosed environment of an underground parking garage, hydrogen leaks can lead to serious safety problems such as fires, explosions, and personal injury. Therefore, the ability to accurately predict the spread range and severity of hydrogen leaks is particularly important. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting hydrogen leakage and diffusion in large underground parking garages with induced ventilation and to conduct risk assessment, so as to solve the problem that the existing technology is insufficient in predicting leakage in induced ventilation and large underground parking lots.

[0004] To achieve the above objectives, the present invention provides the following solution: a method for predicting and assessing the risk of hydrogen leakage in a large underground parking garage based on the Informer model, comprising the following steps:

[0005] Step 1: Construct a full-scale simulation model of the underground parking garage using Fluent; Place sensors around the leak point to be evaluated to collect hydrogen concentration data around the vehicle.

[0006] Step 2: Set initial and boundary conditions, and perform Fluent numerical simulation on the simulation model from Step 1 to obtain hydrogen concentration data and hydrogen concentration cloud maps at different times.

[0007] Step 3: Extract features from the hydrogen concentration obtained in Step 2 using a deep neural network and input them into Informer for model training to obtain a hydrogen concentration cloud map prediction model.

[0008] The deep neural network comprises fully connected neural layers, remodeled neural layers, and deconvolutional layers. The fully connected neural layers are used to extract the hidden key feature H1.

[0009] H1 = Relu(ω1*X + b1)

[0010] Where ReLU is the activation function, w1 and b1 are the weights and biases of the fully connected neural layer, respectively, and X is the hydrogen concentration of the sensor.

[0011] Key feature H1 is transformed into two-dimensional feature H2 through non-parametric reshaping of the neural layer:

[0012] H2 = Reshape(H1)

[0013] Two-dimensional feature H2 is used to extract key two-dimensional features through a deconvolution layer, generating the spatial hydrogen concentration Y:

[0014]

[0015] in This represents the convolution operation; ω2 and b2 are the weights and biases of the deconvolution layer, respectively.

[0016] The Informer's encoder incorporates a probabilistic sparse self-attention mechanism;

[0017] Step 4: Input the hydrogen concentration obtained by actual measurement or simulation from the sensor into the relational model in Step 3 to predict the hydrogen concentration at different times and the hydrogen concentration cloud map.

[0018] Step 5: Based on the prediction results in Step 4, calculate the TNT equivalent of the hydrogen concentration and the radii of the dead zone, the seriously injured zone, and the slightly injured zone at the leak point.

[0019] The numerical simulation process is as follows: a model of the underground parking garage is established to obtain the internal fluid domain of the underground parking garage. Each induced fan is set as an internal fluid domain, different boundaries are grouped and established, and the overall topology is set to be shared.

[0020] After the model is established, the model is meshed. Fluent numerical simulation can be divided into the following three steps: (1) preprocessing stage, (2) solution stage, (3) post-processing stage. In the preprocessing stage, Fluent Meshing is used to mesh the model with hexahedral meshes, and local meshing is performed at the leak point, vehicle and surrounding pillars, nozzle of the induced draft fan and air inlet of the fan. In the solution stage, Fluent is used to select the solution equation, set the fluid material and properties, set the boundary conditions and solution control parameters, and set the cloud map and sensor position. For post-processing, after obtaining the results, the saved file is imported into CFD for post-processing.

[0021] Step 2: Modeling and Simulation Analysis of Hydrogen Leakage and Diffusion. First, the Fluent environment is set up.

[0022] After importing the model into Fluent, the units were first converted. In the General settings, transient simulation was selected. Due to the light weight of hydrogen, gravity has a significant impact on the leakage diffusion process. Therefore, the effect of gravity needs to be considered, and the gravitational acceleration in the vertical direction was set to -9.8 m / s². 2 .

[0023] A turbulence model is adopted. In the Model settings, the energy equation, κ-ε equation and component transport equation are enabled. The component transport equation is set to inlet diffusion and the components are defined as hydrogen and air. The mixed gas is set as an ideal gas.

[0024] The initial conditions for the numerical simulation (orifice size of the leak, hydrogen tank pressure, wind speed) were set. Boundary conditions were set as follows: the fan nozzle and fan contact surface were set to the interior surface; the induced draft fan nozzle was set to a velocity inlet; the air inlet and outlet of the ventilation system were set to velocity inlets; the garage entrance and exit were set to pressure outlets; backflow suppression was configured; and the leak outlet was set to a mass flow inlet. The mass flow rate of the leak outlet was calculated using the following formula:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] Where ρ1 is the initial hydrogen concentration, P1 is the initial pressure, b is the Abel-Noble residual capacity coefficient, and R H2 Let ρ be the hydrogen constant, T1 be the initial temperature, κ be the adiabatic index, ρ3, T3, P3, u3 be the hydrogen density, hydrogen temperature, hydrogen pressure, and hydrogen velocity at the leak point, respectively, and q be the hydrogen constant. m d represents the mass flow rate, and d represents the diameter of the leak.

[0032] After setting up, perform global initialization of initial and boundary conditions, and then set the step size for simulation calculation.

[0033] Dataset Construction: The transient hydrogen concentration data from each sensor obtained through numerical simulation were preprocessed. The preprocessing methods included: cleaning the hydrogen concentration data at all unit time points to remove outliers; converting all operating conditions and environmental variables into numerical features; and performing Z-score normalization on the preprocessed dataset using the mean and standard deviation. The dataset was then divided into training, validation, and test sets. The calculation formula is as follows:

[0034]

[0035] Where X is the original dataset, mean(x) is the mean, and std(x) is the standard deviation of the data. The transient hydrogen concentration data is concatenated with the features. For each time step of the input sequence, hydrogen concentration features are extracted and embedded. The transformed feature vector is passed as input to the encoder layer of the model. Multiple Transformer encoder layers are used to extract the feature representation of the input sequence. Each encoder layer contains a multi-head attention mechanism and a feedforward neural network.

[0036] The feature sequence is input into the Informer for encoding, and a probabilistic sparse self-attention mechanism is introduced. The calculation formula is as follows:

[0037] Self-attention score:

[0038]

[0039] Self-attention weight:

[0040] Attention(Q,K,V)=softmax(Score(Q,K))V

[0041] Where Q represents the query vector, K represents the key vector, and V represents the value vector;

[0042] For each query, a portion of the key is randomly sampled, and the sparsity score of each query is calculated. The N queries with the highest sparsity scores are selected, and the dot product of the N queries and all keys is calculated to obtain the attention result. The remaining LN queries are not calculated. The average of the inputs to the Self-attention layer is used as the output.

[0043] The encoder's output and historical predictions are used as inputs to the decoder, which consists of multiple decoder blocks. The input to each decoder block is the predicted hydrogen concentration from the previous time step. After passing through a multi-head self-attention mechanism, the output of the decoder is fused with the encoder's output through an encoder-decoder attention mechanism, and then processed through a fully connected feed-forward network.

[0044] FFN(X)=ReLU(XW1+b1)W2+b2

[0045] Where W1 and W2 represent weight matrices, and b1 and b2 represent bias vectors;

[0046] The model is trained and its parameters optimized using a pre-defined dataset. During training, the transient hydrogen concentration obtained through numerical simulation in step 2 is used as the target sequence and input into the model along with other input features. During training, a loss function is used to measure the difference between the predicted and true sequences. The loss function uses the mean squared error (MSE) formula as follows:

[0047]

[0048] After training, the trained Informer model can be used to predict new input features and obtain predictions of hydrogen leakage and diffusion.

[0049] The TNT equivalent is calculated using the predicted transient hydrogen concentration data at different times to conduct a risk assessment of hydrogen leakage in the underground parking garage. The calculation formula is as follows:

[0050]

[0051] Among them W TNT α is the equivalent mass of TNT; α is the explosion efficiency coefficient of the combustible vapor cloud; A is the ground explosion coefficient; W f It is the mass of hydrogen; Q f It is the heat of combustion of hydrogen; Q TNT It is the explosive energy per unit mass of TNT explosive.

[0052] The radius of death, radius of serious injury, and radius of minor injury are calculated based on the obtained TNT equivalent mass. This allows for the division of the injury area into death zone, serious injury zone, and minor injury zone based on the degree of harm suffered. The formula for calculating the danger radius is:

[0053]

[0054]

[0055]

[0056] Where P0 is the overpressure value.

[0057] Based on the risk assessment results, develop corresponding response measures and emergency plans, such as improving ventilation and evacuating personnel, to minimize risks and harm.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] This invention provides a method for predicting and assessing hydrogen leakage in large underground parking garages based on the Informer model. This prediction and risk assessment method uses the hydrogen concentration of the previous time series as input to predict the hydrogen concentration cloud map of the next time series. It can accurately predict the hydrogen leakage concentration and hydrogen concentration cloud map in large underground induced ventilation parking garages, solving the problems of complex and time-consuming calculations and inability to predict diffusion processes in short timeframes in traditional numerical simulation methods. It also overcomes the shortcomings of conventional experimental methods, which suffer from large errors and inaccurate predictions in complex environments. The hydrogen leakage prediction and risk assessment method constructed in this invention can perform high-precision and high-efficiency predictions for long-term time series, realizing the spatiotemporal evolution of hydrogen leakage, which is helpful for emergency response to hydrogen leakage accidents and can provide some guidance for the layout of induced ventilation fans and other facilities in large underground parking garages. Attached Figure Description

[0060] Figure 1 A three-dimensional view of a simulation model of an underground parking garage for an embodiment of the present invention.

[0061] Figure 2 This is an overhead view of the underground parking garage.

[0062] The triangle represents the location of the induced draft fan, and the rectangle represents the load-bearing column inside the garage.

[0063] Figure 3 This is a flowchart of a method for predicting and assessing hydrogen leakage in a large underground parking garage based on the Informer model.

[0064] Figure 4 This is a diagram illustrating the prediction steps of the Informer model.

[0065] Figure 5 The figures shown are numerical simulation results and prediction results of the simulation model in the embodiment of the present invention.

[0066] In the figure, a is the result of numerical simulation, and b is the result of prediction. Detailed Implementation

[0067] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and detailed descriptions. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0068] This embodiment simulates an underground parking garage ventilation system that uses induced ventilation, consisting of two air inlets, two air outlets, and 60 induced ventilation fans. Induced ventilation fans, also known as jet fans or relay fans, have a small air volume of their own and transfer air through induction. Induced ventilation fans are commonly used in underground parking garage ventilation systems to agitate and eliminate localized dead air zones, improving localized air quality while saving space.

[0069] Figure 3 A flowchart illustrating a method for predicting and assessing hydrogen leakage in a large underground parking garage based on the Informer model is shown. This prediction and risk assessment method specifically includes the following steps:

[0070] Step 1: Conduct a simulation experiment of a hydrogen leak scenario in an underground parking garage, including: constructing a full-scale simulation model of the underground parking garage, with a computational domain of 205m*165m*3.4m, a fan height of 2.9m, an 8m interval between adjacent fans, fan dimensions of 0.6m*0.5m*0.26m, a leak outlet height of 0.15m above the ground, and orifice sizes of 0.5mm, 1mm, and 2mm; setting initial conditions, with the leak outlet located at... Figure 1 The initial conditions for the leak were the diameter of the leak vent, the hydrogen tank pressure, and the wind speed at the bottom of the vehicle. Twenty sensors (spaced 7m apart in a square) were evenly placed on the ceiling directly above the leak vent (3.4m high) to collect data on the hydrogen concentration around the vehicle during the leak process. Figure 1 and 2 (As shown).

[0071] Step 2: Set initial and boundary conditions, and perform Fluent numerical simulation on the simulation model from Step 1 to obtain hydrogen concentration data and hydrogen concentration cloud maps at different times.

[0072] The numerical simulation process is as follows: Based on the underground parking garage simulation model established in step 1, the internal fluid domain of the underground parking garage is obtained. Each induced draft fan is set as an internal fluid domain, and different boundaries are grouped together, with the overall topology set to share. The model is then meshed; local refinement is applied at the leak point and the induced draft fan using a hexahedral mesh.

[0073] After importing the meshed model into Fluent, first convert the units. In General, select Transient Simulation. Since hydrogen is lightweight, gravity has a significant impact on the leakage diffusion process. Therefore, the effect of gravity needs to be considered. Set the gravitational acceleration in the Y direction to -9.8 m / s². 2 .

[0074] A turbulence model is adopted. In the Model settings, the energy equation, κ-ε equation and component transport equation are enabled. The component transport equation is set to inlet diffusion and the components are defined as hydrogen and air. The mixed gas is set as an ideal gas.

[0075] Initial conditions for numerical simulation: initial temperature T1 is 288.15 K.

[0076] Boundary conditions were set as follows: the nozzles of the induced draft fan were set as velocity inlets, with air velocities of 5 m / s, 10 m / s, and 15 m / s selected at the nozzles to discuss the impact of different air velocities on hydrogen leakage and diffusion. The air inlets and outlets of the ventilation system were also set as velocity inlets with air velocities of 5 m / s and 10 m / s, respectively. The outlets and inlets of the garage were set as pressure outlets with backflow suppression. The leak outlet was set as a mass flow inlet. The internal hydrogen pressure of domestic hydrogen tanks is typically 35 MPa, and the Abel-Noble residual capacity coefficient b = 7.69 * 10^6 m / s. -3 Hydrogen constant R H2 =4.12424*10 3 The mass flow rates at the leak point were calculated to be 0.0045 kg / s, 0.018 kg / s, and 0.073 kg / s, respectively, with a thermal index κ of 1.4. After initializing the initial and boundary conditions, the time step was set to 0.01 s and the number of time steps was 10000. The spatiotemporal evolution of hydrogen over 100 s was calculated to obtain the hydrogen concentration and hydrogen concentration cloud map.

[0077] Step 3: Extract features from the hydrogen concentration obtained from the numerical simulation in Step 2 using a convolutional neural network, input them into the Informer model for training, and perform predictions under different operating conditions to achieve spatiotemporal evolution prediction of random hydrogen leakage (e.g., Figure 4 (As shown).

[0078] Dataset Construction: The transient hydrogen concentration data of each sensor obtained from numerical simulation were preprocessed. The preprocessing method was as follows: data cleaning was performed on the hydrogen concentration data at all unit time points to remove outliers; the wind speed, orifice size and leakage location of the induced fan were converted into numerical features; the preprocessed two-dimensional hydrogen concentration dataset was normalized by Z-score using the mean and standard deviation of the following formula; and the dataset was divided into training set (70%), validation set (10%) and test set (20%).

[0079]

[0080] Where X is the original dataset, mean(x) is the mean, and std(x) is the standard deviation of the data; the transient hydrogen concentration data is concatenated with the features.

[0081] The transient hydrogen concentration data from each sensor was analyzed using a deep neural network to extract features, resulting in a feature sequence of hydrogen concentration.

[0082] Deep neural networks consist of fully connected neural layers, remodeled neural layers, and fan convolutional layers. The fully connected neural layers are used to extract the hidden key features H1.

[0083] H1 = Relu(ω1*X + b1)

[0084] Where ReLU is the activation function, w1 and b1 are the weights and biases of the fully connected neural layer, respectively, and X is the hydrogen concentration of the sensor.

[0085] Key feature H1 is transformed into two-dimensional feature H2 through non-parametric reshaping of the neural layer:

[0086] H2 = Re shape(H1)

[0087] Two-dimensional feature H2 is used to extract key two-dimensional features through a deconvolution layer, generating the spatial hydrogen concentration Y:

[0088]

[0089] in ω represents the convolution operation; ω2 and b2 are the weights and biases of the deconvolution layer, respectively.

[0090] The feature sequence is input into the Informer encoder for encoding, and a probabilistic sparse self-attention mechanism is introduced. The calculation formula is as follows:

[0091] Self-attention score:

[0092]

[0093] Self-attention weight:

[0094] Attention(Q,K,V)=softmax(Score(Q,K))V

[0095] Where Q represents the query vector, K represents the key vector, and V represents the value vector;

[0096] For each query, a portion of the key is randomly sampled, and the sparsity score of each query is calculated. The N queries with the highest sparsity scores are selected, and the dot product of the N queries and all keys is calculated to obtain the attention result. The remaining LN queries are not calculated. The mean (mean(V)) of the input of the Self-attention layer is taken as the output.

[0097] After feature extraction and embedding of hydrogen concentration at each time step of the input sequence, the resulting transformed feature vector is passed as input to the encoder layer of the model. Multiple Transformer encoder layers are used to extract feature representations of the input sequence, each containing a multi-head attention mechanism and a feedforward neural network. The feature sequence is then input into the Informer encoder for encoding, introducing a probable sparse self-attention mechanism. The encoder output and historical predictions are used as input to the decoder, which consists of three decoder blocks. The input to each decoder block is the predicted hydrogen concentration from the previous time step. After passing through the multi-head self-attention mechanism, this is fused with the encoder output through an encoder-decoder attention mechanism, and then processed by a fully connected feedforward network.

[0098] FFN(X)=ReLU(XW1+b1)W2+b2

[0099] Where W1 and W2 represent weight matrices, and b1 and b2 represent bias vectors;

[0100] The model is trained and its parameters optimized using a pre-defined dataset. During training, the transient hydrogen concentration obtained through numerical simulation in step 2 is used as the target sequence and input into the model along with other input features. During training, a loss function is used to measure the difference between the predicted and true sequences; the loss function is the mean squared error (MSE).

[0101]

[0102] After training, the trained Informer model is used to predict new input features. Figure 5The graph displays a comparison between the simulated hydrogen concentration data (5-10 seconds) and the numerical simulation results for hydrogen concentration (11-15 seconds). In each sub-graph, the light gray area represents the hydrogen diffusion range. The comparison shows that the model's predictions and simulation results for hydrogen diffusion are highly consistent.

[0103] The TNT equivalent is calculated using the predicted transient hydrogen concentration data at different times to conduct a risk assessment of hydrogen leakage in the underground parking garage. The calculation formula is as follows:

[0104]

[0105] Among them W TNT α is the equivalent mass of TNT; α is the explosion efficiency coefficient of the combustible vapor cloud; A is the ground explosion coefficient; W f It is the mass of hydrogen; Q f It is the heat of combustion of hydrogen; Q TNT It is the explosive energy per unit mass of TNT explosive.

[0106] The radius of death, radius of serious injury, and radius of minor injury are calculated based on the obtained TNT equivalent mass. This allows for the division of the injury area into death zone, serious injury zone, and minor injury zone based on the degree of harm suffered. The formula for calculating the danger radius is:

[0107]

[0108]

[0109]

[0110] This simulation yielded W after 20 seconds of leakage. f It is 1.46 kg, and W is calculated to be... TNT The weight is 3.18 kg. In the event of an explosion, the fatal radius would be 1.62 m, the radius of serious injury 3.78 m, and the radius of minor injury 6.43 m. Based on the risk assessment results, corresponding response measures and emergency plans should be developed, such as enhanced ventilation and personnel evacuation, to minimize risks and hazards.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting and assessing the risk of hydrogen leakage in large underground parking garages based on the Informer model, characterized in that, Includes the following steps: Step 1: Construct a full-scale simulation model of the underground parking garage using Fluent; Place sensors around the leak point to be evaluated to collect hydrogen concentration data around the vehicle. Step 2: Set initial and boundary conditions, perform Fluent numerical simulation on the simulation model from Step 1, and obtain hydrogen concentration data and hydrogen concentration cloud map at different times. Step 3: Extract features from the hydrogen concentration obtained in Step 2 using a deep neural network, and then input them into Informer for model training to establish a prediction model for hydrogen concentration and hydrogen concentration cloud map. The deep neural network comprises fully connected neural layers, remodeled neural layers, and fan convolutional layers. The fully connected neural layers are used to extract the hidden key feature H1. H1 = Relu(ω1*X + b1) Where ReLU is the activation function, w1 and b1 are the weights and biases of the fully connected neural layer, respectively, and X is the hydrogen concentration of the sensor. Key feature H1 is transformed into two-dimensional feature H2 through non-parametric reshaping of the neural layer: H2 = Reshape(H1) Two-dimensional feature H2 is used to extract key two-dimensional features through a deconvolution layer, generating the spatial hydrogen concentration Y: in This represents the convolution operation; ω2 and b2 are the weights and biases of the deconvolution layer, respectively. The Informer's encoder incorporates a probabilistic sparse self-attention mechanism; Step 4: Input the hydrogen concentration obtained by actual measurement or simulation from the sensor into the prediction model in Step 3 to predict the hydrogen concentration at different times and the hydrogen concentration cloud map. Step 5: Based on the prediction results in Step 4, calculate the TNT equivalent of the hydrogen concentration and the radii of the dead zone, the severely injured zone, and the slightly injured zone at the leak point.

2. The method for predicting and assessing the risk of hydrogen leakage in a large underground parking garage based on the Informer model as described in claim 1, characterized in that: The calculation formula for the probabilistic sparse self-attention mechanism is as follows: Self-attention score: Self-attention weights: Attention(Q,K,V)=softmax(Score(Q,K))V Where Q represents the query vector, K represents the key vector, and V represents the value vector; for each query, a portion of the key is randomly sampled, and the sparsity score of each query is calculated. The N queries with the highest sparsity scores are selected, and only the dot product of the N queries and all keys is calculated to obtain the attention result. The remaining LN queries are not calculated. The mean (V) of the input to the Self-attention layer is taken as the output.

3. The method for predicting and assessing the risk of hydrogen leakage in a large underground parking garage based on the Informer model as described in claim 1, characterized in that: In step 3, the encoder output and historical predictions are used as inputs to the decoder. The decoder consists of multiple decoder blocks, and the input to each decoder block is the hydrogen concentration prediction from the previous time step. After passing through a multi-head self-attention mechanism, the hydrogen concentration is fused with the encoder output using an encoder-decoder attention mechanism. Then, it is processed through a fully connected feedforward network to obtain the predicted hydrogen concentration for the current time step. FFN(X)=ReLU(XW1+b1)W2+b2 Where W1 and W2 represent weight matrices, and b1 and b2 represent bias vectors.

4. The method for predicting hydrogen leakage in large underground parking garages based on the Informer model according to claim 1, characterized in that: In step 3, during model training, a loss function is used to measure the difference between the predicted sequence and the true sequence. The loss function uses mean squared error.

5. The method for predicting and assessing hydrogen leakage in large underground parking garages based on the Informer model according to claim 1, characterized in that: Step 2 involves performing mesh independence analysis and model verification, selecting the solution equation, setting the fluid material and properties, setting boundary conditions, solving the control parameters, and then performing calculations. A turbulence model is adopted. In the Model settings, the energy equation, κ-ε equation and component transport equation are enabled. The component transport equation is set to inlet diffusion and the components are defined as hydrogen and air. The mixed gas is set as an ideal gas. The induced draft fan nozzles, the air inlet and outlet of the ventilation system are set as velocity inlets, the garage outlet and inlet are set as pressure outlets, backflow suppression is implemented, and the leak outlet is set as a mass flow inlet. The mass flow rate of the leak outlet is calculated using the following formula: Where ρ1 is the initial hydrogen concentration, P1 is the initial pressure, b is the Abel-Noble residual capacity coefficient, and R H2 Let ρ be the hydrogen constant, T1 be the initial temperature, κ be the adiabatic index, ρ3, T3, P3, u3 be the hydrogen density, hydrogen temperature, hydrogen pressure, and hydrogen velocity at the leak point, respectively, and q be the hydrogen constant. m Here, d represents the mass flow rate, and d represents the diameter of the leak. After setting up, perform global initialization of initial and boundary conditions, and then set the step size for simulation calculation.

6. The method for predicting hydrogen leakage in large underground parking garages based on the Informer model according to claim 1, characterized in that: Step 5 uses the predicted transient hydrogen concentration data at different times to calculate the TNT equivalent using the following formula: Among them, W TNT α is the equivalent mass of TNT; α is the explosion efficiency coefficient of the combustible vapor cloud; A is the ground explosion coefficient; W f Q is the mass of hydrogen gas; f It is the heat of combustion of hydrogen; Q TNT It is the explosive energy per unit mass of TNT explosive; The following formula is used to calculate the radius of death, radius of serious injury, and radius of minor injury at the leak point to be evaluated:

Images