Rapid deduction method and system for hydrogen leakage consequence of hydrogen refueling station

Abstract

The invention discloses a hydrogen refueling station hydrogen leakage consequence rapid deduction method, and belongs to the technical field of hydrogen refueling station safety monitoring. The method comprises the following steps: establishing a hydrogen refueling station hydrogen leakage-diffusion numerical model; arranging a plurality of hydrogen detectors in the model for monitoring hydrogen concentration time sequence data; setting a hydrogen concentration slice to obtain hydrogen concentration field time sequence data; simulating different leakage scenes and the like, performing numerical simulation calculation according to the model to obtain hydrogen leakage consequence data in the hydrogen refueling station, and establishing a hydrogen leakage consequence database of the hydrogen refueling station; dividing data stored in the database into a training set, a verification set and a test set for prediction model training; training the long-short-term memory neural network model to predict the hydrogen concentration of the hydrogen detector at the future moment; and training the transposed convolutional neural network model to deduce the hydrogen concentration field at the future moment. According to the method, the consequence of the hydrogen leakage accident of the hydrogen refueling station can be quickly predicted and deduced, so that guidance can be provided for emergency measures and personnel evacuation.

Description

Technical Field

[0001] The present invention belongs to the technical field of safety monitoring of hydrogen refueling stations, and in particular relates to a method and system for quickly deducing the consequences of hydrogen leakage in a hydrogen refueling station. Background Art

[0002] At present, hydrogen is the main body of clean energy, and the development and utilization of hydrogen energy has become an important direction of the new round of energy technology transformation. As an important infrastructure for the development of the hydrogen energy industry and hydrogen fuel vehicles, the safe operation of hydrogen refueling stations has been promoted to the top priority of the development of the hydrogen energy industry. Compared with other clean energy sources, hydrogen has unfavorable characteristics such as flammability and explosion, low ignition energy, large buoyancy, and invisible flames in the sun. Once the hydrogen stored in the hydrogen refueling station leaks, it is easy to cause an explosion. Relying on scientific and reasonable prediction and deduction methods to quickly predict and deduce the diffusion range of leaked hydrogen is of great significance for warning potential hydrogen fire or explosion risks and guiding emergency response and personnel evacuation.

[0003] The existing hydrogen leakage alarm, early warning and prediction technologies for hydrogen refueling stations mainly include the following:

[0004] (1) Traditional alarm method: by deploying a certain number of hydrogen concentration detectors in the tunnel (pipeline corridor), and transmitting the time-series gas concentration data obtained by the detectors to the computer, when a certain threshold is reached, an alarm signal is issued to indicate the occurrence of hydrogen leakage. For example, the Chinese utility model patent with publication number CN218441824U discloses a hydrogen refueling station safety monitoring system, including a monitoring module, a control module and an early warning module. Among them, the monitoring module includes a combustible gas sensor for collecting the concentration of hydrogen, so as to facilitate timely detection of whether there is a hydrogen leak. The control module is connected to the monitoring module and the early warning module, and is used to receive monitoring data from the monitoring module. When a hydrogen leak is found, the early warning function of the early warning module is activated.

[0005] (2) Robot inspection and alarm technology: China Utility Model Patent Publication No. CN213479860U discloses a hydrogen station hydrogen process pipeline intelligent inspection robot system, which determines whether there is a hydrogen leak based on the change in hydrogen concentration detected by the sensor. China Invention Patent Application Publication No. CN117532634A discloses a hydrogen station intelligent inspection robot system, which can achieve comprehensive monitoring of the entire hydrogen station through the collaborative operation of multiple robots. Once a fire risk is detected, an alarm signal is immediately sent, and the monitoring data (such as the location of the fire risk point, time, detection results, etc.) is transmitted to the remote monitoring platform in real time.

[0006] (3) Risk assessment and early warning technology: A Chinese invention patent with authorization announcement number CN114707916B discloses a method for quantitatively assessing hydrogen leakage safety accidents at hydrogen refueling stations. According to the source of the safety accident, a risk assessment model is used to conduct a quantitative risk assessment of hydrogen leakage safety accidents at hydrogen refueling stations, obtain assessment results, determine risk causative factors, and provide guidance for improving the safety level of hydrogen refueling stations.

[0007] (4) Hydrogen leakage diffusion range prediction technology: The Chinese invention patent with authorization announcement number CN113128755B discloses a method and system for predicting the diffusion range of liquid hydrogen leakage. Different liquid hydrogen evaporation and diffusion models are established to deduce the spatiotemporal distribution characteristics of hydrogen concentration in the air after liquid hydrogen leakage, predict hydrogen explosion hazard areas and issue an alarm for ranges exceeding the safety threshold.

[0008] Although existing technologies provide guidance for improving the safety level of hydrogen refueling stations, there are still deficiencies. First, traditional alarm methods and robot inspection alarm technology rely on real-time concentration information recorded by hydrogen detectors to determine whether a hydrogen leakage accident has occurred, which is essentially an alarm technology. These methods cannot predict the development trend of hydrogen leakage, nor can they present the distribution of hydrogen concentration. Second, risk assessment and early warning technology is a hazard source search technology that cannot predict the development behavior of hydrogen leakage, nor can it deduce the consequences of accidents. Third, the hydrogen leakage diffusion range prediction technology is an empirical formula for describing the diffusion range of hydrogen leakage, which is summarized based on a large amount of experimental data. Although this type of method can obtain the risk value or impact range after the leak in a shorter time, the formula has a limited scope of application and cannot predict the evolution of hydrogen concentration under complex conditions such as different leakage locations and environmental factors. In hydrogen leakage accidents at hydrogen refueling stations, it is very important to know the future distribution trend of the leaked hydrogen concentration, which can provide guidance for emergency measures and personnel evacuation. Summary of the invention

[0009] Based on the above technical problems, the present invention proposes a method and system for quickly deducing the consequences of hydrogen leakage at a hydrogen refueling station, so as to achieve rapid prediction and deduction of the consequences of hydrogen leakage accidents at hydrogen refueling stations.

[0010] The technical solution adopted by the present invention is:

[0011] A method for rapidly simulating the consequences of hydrogen leakage at a hydrogen refueling station comprises the following steps:

[0012] Step S1, establishing a hydrogen leakage-diffusion numerical model of a hydrogen refueling station;

[0013] Step S2, arranging a plurality of hydrogen detectors in the hydrogen leakage-diffusion numerical model of the hydrogen refueling station to monitor the hydrogen concentration time series data at the location of the hydrogen detectors under the condition of hydrogen leakage and diffusion; setting the hydrogen concentration slice to obtain the hydrogen concentration field time series data under the condition of hydrogen leakage and diffusion;

[0014] Step S3, simulating different leakage scenarios, leakage modes and influencing factors, and obtaining hydrogen leakage consequence data in the hydrogen refueling station according to the hydrogen leakage-diffusion numerical model of the hydrogen refueling station and the hydrogen detectors arranged in the model;

[0015] Step S4, pre-processing the hydrogen leakage consequence data in the hydrogen refueling station obtained in step S3, and establishing a hydrogen leakage consequence database for the hydrogen refueling station;

[0016] Step S5, dividing the data in the hydrogen refueling station hydrogen leakage consequence database into a training set, a validation set and a test set for prediction model training;

[0017] Step S6: construct a hydrogen concentration prediction model framework based on the long short-term memory neural network, and train the long short-term memory neural network according to the data in the hydrogen leakage consequence database of the hydrogen filling station; and predict the hydrogen concentration time series data of the hydrogen detector at the future time through the constructed hydrogen concentration prediction model framework;

[0018] Step S7, building a deduction model framework based on the transposed convolutional neural network, training the transposed convolutional neural network according to the hydrogen concentration time series data and hydrogen concentration field time series data of the hydrogen detector at the future moment; and deducing the hydrogen concentration field at the future moment through the built deduction model framework;

[0019] Step S8: When hydrogen leaks at the hydrogen refueling station, the hydrogen concentration prediction model framework constructed in step S6 is input according to the data measured by the hydrogen detector, and then combined with the deduction model framework constructed in step S7 to quickly deduce the consequences of hydrogen leakage at the hydrogen refueling station.

[0020] In the above step S1: according to the physical characteristics of the hydrogen refueling station (geometric features, material properties of the building and equipment, etc.), a full-scale hydrogen leakage-diffusion numerical model of the hydrogen refueling station is established using CFD simulation software.

[0021] In the above step S2: the number and coordinates of hydrogen detectors are set to be consistent with the actual scene of the hydrogen refueling station; the hydrogen concentration slicing is a function in the CFD simulation software, which can output the two-dimensional distribution result of hydrogen concentration in a specific plane for training the transposed convolutional neural network.

[0022] In the above step S3: the hydrogen leakage consequence data in the hydrogen filling station includes the hydrogen concentration time series data at the location of the hydrogen detector and the hydrogen concentration field time series data at the slice location.

[0023] In the above step S3: the leakage scenario refers to the potential hydrogen leakage area in the hydrogen refueling station, including the transmission pipeline, hydrogen storage area, compression area, and hydrogenation area; the leakage mode refers to the amount of hydrogen leakage, including micro-leakage, small flow leakage, large flow leakage and transient large-scale leakage; the influencing factors refer to the environmental factors that affect the distribution of leakage hydrogen concentration, including wind speed and wind direction indicators.

[0024] In the above step S6: the long short-term memory neural network is a time recurrent neural network, which is used to capture the long-term dependencies in the hydrogen concentration time series data, and output the hydrogen concentration time series data at a future moment by inputting the historical hydrogen concentration data.

[0025] In the above step S6: the long short-term memory neural network includes an input layer for receiving external data and passing it to the hidden layer;

[0026] The hidden layer is responsible for processing the input data and updating its own information by combining the current input and the state of the previous time step, thereby capturing the long-term dependencies in the time series. As time goes by, the hidden layer gradually accumulates and updates important historical information.

[0027] And the output layer is used to receive the final state of the hidden layer and transform it into the actual output.

[0028] In the above step S7: the transposed convolutional neural network is an upsampling convolutional neural network, which is used to gradually deduce the discrete hydrogen concentration monitoring point data to the two-dimensional hydrogen concentration field distribution; by inputting the future hydrogen concentration time series data output by the long short-term memory neural network, the two-dimensional hydrogen concentration field distribution is output, and the consequences of hydrogen leakage at the hydrogen refueling station are quickly deduced.

[0029] In the above step S7: the transposed convolutional neural network includes an input layer for receiving the future hydrogen concentration time series data output from the long short-term memory neural network, and these data are one-dimensional vectors;

[0030] The fully connected layer is used to expand the dimension and increase the amount of information;

[0031] The reshaping layer is used to convert one-dimensional data into a low-resolution two-dimensional form; through the reshaping process, the data can better adapt to subsequent convolution processing;

[0032] The transposed convolution layer is used to gradually improve the spatial resolution of the model. Through multiple upsampling, a higher resolution hydrogen concentration image or scene is restored. The transposed convolution layer not only helps to restore spatial information, but also processes the nonlinear characteristics in the data through the activation function layer, further enhancing the model's ability to fit complex patterns. Finally, after multiple layers of convolution and upsampling, the required high-resolution output image or signal is generated.

[0033] and an output layer for outputting a high-resolution output image or signal.

[0034] The present invention also provides a rapid deduction system for the consequences of hydrogen leakage in a hydrogen refueling station, which includes a perception module, a communication module, a data storage module, a data processing module and a leakage consequence display module; the perception module is connected to the data storage module through the communication module, the data storage module is connected to the data processing module, and the data processing module is connected to the leakage consequence display module;

[0035] The sensing module includes a plurality of hydrogen detectors, which are used to collect the concentration of hydrogen leaked from the hydrogen filling station in real time;

[0036] The communication module transmits the real-time hydrogen concentration data monitored by the sensing module to the data storage module;

[0037] The data storage module is used to store the real-time hydrogen concentration data collected by the hydrogen detector, and to store the model prediction and deduction results;

[0038] The data processing module includes a trained hydrogen leakage concentration real-time prediction submodule and a hydrogen leakage consequence rapid deduction submodule; the hydrogen leakage concentration real-time prediction submodule processes the hydrogen concentration time series data of the real-time hydrogen detector position recorded in the data storage module, and outputs the hydrogen concentration at a certain moment in the future; the hydrogen leakage consequence rapid deduction submodule processes the hydrogen concentration at a certain moment in the future predicted by the hydrogen leakage concentration real-time prediction submodule, and outputs the hydrogen concentration field at a certain moment in the future;

[0039] The leakage consequence display module receives the hydrogen concentration field information output by the data processing module, and displays the distribution of hydrogen concentration in the hydrogen filling station.

[0040] The beneficial technical effects of the present invention are:

[0041] The method and system for rapidly deducing the consequences of hydrogen leakage at a hydrogen refueling station of the present invention can obtain the future distribution trend of the concentration of leaked hydrogen in the event of a hydrogen leakage accident at a hydrogen refueling station, thereby realizing rapid prediction and deduction of the consequences of a hydrogen leakage accident at a hydrogen refueling station, thereby providing guidance for emergency measures and personnel evacuation.

[0042] Specifically:

[0043] (1) The present invention establishes a hydrogen concentration prediction model for a hydrogen refueling station. The model is built based on a long short-term memory neural network, so that it can process discrete time series data from hydrogen detectors and predict the hydrogen concentration at the location of the hydrogen detector in the next few hours, effectively realizing early forecasting.

[0044] (2) The present invention combines two deep learning methods, long short-term memory neural network and transposed convolutional neural network, which not only improves the prediction accuracy of hydrogen concentration and realizes the advance forecast of hydrogen concentration, but also optimizes the spatial deduction efficiency of hydrogen distribution and realizes the function of deducing hydrogen distribution.

[0045] (3) The system of the present invention realizes data transmission, storage, model prediction and deduction, and result display, and displays the deduction results through a user-friendly graphical interface, so that operators can interact with the model deduction results intuitively. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0047] Figure 1 It is a schematic flow chart of a method for rapidly deducing the consequences of hydrogen leakage at a hydrogen refueling station according to the present invention;

[0048] Figure 2 This is an example diagram of a three-dimensional virtual model of a hydrogen refueling station built using CFD simulation software;

[0049] Figure 3 It is a schematic diagram of the sampling principle of the sliding window method;

[0050] Figure 4 This is an example of reducing the RGB cloud image and converting it into a grayscale cloud image;

[0051] Figure 5 This is an example diagram of the hydrogen leakage concentration prediction model architecture;

[0052] Figure 6 This is an example diagram of the prediction results of the hydrogen leakage concentration prediction model;

[0053] Figure 7 This is an example diagram of the hydrogen leakage concentration field simulation model architecture;

[0054] Figure 8 This is an example diagram of the deduction results of the hydrogen leakage concentration field deduction model;

[0055] Fig. 9 It is a structural principle block diagram of the system for rapidly simulating the consequences of hydrogen leakage in a hydrogen refueling station according to the present invention. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the purpose, technical solutions and implementation details of the present invention, so as to fully demonstrate the practicality of the present invention, the following will be further described in detail in conjunction with specific embodiments and drawings. Obviously, the embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0057] like Figure 1 As shown, a method for quickly simulating the consequences of hydrogen leakage at a hydrogen refueling station includes the following steps:

[0058] Step S01: Establish a full-scale three-dimensional virtual model of a hydrogen refueling station, specifically:

[0059] CFD simulation software was used to build a virtual three-dimensional model of a hydrogen refueling station where the method was applied to simulate the consequences of a hydrogen leakage accident. Figure 2 As shown, the hydrogen refueling station includes functional areas such as hydrogen refueling area, compression area, hydrogen storage area, hydrogen unloading area and hydrogen unloading area, as well as buildings such as machine room and office area, to ensure that the three-dimensional model is consistent with the actual location of the hydrogen refueling station and the area is the same.

[0060] That is, a 3D model was built using simulation software based on the geometric features of the hydrogen refueling station, including the material properties of the building and equipment, etc. This model is consistent with the details of the hydrogen refueling station.

[0061] Step S02: Arrange hydrogen detectors and concentration slices, specifically:

[0062] Use CFD simulation software to set hydrogen concentration monitoring points to represent hydrogen detectors in real scenarios, ensuring that the virtual monitoring points are in the same position and function as real hydrogen detectors. Use CFD simulation software to set slices at key plane positions of hydrogen refueling stations where this method is intended to be used to deduce accident consequences. The slices are used to record hydrogen concentration data at all grid points and output a cloud map of leaked hydrogen concentration distribution.

[0063] Step S03: Carry out a numerical simulation of hydrogen leakage, specifically:

[0064] Collect information on hydrogen leakage accidents, identify high-frequency leakage scenarios and leakage patterns; investigate meteorological factors in the area where the hydrogen refueling station is located, and determine data such as wind direction and wind speed in different seasons and months. Focusing on the leakage diffusion process under different hydrogen leakage scenarios at hydrogen refueling stations, use CFD numerical simulation software to carry out numerical simulations of hydrogen leakage at several hydrogen refueling stations, and obtain discrete hydrogen detector data and hydrogen concentration distribution data after hydrogen leakage at hydrogen refueling stations.

[0065] Step S04: Pre-process the simulation data and establish a leakage consequence database, specifically:

[0066] Data processing consists of two parts. The first part is to construct a sample set for the long short-term memory neural network, and the second part is to construct a sample set for the transposed convolutional neural network.

[0067] The hydrogen concentration at a certain point in the future output by the long short-term memory neural network is used as the input data of the transposed convolutional neural network, and the grayscale image of the hydrogen concentration at that point in time is used as the sample label of the model.

[0068] Step S05: Divide the data into a training set, a validation set, and a test set, specifically:

[0069] Before training the LSTM neural network and the transposed convolutional neural network models, the sample data is randomly divided into training set, validation set and test set. There is no fixed ratio for the division of training set, validation set and test set, usually they can account for 70%, 15% and 15%.

[0070] Step S06: Building a hydrogen concentration prediction model architecture based on the long short-term memory neural network, and training the long short-term memory neural network according to the hydrogen concentration field time series data to predict the hydrogen concentration of the hydrogen detector at a future time.

[0071] That is to say, the data used to train the hydrogen concentration prediction model architecture is the hydrogen concentration time series data at the location of the hydrogen detector in the consequence database.

[0072] The LSTM neural network is a time recurrent neural network used to capture long-term dependencies in hydrogen concentration time series data. It contains an input layer, a hidden layer, and an output layer. The main function of the input layer is to receive external data and pass it to the hidden layer. The hidden layer is responsible for processing the input data and updating its own information by combining the current input and the state of the previous time step, thereby capturing long-term dependencies in the time series. Over time, the hidden layer gradually accumulates and updates important historical information. Finally, the output layer receives the final state of the hidden layer and converts it into actual output.

[0073] In the present invention, the historical hydrogen concentration data recorded by the discrete hydrogen sensor is input into the trained long short-term memory neural network to output the hydrogen concentration at the future moment.

[0074] Furthermore, a real-time prediction model for hydrogen concentration was built using Python programming language based on the popular deep learning framework PyTorch. It includes building a model framework, designing long and short-term memory network layer parameters, and selecting appropriate training parameters for training. During the training process, the training is considered completed when the loss value is acceptable.

[0075] Figure 5 The following diagram shows an example of the hydrogen leakage concentration prediction model architecture. Figure 5As shown in the figure, the input layer of the model framework has 23 units for inputting detector data, the hidden layer has 2 layers of long short-term memory neural network, and the output layer consists of 1 fully connected layer with 23 units.

[0076] Two layers of LSTM networks have 100 and 50 neurons respectively. A random dropout layer with a dropout rate of 0.10 is added at the end of the first and second LSTM layers to reduce the risk of overfitting. The activation function and optimization function use "Tanh" and "Adam" respectively.

[0077] Appropriate training parameters were selected, specifically, mean square error (MSE) and coefficient of determination (R2) were used as loss function and indicator to evaluate the performance of the model after each iteration; the training batch size was set to 32; the initial learning rate and training cycle were set to 0.001 and 1000. Figure 6 The following figure shows the prediction results of the hydrogen leakage concentration prediction model: Figure 6 As shown in the figure, the red line represents the hydrogen concentration time series data recorded by the detector, and the blue line represents the hydrogen concentration time series data predicted by the model. In other words, the blue line is the model prediction result, and the red line represents the data recorded by the sensor. Because the model can make predictions based on the data recorded by the sensor, the time point of the blue line is ahead of the red line. The interval between the two represents the lead time. The lead time in the figure is 2.5 seconds.

[0078] Step S07: Building a deduction model architecture based on the transposed convolutional neural network, and training a rapid deduction model for the consequences of hydrogen leakage.

[0079] According to the hydrogen concentration time series data of the hydrogen detector at future moments (the output of the hydrogen concentration prediction model architecture) and the hydrogen concentration field time series data (slice output leakage hydrogen concentration distribution cloud map), the transposed convolutional neural network model is trained to deduce the hydrogen concentration field at future moments.

[0080] The transposed convolutional neural network is an upsampling convolutional neural network for gradually extrapolating discrete hydrogen concentration monitoring point data to a two-dimensional hydrogen concentration field distribution. The transposed convolutional neural network architecture consists of multiple levels, including an input layer, a fully connected layer, a reshaping layer, a transposed convolutional layer, and an output layer. Each layer plays an important role in the data processing process, gradually realizing the deduction from low resolution to high resolution. First, the input layer receives hydrogen concentration data from the sensor, which is a one-dimensional vector. Then, the data is dimensionally expanded through a fully connected layer to increase its information content, and then the one-dimensional data is converted into a low-resolution two-dimensional form through a reshaping layer. Through this reshaping process, the data can better adapt to subsequent convolution processing. Under the action of the transposed convolutional layer, the model gradually improves the spatial resolution, and restores a higher resolution hydrogen concentration image or scene through multiple upsampling. In this process, the transposed convolutional layer not only helps to restore spatial information, but also processes the nonlinear features in the data through the activation function layer, further enhancing the model's ability to fit complex patterns. Finally, after multiple layers of convolution and upsampling, the network generates the required high-resolution output image or signal to complete the deduction and consequence prediction of hydrogen concentration. By inputting the hydrogen concentration data at the future moment output by the long short-term memory neural network model, the two-dimensional concentration field distribution of hydrogen is output, and the consequences of hydrogen leakage at the hydrogen filling station are quickly deduced.

[0081] Furthermore, the hydrogen concentration field deduction model was built using Python programming language based on the popular deep learning framework PyTorch. It includes building the model framework, designing the transposed convolutional neural network layer parameters, and selecting appropriate training parameters for training. During the training process, the training is considered completed when the loss value is acceptable.

[0082] Figure 7 The following diagram shows an example of the hydrogen leakage concentration prediction model architecture. It contains 1 input layer, 4 fully connected layers, 1 reshape layer, 7 TCNN layers and 1 CNN layer. In addition, a random inactivation layer (dropout) is added before the reshape layer.

[0083] like Figure 7As shown in the figure, the input layer of the hydrogen leakage concentration field deduction model architecture consists of a 1×23 dimensional vector, corresponding to the hydrogen concentration data at a certain future moment output by the hydrogen leakage concentration prediction model. The input data is processed by four fully connected layers (Dense layer) to increase the dimension of the input parameters. At the same time, a random inactivation layer (Dropout layer) is used after the first fully connected layer to prevent overfitting during training. The reshape layer converts the one-dimensional data into a three-dimensional tensor. Then, the resolution of the hydrogen concentration data is improved through a series of transposed convolutional neural network layers (TCNN). The output layer is a CNN layer, which inputs a grayscale image of 50×110 pixels and has 1 image channel, corresponding to the hydrogen concentration.

[0084] The TCNN layer contains many parameters, which are used to control the size of the output image of the TCNN layer according to formula (4).

[0085] Output size = (input size - 1) × S - 2 × Pad + K (4)

[0086] Among them, S is the stride, K is the convolution kernel size, and Pad is the padding size. In addition, Channel controls the channel of the output image. The design of the parameters depends on the technical requirements of the hydrogen refueling station where the method is intended to be used to deduce the accident consequences.

[0087] Appropriate training parameters were selected for training. Specifically, mean square error (MSE) and coefficient of determination (R2) were used as loss functions and indicators to evaluate the performance of the model after each iteration. The training batch size was set to 32. The initial learning rate and training cycle were set to 0.001 and 1000.

[0088] Figure 8 An example diagram of the deduction results of the hydrogen leakage concentration field deduction model is shown, which includes the model deduction results at some time points. It is a grayscale diagram of the hydrogen concentration distribution at future moments output by the model.

[0089] Step S08: deploying the trained model to the data processing module in the rapid simulation system of the consequences of hydrogen leakage at the hydrogen filling station, specifically:

[0090] The process of embedding the trained model into the data processing module includes the following steps: First, export the trained PyTorch model in .pt or .pth format and upload it to the server where the data processing module is located. Then, configure the PyTorch runtime environment in the data processing module and ensure that the dependent libraries are correctly installed. When loading the model, use the torch.load() or torch.jit.load() method to import the model into memory and deploy it as an inference service.

[0091] In summary, the advance deduction model established in the method of the present invention includes two sub-models. The first is a hydrogen concentration prediction model based on a long short-term memory network. This model uses data from discrete sensors to predict the hydrogen concentration recorded by the sensor at a certain moment in the future (such as 2.5 seconds in advance). The predicted hydrogen concentration is discrete. The second is a deduction model architecture based on a transposed convolutional neural network. This deduction model uses the output of the first model as input, and then deduces a two-dimensional hydrogen concentration field. That is, the first model is responsible for advance prediction, and the second model is responsible for deducing from discrete sensor data to a two-dimensional concentration field.

[0092] In the above method: for the long short-term memory neural network, the linear interpolation method is first used to generate data records with uniform time intervals, and then the sliding window method is used to reconstruct the hydrogen concentration detector time series data to meet the sample structure requirements for the long short-term memory neural network.

[0093] The following example will specifically illustrate how to establish a training sample for a long short-term memory neural network. This example is called Example 1, which is simulated using FLACS simulation software. The hydrogen leak lasts for 20 seconds and 23 hydrogen detectors are set.

[0094] Step 1: The hydrogen concentration time series data collected from the simulation software refers to the raw data. For each simulation, several hydrogen detectors recorded the hydrogen concentration from 0 to 20 seconds and stored it in a data table. The dimension of the raw data (rows × columns) = 10,294 × 23. Each row reflects the temperature measurement at a certain time point, which is composed of measurement data from 23 detectors. Each column represents the hydrogen concentration time series data recorded by a detector, which is composed of hydrogen concentration data within 20s. Because the time interval of the data recorded by FLACS is not fixed, that is, the detector does not record data changes at a fixed sampling frequency. Before further processing the data, it is necessary to reconstruct the data into sequence information with uniform time intervals. The specific approach is: first, delete the repeated time items; second, assume that the change between two data points is linear, and generate data with equal time intervals based on linear interpolation. This provides a more consistent and coherent data set for subsequent data analysis and modeling. The time interval after data reshaping is 0.05 seconds. The data dimension is 401 × 23, representing an array of 401 time points and 23 detectors.

[0095] Step 2: Reconstruct the data structure to match the training mechanism of the LSTM neural network. The sliding window method is a common method for generating a sample database for the LSTM neural network. Figure 3 The schematic diagram of the sliding window sampling principle is shown. As shown in the figure, the window size represents the number of time points of each sample, that is, the length of the window. Figure 3In the example, the window width is set to 50, that is, each input sample contains hydrogen concentration data of 50 time points (duration 2.5 seconds). In addition, the sliding window method also includes a sliding step. The sliding step represents the step size of a single slide on the time series, which determines the time interval and overlap between samples. In Example 1, the sliding step is set to 1. Therefore, the first three input samples are data of 0-2.45 seconds, 0.05–2.50 seconds, and 0.1–2.55 seconds, respectively. The input data Xn of the nth sliding window can be written as:

[0096]

[0097] To make the training objective clear, each sample is assigned a corresponding output label, which is the hydrogen concentration prediction at a certain point in the future. The lead time parameter determines the length of the prediction time. In Example 1, the lead time = 2.5 seconds, so the labels of the first three samples are the concentration data at 4.95 seconds, 5.0 seconds, and 5.05 seconds. The output of the nth time window is labeled Yn:

[0098] Y n =[T 1,4.9+0.05n ,T 2,4.9+0.05n ,T 3,4.9+0.05n ,…T 23,4.9+0.05n ] (2)

[0099] Step 3: Normalize the concentration data using the normalization function of formula (3) to ensure that all measurements are presented in a similar scale.

[0100]

[0101] In this way, the sample database of the long short-term memory neural network is constructed.

[0102] For the transposed convolutional neural network, Figure 4 This is an example of reducing the RGB cloud image and converting it to a grayscale cloud image. Figure 4 As shown in Figure 1, its data preprocessing part includes operations such as reducing the scale of the cloud image and converting the RGB cloud image to a grayscale image. Take Example 1 as an example to illustrate, specifically:

[0103] Step 1: Assuming that the original cloud image size is 698×1535 pixels, the cloud image scale is reduced to 50×110 pixels (resolution = 1m) without changing the image aspect ratio.

[0104] Step 2: The original RGB image has three channels. The RGB image is grayscaled so that it has only one channel.

[0105] In this way, the sample database of the transposed convolutional neural network is constructed.

[0106] The present invention also provides a rapid deduction system for the consequences of hydrogen leakage in a hydrogen refueling station, such as Fig. 9 As shown, the system mainly includes: a perception module 1, a communication module 2, a data storage module 3, a data processing module 4 and a leakage consequence display module 5.

[0107] The sensing module 1 is mainly composed of several hydrogen detectors in the hydrogen filling station, equipped with high-sensitivity sensors to monitor the hydrogen concentration in real time. The function of this module is to quickly detect and trigger an alarm when hydrogen leaks, and transmit the monitoring data to the communication module in real time.

[0108] The communication module 2 includes wireless and wired transmission units, which are responsible for quickly and accurately transmitting the hydrogen concentration data detected by the sensing module to the data storage module. This module ensures the security and integrity of the data during transmission and prevents data loss.

[0109] The data storage module 3 is composed of a server, a storage device and a network device, and is responsible for receiving, storing and managing hydrogen concentration data. The server controls the real-time storage and transmission of data, the storage device is responsible for the persistent preservation of data, and the network device ensures the smooth flow of data.

[0110] The data processing module 4 is embedded with a submodule for real-time prediction of hydrogen leakage concentration and a submodule for rapid deduction of the consequences of hydrogen leakage. In other words, the data processing module 4 includes two submodules. These two submodules are models built based on deep learning algorithms, namely, a submodule for real-time prediction of hydrogen leakage concentration, corresponding to a hydrogen concentration prediction model architecture built based on a long short-term memory neural network; and a submodule for rapid deduction of the consequences of hydrogen leakage, corresponding to a deduction model architecture built based on a transposed convolutional neural network.

[0111] By writing an inference script and calling the model for real-time inference, the required prediction results are generated. Specifically, the real-time prediction submodule of hydrogen leakage concentration calls the pre-trained long short-term memory neural network to calculate the trend of hydrogen concentration changes at a specific moment in the future based on the time series data of the hydrogen detector recorded in the data storage module. Subsequently, the hydrogen leakage consequence rapid deduction submodule receives the data output by the prediction submodule, calls the pre-trained transposed convolutional neural network to further deduce and output the hydrogen concentration distribution at that moment. After the deduction is completed, the results are passed to the leakage consequence display module through the internal data interface.

[0112] The leakage consequence display module is mainly composed of hardware and software, among which the LCD display is used to intuitively display the distribution of leaked hydrogen in the hydrogen refueling station; the workstation is responsible for running the display system software to ensure that the results are displayed in a low-latency manner; the keyboard and mouse are used for operation and view adjustment. The software part includes a graphical user interface, which provides an operation entry and supports view switching and data display; the data visualization tool is used to draw the distribution map of the hydrogen concentration field and time series animation to help intuitively display the diffusion trend. This module receives the hydrogen concentration field deduction results output by the data processing module in real time, and presents them through a variety of visualization methods to provide clear risk information support for operators.

[0113] The method for rapidly deducing the consequences of hydrogen leakage at a hydrogen refueling station of the present invention uses the PyTorch deep learning framework, and by calling its built-in long short-term memory neural network model and transposed convolutional neural network model, it can realize rapid deduction of the consequences of hydrogen leakage at a hydrogen refueling station. Moreover, the deduction results are highly accurate and reliable, thus providing guidance for emergency measures and personnel evacuation.

[0114] Parts not described in the above implementation manner can be implemented by adopting or drawing on existing technologies.

[0115] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above implementation modes. Any changes, modifications, additions or substitutions made by those skilled in the art within the essential scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A method for rapidly simulating the consequences of hydrogen leakage at a hydrogen refueling station, characterized in that The following steps are involved: Step S1, establishing a hydrogen leakage-diffusion numerical model of a hydrogen refueling station; Step S2, arranging a plurality of hydrogen detectors in the hydrogen leakage-diffusion numerical model of the hydrogen refueling station to monitor the hydrogen concentration time series data at the location of the hydrogen detectors under the condition of hydrogen leakage and diffusion; setting the hydrogen concentration slice to obtain the hydrogen concentration field time series data under the condition of hydrogen leakage and diffusion; Step S3, simulating different leakage scenarios, leakage modes and influencing factors, and obtaining hydrogen leakage consequence data in the hydrogen refueling station according to the hydrogen leakage-diffusion numerical model of the hydrogen refueling station and the hydrogen detectors arranged in the model; Step S4, pre-processing the hydrogen leakage consequence data in the hydrogen refueling station obtained in step S3, and establishing a hydrogen leakage consequence database for the hydrogen refueling station; Step S5, dividing the data in the hydrogen refueling station hydrogen leakage consequence database into a training set, a validation set and a test set for prediction model training; Step S6: construct a hydrogen concentration prediction model framework based on the long short-term memory neural network, and train the long short-term memory neural network according to the data in the hydrogen leakage consequence database of the hydrogen filling station; and predict the hydrogen concentration time series data of the hydrogen detector at the future time through the constructed hydrogen concentration prediction model framework; Step S7, building a deduction model framework based on the transposed convolutional neural network, training the transposed convolutional neural network according to the hydrogen concentration time series data and hydrogen concentration field time series data of the hydrogen detector at the future moment; and deducing the hydrogen concentration field at the future moment through the built deduction model framework; Step S8: When hydrogen leaks at the hydrogen refueling station, the hydrogen concentration prediction model framework constructed in step S6 is input according to the data measured by the hydrogen detector, and then combined with the deduction model framework constructed in step S7 to quickly deduce the consequences of hydrogen leakage at the hydrogen refueling station.

2. A method for rapid deduction of the consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S1: according to the physical characteristics of the hydrogen refueling station, a full-scale hydrogen leakage-diffusion numerical model of the hydrogen refueling station is established using CFD simulation software.

3. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 2, characterized in that: In step S2: the number and coordinates of hydrogen detectors are set to be consistent with the actual scene of the hydrogen refueling station; the hydrogen concentration slicing is a function in the CFD simulation software, which can output the two-dimensional distribution results of the hydrogen concentration in a specific plane for training the transposed convolutional neural network.

4. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S3: the hydrogen leakage consequence data in the hydrogen filling station includes the hydrogen concentration time series data at the location of the hydrogen detector and the hydrogen concentration field time series data at the slice location.

5. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S3: the leakage scenario refers to the potential hydrogen leakage area in the hydrogen refueling station, including the transmission pipeline, hydrogen storage area, compression area, and hydrogenation area; the leakage mode refers to the amount of hydrogen leakage, including micro-leakage, small flow leakage, large flow leakage and transient large-scale leakage; the influencing factors refer to the environmental factors that affect the distribution of leakage hydrogen concentration, including wind speed and wind direction indicators.

6. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S6: the long short-term memory neural network is a time recurrent neural network, which is used to capture the long-term dependency in the hydrogen concentration time series data, and output the hydrogen concentration time series data at a future moment by inputting the historical hydrogen concentration data.

7. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S6: the long short-term memory neural network includes an input layer for receiving external data and passing it to the hidden layer; The hidden layer is responsible for processing the input data and updating its own information by combining the current input and the state of the previous time step, thereby capturing the long-term dependencies in the time series. As time goes by, the hidden layer gradually accumulates and updates important historical information. And the output layer is used to receive the final state of the hidden layer and transform it into the actual output.

8. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S7: the transposed convolutional neural network is an upsampling convolutional neural network, which is used to gradually deduce the discrete hydrogen concentration monitoring point data to the two-dimensional hydrogen concentration field distribution; by inputting the future hydrogen concentration time series data output by the long short-term memory neural network, the two-dimensional hydrogen concentration field distribution is output, and the consequences of hydrogen leakage at the hydrogen refueling station are quickly deduced.

9. A method for rapid deduction of consequences of hydrogen leakage at a hydrogen refueling station according to claim 1, characterized in that: In step S7: the transposed convolutional neural network includes an input layer for receiving the hydrogen concentration time series data at a future moment outputted from the long short-term memory neural network, wherein the data is a one-dimensional vector; The fully connected layer is used to expand the dimension and increase the amount of information; A reshape layer, which converts one-dimensional data into a low-resolution two-dimensional form; Through the reshaping process, the data can better adapt to the subsequent convolution processing; The transposed convolution layer is used to gradually improve the spatial resolution of the model. Through multiple upsampling, a higher resolution hydrogen concentration image or scene is restored. The transposed convolution layer not only helps to restore spatial information, but also processes the nonlinear characteristics in the data through the activation function layer, further enhancing the model's ability to fit complex patterns. Finally, after multiple layers of convolution and upsampling, the required high-resolution output image or signal is generated. and an output layer for outputting a high-resolution output image or signal.

10. A rapid simulation system for the consequences of hydrogen leakage at a hydrogen refueling station, characterized in that: It includes a perception module, a communication module, a data storage module, a data processing module and a leakage consequence display module; the perception module is connected to the data storage module through the communication module, the data storage module is connected to the data processing module, and the data processing module is connected to the leakage consequence display module; The sensing module includes a plurality of hydrogen detectors, which are used to collect the concentration of hydrogen leaked from the hydrogen filling station in real time; The communication module transmits the real-time hydrogen concentration data monitored by the sensing module to the data storage module; The data storage module is used to store the real-time hydrogen concentration data collected by the hydrogen detector, and to store the model prediction and deduction results; The data processing module includes a trained hydrogen leakage concentration real-time prediction submodule and a hydrogen leakage consequence rapid deduction submodule; the hydrogen leakage concentration real-time prediction submodule processes the hydrogen concentration time series data of the real-time hydrogen detector position recorded in the data storage module, and outputs the hydrogen concentration at a certain moment in the future; the hydrogen leakage consequence rapid deduction submodule processes the hydrogen concentration at a certain moment in the future predicted by the hydrogen leakage concentration real-time prediction submodule, and outputs the hydrogen concentration field at a certain moment in the future; The leakage consequence display module receives the hydrogen concentration field information output by the data processing module, and displays the distribution of hydrogen concentration in the hydrogen filling station.

Images