Generative adversarial network-based underground cavern harmful gas diffusion rapid prediction method

Abstract

The invention discloses an underground cavern harmful gas diffusion rapid prediction method based on a generative adversarial network, and the method comprises the steps: building a three-dimensional geometric model of an underground cavern, and carrying out the numerical simulation of pollutant diffusion through a computational fluid dynamics (CFD) method, and obtaining a simulation result; key parameter data are extracted, and a generative adversarial network model is improved based on a Pix2Pix architecture; in improved generative adversarial network model training, a key data set is used to pre-train a generator, and then the key data set and a discriminator are combined to train and optimize the GAN; optimizing a loss function, and adding perception loss and feature matching loss; and utilizing the trained improved generative adversarial network model to predict pollutant concentration distribution so as to obtain a prediction result of a high-resolution two-dimensional image of the pollutant concentration distribution. According to the method, the CFD numerical simulation and the machine learning algorithm are integrated, the pollutant diffusion prediction process is optimized, and rapid and accurate prediction of pollutant diffusion in the underground powerhouse construction period is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of underground plant construction and environmental engineering technology, and in particular to a method for quickly predicting the diffusion of harmful gases in underground caverns based on a generative adversarial network. Background Art

[0002] In the study of pollutant diffusion in underground caverns, existing technologies mainly use two methods: computational fluid dynamics (CFD) simulation and machine learning algorithms. (1) CFD simulation can accurately predict the flow field and pollutant concentration distribution in underground caverns through numerical simulation (such as Reynolds average method RANS and large eddy simulation LES), and is suitable for the analysis of complex geometric structures and boundary conditions. The CFD method is based on the Navier-Stokes equation and can capture subtle changes in fluid flow. Especially in a closed and complex environment such as an underground cavern, it can accurately simulate the dynamic process of air flow, temperature distribution and pollutant diffusion. The advantage of CFD simulation lies in its high precision and reliability. It can be verified by experimental data (such as wind tunnel experiments or field measurements) to ensure the accuracy of its results. CFD simulation can simulate the impact of different ventilation schemes on pollutant diffusion and help engineers select the optimal design scheme. In addition, CFD simulation can also combine multi-physics field coupling technology to simultaneously analyze the interaction between temperature field, pressure field and pollutant concentration field, providing a comprehensive solution for complex scenarios. (2) Machine learning algorithms can quickly predict the pollutant diffusion process by learning a large amount of CFD or measured data, significantly reducing computing cost and time. Machine learning algorithms, driven by data, can capture complex nonlinear relationships and achieve near-real-time predictions in specific scenarios. For example, convolutional neural networks (CNNs) can rapidly generate pollutant concentration maps by learning the spatial characteristics of flow fields. Generative adversarial networks (GANs), through adversarial training of generators and discriminators, can produce highly accurate flow field predictions. These algorithms not only process high-dimensional data but also achieve high-precision predictions on coarse grids through super-resolution techniques, significantly improving computational efficiency. The application of machine learning algorithms in the study of pollutant diffusion in underground caverns has provided a highly effective technical approach for engineering practice, demonstrating significant advantages in scenarios requiring rapid evaluation of multiple design options or real-time monitoring of pollutant diffusion. In recent years, some studies have also attempted to combine CFD simulation with machine learning algorithms, further improving prediction efficiency and accuracy by generating high-precision data for model training or optimizing the computational process. For example, using high-fidelity data generated by CFD simulations to train deep learning models can significantly improve the prediction accuracy of machine learning algorithms. Furthermore, machine learning algorithms can be used to optimize meshing and boundary condition settings in CFD simulations, reducing computational time and resource consumption. This hybrid method not only improves computational efficiency, but also ensures that the prediction results conform to physical laws, providing a new technical path for the study of pollutant diffusion in underground caverns.

[0003] Although CFD simulation and machine learning algorithms have made significant progress in the study of pollutant diffusion in underground caverns, the following problems still exist: (1) Although CFD simulation has high accuracy, it is computationally intensive and time-consuming, especially when dealing with complex geometric structures and large-scale computational domains. The demand for computing resources is huge and it is difficult to meet the needs of real-time feedback. For example, in the design optimization process of underground caverns, engineers usually need to evaluate multiple ventilation schemes or pollutant diffusion scenarios, and each CFD simulation calculation may take hours or even days, which seriously limits its application in rapid iterative design. (2) The setting and verification process of CFD models is complicated and requires professional experience and knowledge, which to a certain extent limits its wide application. For example, meshing, boundary condition setting and turbulence model selection in CFD simulation require engineers to have a deep fluid mechanics background and rich practical experience, otherwise the calculation results may be inaccurate or unreliable. In addition, CFD simulation often requires fine meshing when dealing with complex geometric structures, which further increases the demand for computing resources and computing time. (3) Although machine learning algorithms have fast calculation speed, their prediction accuracy depends on the quality and quantity of training data, and the lack of physical constraints may result in prediction results that do not conform to the actual fluid mechanics laws. For example, machine learning models may not be able to accurately capture the complex flow phenomena in underground caverns caused by temperature gradients or pressure changes, or may not be able to handle boundary conditions or geometric structures that are not included in the training data. At the same time, the training of machine learning models requires a large amount of high-quality data, and obtaining this data usually takes a lot of time and resources. For example, in the study of pollutant diffusion in underground caverns, it may take weeks or even months to obtain high-precision CFD simulation data or field measurement data, which to some extent limits the application scope of machine learning algorithms. (4) Machine learning models have poor interpretability and are difficult to directly apply to complex scenarios in engineering practice. The prediction results provided by machine learning models may lack clear physical meaning and are difficult for engineers to understand and apply. (5) Although some studies have attempted to combine CFD simulation with machine learning algorithms to make up for the shortcomings of a single method, this hybrid approach still faces some challenges. For example, how to ensure that the machine learning model can fully learn the physical laws in CFD simulation during training, and how to balance computational efficiency and prediction accuracy in practical applications, are still key issues that need to be solved. (6) With the continuous expansion of underground caverns and the increase in design complexity, the limitations of existing technologies in dealing with large-scale computational domains or complex geometric structures are becoming increasingly prominent. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] In response to the problems in the existing technology of pollutant diffusion prediction during the construction period of underground power plants, such as long time consumption, high computational cost and difficulty in achieving rapid prediction, the present invention provides a rapid prediction method for pollutant diffusion during the construction period of underground power plants based on a generative adversarial network. It combines computational fluid dynamics (CFD) numerical simulation with deep learning algorithms to obtain near-real-time predictions of pollutant concentration distribution through an integrated multi-factor data set.

[0006] Another object of the present invention is to propose a rapid prediction system for the diffusion of harmful gases in underground caverns based on a generative adversarial network.

[0007] To achieve the above objectives, the present invention proposes a method for rapidly predicting the diffusion of harmful gases in underground caverns based on a generative adversarial network, comprising:

[0008] A three-dimensional geometric model of the underground cavern was established, and numerical simulation of pollutant diffusion was performed using the computational fluid dynamics (CFD) method to obtain simulation results.

[0009] Extract key parameter data from CFD-POST simulation results, visualize them through cloud charts, and convert the visualized 3D data into 2D images for processing to obtain key data sets;

[0010] An improved generative adversarial network model is based on the Pix2Pix architecture. In the improved generative adversarial network model, the generator adopts a U-Net structure, the discriminator uses a PatchGAN architecture, a time channel and a wind speed channel are added to the network input layer, and an attention mechanism improvement module is introduced.

[0011] In the training of the improved generative adversarial network model, the generator is pre-trained using key datasets, and then jointly trained with the discriminator to optimize the GAN. The loss function is optimized, incorporating perceptual loss and feature matching loss, and dynamic learning rate adjustment and early stopping strategies are used to prevent overfitting.

[0012] The trained improved generative adversarial network model is used to predict the pollutant concentration distribution to obtain the prediction results of a high-resolution two-dimensional image of the pollutant concentration distribution.

[0013] The method for rapid prediction of harmful gas diffusion in underground caverns based on a generative adversarial network according to an embodiment of the present invention may also have the following additional technical features:

[0014] In one embodiment of the present invention, key parameter data is extracted from CFD-POST simulation results, visualized in a cloud diagram format, and the visualized three-dimensional data is converted into a two-dimensional image for processing to obtain a key data set, including:

[0015] Extract key parameter data from CFD-POST simulation results and visualize the key parameter data through cloud charts to obtain visualized three-dimensional data; wherein the key parameter data includes pollutant diffusion concentration information;

[0016] The generated three-dimensional visual data is processed so that the data format is adapted to the Pix2Pix model input, and key data sets are constructed for different prediction requirements. The key data sets are preprocessed and divided into training sets, validation sets, and test sets; wherein the key data sets include optimization mode data sets and monitoring mode data sets.

[0017] In one embodiment of the present invention, key parameter data is extracted from CFD-POST simulation results, and the key parameter data is visualized in a cloud diagram to obtain visualized three-dimensional data, including:

[0018] Use CFD-POST as a post-processing tool for Fluent calculation results to extract pollutant diffusion concentration data at a preset pedestrian level;

[0019] A cloud map is generated based on the extracted pollutant diffusion concentration data, and the CO concentration scale is unified when displayed to obtain visual three-dimensional data.

[0020] In one embodiment of the present invention, the generated three-dimensional visualized data is processed so that the data format is adapted to the Pix2Pix model input, and key data sets are constructed for different prediction requirements. The key data sets are preprocessed and divided into training sets, validation sets, and test sets, including:

[0021] The generated cloud image is resized to 1200*600 to meet the input requirements of the Pix2Pix model.

[0022] Design an optimized model dataset: Map the model height to two dimensions and display it in grayscale. Mark the smoke throwing area and the air duct outlet on the map. Combine this with simulated pollutant distribution cloud maps at pedestrian level at different ventilation times. Process the image pixel size to obtain a dataset.

[0023] Design the monitoring model data set: when labeling, the pollutant concentration is also marked to obtain the monitoring model data set;

[0024] The optimization mode dataset and the monitoring mode dataset are normalized and preprocessed, and divided into training set, validation set and test set.

[0025] In one embodiment of the present invention, a generative adversarial network model is improved based on the Pix2Pix architecture. In the improved generative adversarial network model, the generator adopts a U-Net structure, the discriminator uses a PatchGAN architecture, a time channel and a wind speed channel are added to the network input layer, and an attention mechanism improvement module is introduced, including:

[0026] The Generative Adversarial Network (GAN) is improved based on the Pix2Pix framework, and the objective function is optimized through the minimax game. The generator adopts the U-Net structure and uses the encoder-decoder architecture to achieve feature extraction and high-resolution image generation; the discriminator adopts the PatchGAN structure:

[0027]

[0028] Where G is the generator, D is the discriminator, G(Z) is the generated data, x~pdata is the sampled data in the real data, x~pz is the sampled noise in the random noise distribution, D(x) is the probability that the discriminator judges that the real data x is true, and D(G(z)) is the probability that the discriminator judges that the generated data G(z) is true;

[0029] Add additional time channel and wind speed channel to the input layer;

[0030] Introducing an attention mechanism module into the generator’s jump connection to enhance the model’s ability to focus on key areas:

[0031] α=σ(W a [F enc ,F dec ]+b a )

[0032] Where Fenc and Fdec are the feature maps of the encoder and decoder respectively, σ is the Sigmoid function, and α is the attention weight;

[0033] Optimize network depth and width according to data complexity to avoid overfitting or underfitting; compare the effects of different numbers of layers and convolution kernels through experiments to select the optimal model structure.

[0034] In one embodiment of the present invention, in the training of the improved generative adversarial network model, the generator is pre-trained using a key data set, and then the GAN is optimized by joint training with the discriminator; the loss function is optimized, perceptual loss and feature matching loss are added, and dynamic learning rate adjustment and early stopping strategies are used to prevent overfitting, including:

[0035] Use key data sets to pre-train the generator separately to initially learn the mapping relationship between CFD simulation data. Then jointly train the generator and discriminator to optimize the generation effect through adversarial learning.

[0036] Based on the traditional adversarial loss GAN Loss, we introduce the perceptual loss and feature matching loss;

[0037] Adopt an adaptive learning rate adjustment strategy and monitor the PSNR peak signal-to-noise ratio and SSIM structural similarity indicators on the validation set;

[0038] Design a visual monitoring module to record the changes in loss function and the quality of generated samples in real time to adjust hyperparameters and optimize training strategies.

[0039] In one embodiment of the present invention, the generator is pre-trained separately to initially learn the mapping relationship of CFD simulation data, and then the generator and discriminator are jointly trained to optimize the generation effect through adversarial learning, including:

[0040] Perform generator pre-training and optimize the generator using only L1 loss. The objective function formula is as follows:

[0041] L pretrain =λ L1 PG(z)-yP1

[0042] Where y is the actual CFD simulation result, λL1=100 is the weight coefficient;

[0043] The discriminator is introduced to conduct joint adversarial training. The total loss function formula is as follows:

[0044] L total =L GAN +λ L L1+λ FM L FM +λ perc L perc

[0045]

[0046] Where D(i) is the feature map of the ii-th layer of the discriminator, and φ is the intermediate layer feature extractor of the pre-trained VGG19 network.

[0047] In one embodiment of the present invention, the pollutant concentration distribution is predicted using a trained improved generative adversarial network model to obtain a prediction result of a high-resolution two-dimensional image of the pollutant concentration distribution, including:

[0048] The error index was calculated on the test set and compared with the traditional CFD simulation results to verify the model's advantages in prediction speed and accuracy:

[0049]

[0050] Build an interactive visualization platform that supports user input of environmental parameters, outputs pollutant diffusion prediction results in real time, and dynamically renders concentration distribution maps;

[0051] Provides multiple visualization methods such as contour maps, heat maps, and dynamic evolution animations to intuitively present pollutant diffusion trends;

[0052] The comparison function between CFD simulation and GAN prediction is integrated into the GUI, supporting error distribution visualization.

[0053] To achieve the above objectives, the present invention further proposes a rapid prediction system for the diffusion of harmful gases in underground caverns based on a generative adversarial network, comprising:

[0054] The simulation module is used to establish a three-dimensional geometric model of the underground cavern and perform numerical simulation of pollutant diffusion using the computational fluid dynamics (CFD) method to obtain simulation results.

[0055] The data set construction module is used to extract key parameter data from CFD-POST simulation results, visualize them through cloud graphs, and convert the visualized three-dimensional data into two-dimensional images for processing to obtain key data sets;

[0056] A generative adversarial network building module is used to improve the generative adversarial network model based on the Pix2Pix architecture. In the improved generative adversarial network model, the generator adopts a U-Net structure, the discriminator uses a PatchGAN architecture, and a time channel and wind speed channel are added to the network input layer. At the same time, an attention mechanism improvement module is introduced.

[0057] The model training optimization module is used to pre-train the generator using key datasets in the training of the improved generative adversarial network model, and then jointly train and optimize the GAN with the discriminator. The module also optimizes the loss function, incorporating perceptual loss and feature matching loss, and adopts dynamic learning rate adjustment and early stopping strategies to prevent overfitting.

[0058] The data prediction module is used to predict the pollutant concentration distribution using the trained improved generative adversarial network model to obtain the prediction results of a high-resolution two-dimensional image of the pollutant concentration distribution.

[0059] The method and system for rapid prediction of harmful gas diffusion in underground caverns based on a generative adversarial network in an embodiment of the present invention optimizes the process of pollutant diffusion prediction by integrating CFD numerical simulation and machine learning algorithms, thereby achieving rapid and accurate prediction of pollutant diffusion during underground plant construction.

[0060] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0062] Figure 1 is a flow chart of a method for rapidly predicting the diffusion of harmful gases in underground caverns based on a generative adversarial network according to an embodiment of the present invention;

[0063] Figure 2 is a logic diagram of a dual-mode data set construction work according to an embodiment of the present invention;

[0064] Figure 3 This is a flowchart of building a rapid pollutant prediction model based on Pix2pixGAN according to an embodiment of the present invention;

[0065] Figure 4 2 is a structural diagram of a rapid prediction system for the diffusion of harmful gases in underground caverns based on a generative adversarial network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0067] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0068] The following describes a method and system for quickly predicting the diffusion of harmful gases in underground caverns based on a generative adversarial network according to an embodiment of the present invention with reference to the accompanying drawings.

[0069] Figure 1 is a flow chart of a method for rapidly predicting the diffusion of harmful gases in underground caverns based on a generative adversarial network according to an embodiment of the present invention. Figure 1 As shown, the method includes but is not limited to the following steps:

[0070] S1, establish a three-dimensional geometric model of the underground cavern, and use the computational fluid dynamics (CFD) method to perform numerical simulation of pollutant diffusion to obtain simulation results.

[0071] Specifically, a three-dimensional geometric model of the underground cavern was first established, and computational fluid dynamics (CFD) methods were used to numerically simulate pollutant diffusion. To ensure the reliability of the simulation results, mesh independence verification was performed, and the optimal meshing scheme was determined by comparing different mesh densities. Simultaneously, model validity verification was conducted, comparing the simulation results with experimental data or theoretical solutions to verify the model's accuracy. This step provides a reliable training data foundation for subsequent deep learning.

[0072] It's understandable that a multi-sensor integrated device primarily consists of sensors, a hub, a Data Transfer Unit (DTU) with an embedded 4G wireless data transmission terminal module, and cables. The sensors convert ambient temperature, humidity, pollutant concentration, and oxygen concentration into electrical signals; the hub houses and manages numerous connection lines and provides power to the various sensors; and the DTU transmits data via a 4G wireless network to the corresponding IP address and port number.

[0073] S2, extracts key parameter data from CFD-POST simulation results, visualizes them through cloud graphs, and converts the visualized three-dimensional data into two-dimensional images for processing to obtain key data sets.

[0074] Specifically, key parameter data, including information such as pollutant diffusion concentration, are extracted from CFD-POST simulation results. Cloud charts are used for visualization, and three-dimensional data are converted into two-dimensional images for processing. In response to different prediction needs, two modes of data sets are constructed: one is a design optimization mode data set, and the other is a monitoring mode data set. The data set needs to be pre-processed by size normalization, CO concentration scale unification, etc., and divided into training set, validation set, and test set. Figure 2 As shown in the figure, a working logic diagram is constructed for the dual-mode dataset. S2 includes the following steps:

[0075] S201: Extract key parameter data from CFD-POST simulation results, including information such as pollutant diffusion concentration. Visualize them using cloud charts to facilitate subsequent data processing and analysis.

[0076] Furthermore, the specific operation of step S201 includes the following steps:

[0077] S2011: Use CFD-POST as a post-processing tool for Fluent calculation results to extract pollutant diffusion concentration data at pedestrian level (about 1.6m).

[0078] S2012: Generate cloud maps for the extracted pollutant diffusion concentration data and unify the CO concentration scale when displaying to ensure data consistency and comparability.

[0079] S202: Process the generated cloud map to make its format suitable for Pix2Pix model input, and construct two data sets for different prediction requirements. For the prediction time, add a time channel to the model input, and for different duct outlet wind speeds, add a ventilation wind speed channel to the model input to enhance the applicability of the model.

[0080] Furthermore, the specific operation of step S202 includes the following steps:

[0081] S2021: Resize the generated cloud image to 1200*600 to meet the input requirements of the Pix2Pix model.

[0082] S2022: Construct data sets for the design optimization and monitoring modes respectively. The paired data sets for the design optimization mode are: the three-dimensional model maps the model height to two dimensions, which is expressed in the form of grayscale values. The smoke throwing area and the air duct outlet are marked in the figure, and the pollutant distribution cloud map simulated by the pedestrian layer height at different ventilation times is spliced together. After processing the pixel size of the image, a set of data sets is obtained. The difference between the monitoring mode and the design optimization model is that the pollutant concentration is also marked during the annotation to reflect the monitoring data, thereby obtaining the data set of the monitoring mode. The data set is preprocessed by normalization and other preprocessing, and divided into training set, validation set and test set to meet the needs of model training and evaluation.

[0083] S3, an improved generative adversarial network model based on the Pix2Pix architecture; in the improved generative adversarial network model, the generator adopts the U-Net structure, the discriminator uses the PatchGAN architecture, the time channel and wind speed channel are added to the network input layer, and an attention mechanism improvement module is introduced.

[0084] Specifically, the generative adversarial network model is improved based on the Pix2Pix architecture. The generator adopts the U-Net structure, and the discriminator uses the PatchGAN architecture. To enhance the ability to extract spatiotemporal features, time channels and wind speed channels are added to the network input layer; at the same time, improved modules such as the attention mechanism are introduced to enhance the model's ability to capture key features. The network depth and width need to be optimized and adjusted according to the complexity of the data. Figure 3 As shown in Figure 3, a flowchart for building a rapid pollutant prediction model based on Pix2pixGAN is provided. S3 includes the following steps:

[0085] S301: Improve the generative adversarial network (GAN) based on the Pix2Pix framework, and optimize the objective function through the minimax game. The generator adopts the U-Net structure and uses its encoder-decoder architecture to achieve feature extraction and high-resolution image generation; the discriminator adopts the PatchGAN structure and enhances the detail discrimination ability through the local receptive field.

[0086]

[0087] Where G is the generator, D is the discriminator, G(Z) is the generated data, x~pdata is the sampled data in the real data, x~pz is the sampled noise in the random noise distribution, D(x) is the probability that the discriminator judges that the real data x is true, and D(G(z)) is the probability that the discriminator judges that the generated data G(z) is true.

[0088] S302: To improve the model's ability to capture spatiotemporal dynamic characteristics, additional time channels and wind speed channels are added to the input layer, enabling the network to learn the temporal evolution of pollutant diffusion and the influence of wind speed.

[0089] S303: Introduce an attention mechanism module into the jump connection of the generator to enhance the model's ability to focus on key areas (such as pollution sources, vents, etc.), reduce irrelevant background interference, and improve prediction accuracy.

[0090] α=σ(W a [F enc ,F dec ]+b a )

[0091] Where Fenc and Fdec are the feature maps of the encoder and decoder respectively, σ is the Sigmoid function, and α is the attention weight, which is used to enhance the feature response of key areas (such as pollution sources and vents).

[0092] S304: Optimize network depth and width based on data complexity to avoid overfitting or underfitting. Experimentally compare the effects of different numbers of layers and convolution kernels to select the optimal model structure.

[0093] S4, in the training of the improved generative adversarial network model, the generator is pre-trained using key data sets, and then jointly trained with the discriminator to optimize the GAN; the loss function is optimized, perceptual loss and feature matching loss are added, and dynamic learning rate adjustment and early stopping strategies are used to prevent overfitting.

[0094] Specifically, a two-stage training strategy is adopted, pre-training the generator and then jointly training the GAN. A dedicated optimization monitoring mode is designed to track the changes in indicators such as PSNR and SSIM in real time. The loss function combination is optimized, adding components such as perceptual loss and feature matching loss on the basis of traditional adversarial loss to balance the accuracy and clarity of the generated results. Dynamic learning rate adjustment and early stopping strategies are used to prevent overfitting. S4 includes the following steps:

[0095] S401: First, pre-train the generator separately to enable it to initially learn the mapping relationship of CFD simulation data; then jointly train the generator and discriminator to optimize the generation effect through adversarial learning.

[0096] Furthermore, the specific operation of step S401 includes the following steps:

[0097] S4011: Perform generator pre-training and optimize the generator using only L1 loss. The objective function formula is as follows:

[0098] L pretrain =λ L1 PG(z)-yP1

[0099] Where y is the actual CFD simulation result, and λL1=100 is the weight coefficient.

[0100] S4012: Introduce the discriminator and perform joint adversarial training. The total loss function formula is as follows:

[0101] L total =L GAN +λ L L1+λ FM L FM +λ perc L perc

[0102]

[0103] Where D(i) is the feature map of the ii-th layer of the discriminator, and φ is the intermediate layer feature extractor of the pre-trained VGG19 network.

[0104] S402: Based on the traditional GAN loss, perceptual loss and feature matching loss are introduced to ensure that the generated results are close to the real data in both global structure and local details.

[0105] S403: Adopt an adaptive learning rate adjustment strategy (such as the Adam optimizer) and monitor the PSNR (peak signal-to-noise ratio) and SSIM (structural similarity) indicators on the validation set. If the performance no longer improves, terminate the training early to prevent overfitting.

[0106] S404: Design a visual monitoring module to record the changes in the loss function and the quality of generated samples in real time, so as to facilitate the adjustment of hyperparameters and optimization of training strategies.

[0107] S5, using the trained improved generative adversarial network model to predict the pollutant concentration distribution to obtain the prediction results of a high-resolution two-dimensional image of the pollutant concentration distribution.

[0108] Specifically, a dedicated graphical user interface (GUI) was developed to integrate model prediction and result display functions. The system can achieve near-real-time prediction of pollutant concentration distribution, significantly shortening the prediction time compared to traditional CFD methods. The prediction results are displayed through various visualization methods such as contour maps and heat maps, and support comparative analysis with CFD results to verify the model's prediction accuracy and reliability. Specifically, S5 includes the following steps:

[0109] S501: Calculate error indicators (such as mean square error (MSE) and peak signal-to-noise ratio (PSNR)) on the test set and compare them with traditional CFD simulation results to verify the model's prediction speed and accuracy advantages.

[0110]

[0111] S502: Build an interactive visualization platform that supports users to input environmental parameters (such as wind speed and pollution source location), output pollutant diffusion prediction results in real time, and dynamically render concentration distribution maps.

[0112] S503: Provides multiple visualization methods such as contour maps, heat maps, and dynamic evolution animations to intuitively present pollutant diffusion trends, facilitating engineering analysis and decision support.

[0113] S504: Integrate the comparison function between CFD simulation and GAN prediction in the GUI, support error distribution visualization, and help users evaluate the reliability of the model under different working conditions.

[0114] In summary, addressing the time-consuming, computationally expensive, and difficult-to-achieve rapid prediction challenges of existing methods for pollutant diffusion prediction during underground powerhouse construction, the present invention optimizes the pollutant diffusion prediction process by integrating CFD numerical simulation with machine learning algorithms, enabling rapid and accurate prediction of pollutant diffusion during underground powerhouse construction. This invention achieves lightweight, portable, and intelligent pollutant diffusion prediction during underground powerhouse construction, resolving the high computational costs, slow prediction speeds, and data lags inherent in traditional prediction methods. It improves the efficiency and reliability of the prediction system and provides effective technical support for the safe construction and operational management of underground powerhouse projects.

[0115] The method for rapidly predicting the diffusion of harmful gases in underground caverns, based on a generative adversarial network (GAN), according to an embodiment of the present invention, combines computational fluid dynamics (CFD) numerical simulation with deep learning algorithms to obtain near-real-time predictions of pollutant concentration distributions using an integrated multi-factor dataset. This method is applicable to fields such as underground power plant construction and environmental engineering, particularly in the prediction and management of pollutant diffusion in underground spaces. A data-driven pollutant diffusion model is established, utilizing generative adversarial network (GAN) technology to rapidly predict pollutant diffusion during underground power plant construction.

[0116] In order to implement the above embodiment, Figure 4 As shown, this embodiment also provides a rapid prediction system 10 for the diffusion of harmful gases in underground caverns based on a generative adversarial network, including:

[0117] The simulation module 100 is used to establish a three-dimensional geometric model of the underground cavern and perform numerical simulation of pollutant diffusion using a computational fluid dynamics (CFD) method to obtain simulation results;

[0118] The data set construction module 200 is used to extract key parameter data from the CFD-POST simulation results, visualize them in a cloud diagram, and convert the visualized three-dimensional data into a two-dimensional image for processing to obtain a key data set;

[0119] A generative adversarial network building module 300 is used to improve the generative adversarial network model based on the Pix2Pix architecture; in the improved generative adversarial network model, the generator adopts a U-Net structure, the discriminator uses a PatchGAN architecture, adds a time channel and a wind speed channel to the network input layer, and introduces an attention mechanism improvement module;

[0120] The model training optimization module 400 is used to pre-train the generator using a key dataset and then jointly train and optimize the GAN with the discriminator in the improved generative adversarial network model training; optimize the loss function, add perceptual loss and feature matching loss, and use dynamic learning rate adjustment and early stopping strategy to prevent overfitting;

[0121] The data prediction module 500 is used to predict the pollutant concentration distribution using the trained improved generative adversarial network model to obtain a prediction result of a high-resolution two-dimensional image of the pollutant concentration distribution.

[0122] The rapid prediction system for harmful gas diffusion in underground caverns, based on a generative adversarial network (GAN), in an embodiment of the present invention combines computational fluid dynamics (CFD) numerical simulation with deep learning algorithms to obtain near-real-time predictions of pollutant concentration distributions using an integrated multi-factor dataset. This invention is applicable to fields such as underground power plant construction and environmental engineering, particularly in the prediction and management of pollutant diffusion in underground spaces. A data-driven pollutant diffusion model is established, utilizing generative adversarial network (GAN) technology to achieve rapid prediction of pollutant diffusion during underground power plant construction.

[0123] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0124] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A rapid prediction method for the diffusion of harmful gases in underground caverns based on generative adversarial networks, characterized by: include: A three-dimensional geometric model of the underground cavern was established, and numerical simulation of pollutant diffusion was performed using the computational fluid dynamics (CFD) method to obtain simulation results. Extract key parameter data from CFD-POST simulation results, visualize them through cloud charts, and convert the visualized 3D data into 2D images for processing to obtain key data sets; An improved generative adversarial network model is based on the Pix2Pix architecture. In the improved generative adversarial network model, the generator adopts a U-Net structure, the discriminator uses a PatchGAN architecture, a time channel and a wind speed channel are added to the network input layer, and an attention mechanism improvement module is introduced. In the training of the improved generative adversarial network model, the generator is pre-trained using key data sets, and then jointly trained with the discriminator to optimize the GAN; Optimize the loss function, add perceptual loss and feature matching loss, and use dynamic learning rate adjustment and early stopping strategy to prevent overfitting; The trained improved generative adversarial network model is used to predict the pollutant concentration distribution to obtain the prediction results of a high-resolution two-dimensional image of the pollutant concentration distribution.

2. The method according to claim 1, characterized in that Extract key parameter data from CFD-POST simulation results and visualize them through cloud charts. Convert the visualized 3D data into 2D images for processing to obtain key data sets, including: Extract key parameter data from CFD-POST simulation results and visualize the key parameter data through cloud charts to obtain visualized three-dimensional data; wherein the key parameter data includes pollutant diffusion concentration information; The generated three-dimensional visual data is processed so that the data format is adapted to the Pix2Pix model input, and key data sets are constructed for different prediction requirements. The key data sets are preprocessed and divided into training sets, validation sets, and test sets; wherein the key data sets include optimization mode data sets and monitoring mode data sets.

3. The method according to claim 2, characterized in that Extract key parameter data from CFD-POST simulation results and visualize them through cloud charts to obtain three-dimensional data, including: Use CFD-POST as a post-processing tool for Fluent calculation results to extract pollutant diffusion concentration data at a preset pedestrian level; A cloud map is generated based on the extracted pollutant diffusion concentration data, and the CO concentration scale is unified when displayed to obtain visual three-dimensional data.

4. The method according to claim 3, characterized in that The generated 3D visualization data is processed to make the data format compatible with the Pix2Pix model input. Key data sets are constructed for different prediction requirements, preprocessed, and divided into training sets, validation sets, and test sets, including: The generated cloud image is resized to 1200*600 to meet the input requirements of the Pix2Pix model. Design an optimized model dataset: Map the model height to two dimensions and display it in grayscale. Mark the smoke throwing area and the air duct outlet on the map. Combine this with simulated pollutant distribution cloud maps at pedestrian level at different ventilation times. Process the image pixel size to obtain a dataset. Design the monitoring model data set: when labeling, the pollutant concentration is also marked to obtain the monitoring model data set; The optimization mode dataset and the monitoring mode dataset are normalized and preprocessed, and divided into training set, validation set and test set.

5. The method according to claim 4, characterized in that The generative adversarial network model is improved based on the Pix2Pix architecture. In the improved generative adversarial network model, the generator adopts the U-Net structure, the discriminator uses the PatchGAN architecture, and the time channel and wind speed channel are added to the network input layer. At the same time, an attention mechanism improvement module is introduced, including: The Generative Adversarial Network (GAN) is improved based on the Pix2Pix framework, and the objective function is optimized through the minimax game. The generator adopts the U-Net structure and uses the encoder-decoder architecture to achieve feature extraction and high-resolution image generation; the discriminator adopts the PatchGAN structure: Where G is the generator, D is the discriminator, G(Z) is the generated data, x~pdata is the sampled data in the real data, x~pz is the sampled noise in the random noise distribution, D(x) is the probability that the discriminator judges that the real data x is true, and D(G(z)) is the probability that the discriminator judges that the generated data G(z) is true; Add additional time channel and wind speed channel to the input layer; Introducing an attention mechanism module into the generator’s jump connection to enhance the model’s ability to focus on key areas: α=σ(W a [F enc ,F dec ]+b a ) Where Fenc and Fdec are the feature maps of the encoder and decoder respectively, σ is the Sigmoid function, and α is the attention weight; Optimize network depth and width according to data complexity to avoid overfitting or underfitting; compare the effects of different numbers of layers and convolution kernels through experiments to select the optimal model structure.

6. The method according to claim 5, characterized in that In the training of the improved generative adversarial network model, the generator is pre-trained using key data sets, and then jointly trained with the discriminator to optimize the GAN; Optimize the loss function, add perceptual loss and feature matching loss, and use dynamic learning rate adjustment and early stopping strategy to prevent overfitting, including: Use key data sets to pre-train the generator separately to initially learn the mapping relationship between CFD simulation data. Then jointly train the generator and discriminator to optimize the generation effect through adversarial learning. Based on the traditional adversarial loss GAN Loss, we introduce the perceptual loss and feature matching loss; Adopt an adaptive learning rate adjustment strategy and monitor the PSNR peak signal-to-noise ratio and SSIM structural similarity indicators on the validation set; Design a visual monitoring module to record the changes in loss function and the quality of generated samples in real time to adjust hyperparameters and optimize training strategies.

7. The method according to claim 6, characterized in that The generator is pre-trained separately to initially learn the mapping relationship of CFD simulation data. The generator and discriminator are then jointly trained to optimize the generation effect through adversarial learning, including: Perform generator pre-training and optimize the generator using only L1 loss. The objective function formula is as follows: L pretrain =λ L1 PG(z)-yP1 Where y is the actual CFD simulation result, λL1=100 is the weight coefficient; The discriminator is introduced to conduct joint adversarial training. The total loss function formula is as follows: L total =L GAN +λ L L1+λ FM L FM +λ perc L perc Where D(i) is the feature map of the ii-th layer of the discriminator, and φ is the intermediate layer feature extractor of the pre-trained VGG19 network.

8. The method according to claim 7, characterized in that The trained improved generative adversarial network model is used to predict the pollutant concentration distribution to obtain a high-resolution two-dimensional image of the pollutant concentration distribution, including: The error index was calculated on the test set and compared with the traditional CFD simulation results to verify the model's advantages in prediction speed and accuracy: Build an interactive visualization platform that supports user input of environmental parameters, outputs pollutant diffusion prediction results in real time, and dynamically renders concentration distribution maps; Provides multiple visualization methods such as contour maps, heat maps, and dynamic evolution animations to intuitively present pollutant diffusion trends; The comparison function between CFD simulation and GAN prediction is integrated into the GUI, supporting error distribution visualization.

9. A rapid prediction system for the diffusion of harmful gases in underground caverns based on generative adversarial networks, characterized by: include: The simulation module is used to establish a three-dimensional geometric model of the underground cavern and perform numerical simulation of pollutant diffusion using the computational fluid dynamics (CFD) method to obtain simulation results. The data set construction module is used to extract key parameter data from CFD-POST simulation results, visualize them through cloud graphs, and convert the visualized three-dimensional data into two-dimensional images for processing to obtain key data sets; A generative adversarial network building module is used to improve the generative adversarial network model based on the Pix2Pix architecture. In the improved generative adversarial network model, the generator adopts a U-Net structure, the discriminator uses a PatchGAN architecture, and a time channel and wind speed channel are added to the network input layer. At the same time, an attention mechanism improvement module is introduced. The model training optimization module is used to pre-train the generator using key datasets and then jointly train and optimize the GAN with the discriminator in the improved generative adversarial network model training; Optimize the loss function, add perceptual loss and feature matching loss, and use dynamic learning rate adjustment and early stopping strategy to prevent overfitting; The data prediction module is used to predict the pollutant concentration distribution using the trained improved generative adversarial network model to obtain the prediction results of a high-resolution two-dimensional image of the pollutant concentration distribution.

Images