Abstract:Toxic or flammable gas release accident will possible form dangerous vapor cloud, which is of significant concerns to industrial safety and social security. In the current stage, several models are available for simulating and predicting the dispersion of hazardous gases to the surrounding environment, but these models all require information about the release source, for example, the release rate or release velocity, which is difficult to acquire in emergency situation. There is a need to find an effective method to predict the real-time gas dispersion for toxic/flammable gas release accident. An integrated method combining a gas detection system, numerical simulations using the PHAST and an artificial neural network (ANN), was developed to aid decision making involved in emergency situation. The purpose of the implementation of the integrated method is to find a way to predict the dispersion of released gases while bypassing the requirement of source information, thus saving plenty of time for emergency response. The integrated method is an extension to the existing consequence analysis tools and is specialized in real-time gas dispersion prediction. The core of the new integrated method is using neural network to correlate the potential gas detector readings with PHAST simulation results, especially the concentration and expected evacuation time for surrounding environmental sensitive areas, before any real accidents happen. When real accident occurs, the correlation between the gas detector readings and PHAST results makes it possible to predict the future dispersion of released gases as soon as the released gas being detected by gas detectors. That is the reason why this integrated method could bypass using source information to predict the released gas dispersion. The method contains four steps. To begin the implementation procedure, risk analysis is needed to distinguish the possible release scenarios, including the release position, release type, release substance and surrounding conditions, weather Author: Bing Wang, Tel: +86-10-62773894, E-mail: wangb07@hotmail.com Corresponding author at: Tsinghua University, Department of chemical engineering, 100084 Beijing, China Tel.: + 86-10-62783109;E-mail address: jinsongzhao@tsinghua.edu.cn conditions are also included. Based on the recognized release scenario, consequence analysis is executed in order to estimate the coverage or influenced area of the released gases through gas dispersion simulation. Several gas dispersion models or consequence analysis tools are available for these dispersion simulation, and the integrated method is suitable for any of them, because the accuracy of the integrated model is significantly determined by the dispersion model used. The third step is to extract the useful data from the simulation results, as in this paper, the results from PHAST. The extraction algorithm could find simulated concentration value of released gas at the points where gas detectors are placed, and also the concentrations at environmental sensitive areas as well as the expected evacuation time for these areas, then these extracted data are prepared in the order that is suitable for neural network training. The last step is to train a BP-network with the prepared data. Some of the data sheet are used to build the inner-connections between neurons, the rest are used to verify the reliability of the integrated method by compare the neural network predictions with the existing PHAST results. If the accuracy is acceptable, then the parameters of the network are saved for further prediction, otherwise the network will be initialized and trained again. A hypothetical liquefied chlorine release scenario was studied and tested by using the integrated method. In the case study, the parameters of gas detector placement were optimized and a local optimal solution of the placement was found, the optimal placement covers a ±20° range based on a 95% scenario relevance level. The prediction of the chlorine concentration and the expected time for evacuation at the off-site environmental sensitive areas proved to be accurate in contrast with the PHAST results for the same scenario. The regression slope, indicating the coincidence between network output and PHAST simulation results is 0.9531 for concentration prediction and 0.9965 for the expected evacuation time prediction, as shown in figure 1, where the slope = 1 means perfect predicting accuracy for the integrated method compared with PHAST. Qualitative analyses estimating whether the off-site sensitive areas are impacted by chlorine plume were carried out in the case study. The integrated method could give reliable results for 98.3% of the possible release scenarios. From the case study, the integrated method for accidental gas release in emergency situation proves to be an effective and reliable way for decision making process. Figure 1 Fitness analysis between neural network estimation and PHAST simulation results for (a) chlorine concentration and (b) plume arrival time (or expected time for evacuation)

Real-time hydrogen release and dispersion modelling of hydrogen refuelling station by using deep learning probability approach

Real time hydrogen plume spatiotemporal evolution forecasting by using deep probabilistic spatial-temporal neural network

Numerical Study of Hydrogen Leakage and Explosion in Hydrogen Refueling Station with Large Volume Hydrogen Storage Vessels

Prediction model for the evolution of hydrogen concentration under leakage in hydrogen refueling station using deep neural networks

Prediction of hydrogen leakage location and intensity in hydrogen refueling stations based on deep learning

Probabilistic real-time deep-water natural gas hydrate dispersion modeling by using a novel hybrid deep learning approach

Probabilistic real-time natural gas jet fire consequence modeling of offshore platforms by hybrid deep learning approach

Hydrogen leakage location prediction at hydrogen refueling stations based on deep learning

Numerical Simulation on the Leakage and Diffusion of Hydrogen in Indoor Environment

Real-time natural gas explosion modeling of offshore platforms by using deep learning probability approach

Hydrogen jet diffusion modeling by using physics-informed graph neural network and sparsely-distributed sensor data

Forecasting PM2.5 Using Hybrid Graph Convolution-Based Model Considering Dynamic Wind-Field to Offer the Benefit of Spatial Interpretability.

An Integrated Prediction Method for Real-Time Accidental Toxic Gas Release in Industrial Plants

Spatiotemporal Predictions of Toxic Urban Plumes Using Deep Learning

Computational fluid dynamics model based artificial neural network prediction of flammable vapor clouds formed by liquid hydrogen releases

Digital Twin-Based Hydrogen Refueling Station (HRS) Safety Model: CNN-Based Decision-Making and 3D Simulation

Development of the Quantitative Property-Consequence Relationship Model for Prediction of Hydrogen Leakage and Dispersion Using Response Surface Method and Artificial Neural Network Approaches

Hydrogen leakage risk assessment for hydrogen refueling stations

Forecasting air pollutant concentration using a novel spatiotemporal deep learning model based on clustering, feature selection and empirical wavelet transform

Deterministic risk assessment of hydrogen leak from a fuel cell truck in a real-scale hydrogen refueling station

Quantifying uncertainty: Air quality forecasting based on dynamic spatial-temporal denoising diffusion probabilistic model