다양한 통계 모델을 이용한 폭발위험범위 예측 모델 개발에 관한 연구

Abstract

The korean occupational safety and health act defines how to evaluate the hazardous Zone of Explosive Gases according to KS(Korean Industrial Standards). In many companies handling combustible gases or flammable liquids, explosion-proof electrical equipment have been installed according to the Korean Industrial Standards(KS C IEC 60079-10-1). In this standard, hazardous area for explosive gas atmospheres has to be classified by the evaluation of the evaporation rate of flammable liquid leakage. The evaporation rate is an important factor to determine the zones classification and hazardous area distance. However, there is no systematic method or rule for the estimation of evaporation rate in this standard and the first principle equations of a evaporation rate are very difficult. Thus, this problem can trigger inaccurate results for evaluating evaporation range. In this study, empirical models for estimating an evaporation rate of flammable liquid have been developed to tackle this problem. Throughout the sensitivity analysis of the first principle equations, it can be found that main factors for the evaporation rate are wind speed and temperature and empirical models have to be nonlinear. Polynomial regression is employed to build empirical models. In addition, Partial Least Square(PLS) is performed to address the multi-collinearity of the regression model. Methanol, benzene, para-xylene and toluene are selected as case studies to verify the accuracy of empirical models. Evaluating extents of hazardous zone of explosive gases is very complex and difficult task. Thus, it is really hard for industrial workplaces to understand these procedures and employ too complicated equations. In this study, Predictive models for the extents of hazardous zone of explosive gases are developed to install the explosion proof apparatus. 1,200 research data sets including 12 combustible gases are generated to train predictive models. The hazardous extent is set to an output variable and 12 variables required in the data generation are set as input variables. Multiple linear regression, principal component regression, and artificial neural network are employed to develop predictive models. Throughout the comparisons of models, mean absolute percentage errors of linear regression, principal component regression, and artificial neural network is 44.2%, 49.3%, 5.5% and root mean square errors is 1.389m, 1.602m, 0.203m respectively. Therefore, it can concluded that the artificial neural network shows the best performance and this is the best optimal model for predicting explosive gas hazardous extents.