| 학회 | 한국화학공학회 |
| 학술대회 | 2017년 봄 (04/26 ~ 04/28, ICC 제주) |
| 권호 | 23권 1호, p.1266 |
| 발표분야 | 화학공정안전 |
| 제목 | Developing real time evacuation alarm system in toxic gas release using CFD and Deep Gaussian Process |
| 초록 | To reduce the damege of toxic gas release accdient, gas dispersion models with high accuracy are necessary to predict toxic gas movement. In the urban area where dispersion obstacles are large and abundant, CFD would be the best choice to simulate and analyze the scenarios of accidental release of toxic chemicals. However, because of the large computation time that need to calculate CFD simulation, it can`t be used in emergency situation. For developing the toxic gas release alarm system, deep Gaussian process surrogate model is developed. Surrogate model can predict the critical region with the toxic gas concentration through the entire map instantaneously after accident occur. Because of large computation cost for simulating the original CFD model, it is hard to construct the pre-calculated database for all scenario. Thus, fully Bayesian treatment, multivariate GP, is used for constructing the surrogate model when data is scarce. GP also integrated with the deep belief networks (DBN) called deep Gaussian process to extract and predict the gas diffusion information from complicated urban area with lots of buildings and contour of the ground. |
| 저자 | 전경우, 나종걸, 한종훈 |
| 소속 | 서울대 |
| 키워드 | 위험성평가 |
| 원문파일 | 초록 보기 |
