Keywords
Scenario Discovery; MCMC; Perfect Storm Scenarios; Water Distribution Networks; DREAM
Start Date
15-9-2020 4:20 PM
End Date
15-9-2020 4:40 PM
Abstract
Scenario Discovery is an approach for finding combinations of uncertainties that explain behaviour of interest of a model. In order to do so, the full uncertainty space is sampled using an independent sampling scheme of choice, generating a large number of scenarios. The outcomes of these scenarios are classified, and subspaces of the uncertainty space with a high density of interesting scenarios are identified. However, this approach is suitable only to problems with a relatively low number of uncertainties since the number of required experiments to adequately characterize the outputs typically scales faster than linear with the number of uncertain factors. Small regions in the uncertainty space containing “perfect storm” scenarios, where a number of small, seemingly unimportant disruptions lead to system failure, are therefore easily missed. This paper presents the use of a dependent sampling approach for scenario discovery. Specifically, we explore the use of DREAM for scenario discovery purposes. DREAM is a Markov Chain Monte Carlo (MCMC) algorithm with multiple chains, which by using a suitable likelihood function can be used to more efficiently sample from those regions in the uncertainty space that are predictive of behaviour of interest. To demonstrate this approach, we explore the vulnerability of a small-scale water distribution system. While in urban water networks, vulnerability assessment has been limited to probability-based component failures (i.e., pipe breakage, leakage, excess demand), the proposed approach assesses the “perfect storm” vulnerability of a standard water distribution system which is modelled through EPANET software within the WNTR package. On the basis of this application to a water distribution system, we propose that the approach may be useful for performing scenario discovery on models with high numbers of uncertain factors.
An approach for Scenario Discovery on high-dimensional problems
Scenario Discovery is an approach for finding combinations of uncertainties that explain behaviour of interest of a model. In order to do so, the full uncertainty space is sampled using an independent sampling scheme of choice, generating a large number of scenarios. The outcomes of these scenarios are classified, and subspaces of the uncertainty space with a high density of interesting scenarios are identified. However, this approach is suitable only to problems with a relatively low number of uncertainties since the number of required experiments to adequately characterize the outputs typically scales faster than linear with the number of uncertain factors. Small regions in the uncertainty space containing “perfect storm” scenarios, where a number of small, seemingly unimportant disruptions lead to system failure, are therefore easily missed. This paper presents the use of a dependent sampling approach for scenario discovery. Specifically, we explore the use of DREAM for scenario discovery purposes. DREAM is a Markov Chain Monte Carlo (MCMC) algorithm with multiple chains, which by using a suitable likelihood function can be used to more efficiently sample from those regions in the uncertainty space that are predictive of behaviour of interest. To demonstrate this approach, we explore the vulnerability of a small-scale water distribution system. While in urban water networks, vulnerability assessment has been limited to probability-based component failures (i.e., pipe breakage, leakage, excess demand), the proposed approach assesses the “perfect storm” vulnerability of a standard water distribution system which is modelled through EPANET software within the WNTR package. On the basis of this application to a water distribution system, we propose that the approach may be useful for performing scenario discovery on models with high numbers of uncertain factors.
Stream and Session
false