Abstract
Risk assessment is a well-established process and is often used to assess the risk of systems in several application domains, e.g., maritime, nuclear power, and aerospace. However, the results from a risk assessment represent a snapshot of the risk picture for the system at a given time and, for dynamic systems whose states can change suddenly, these results become obsolete after a while. Dynamic risk assessment methods consider the ever-changing and uncertain nature of dynamic systems. A subset of these methods consists of simulating how the system states evolve in time while considering rules, events and actions that can lead to accident events with undesirable consequences. This position paper studies the use temporal convolutional networks to predict the probability of accident events given a time-series of events, to assist dynamic probabilistic risk assessment methods and improve their computational performance. The paper’s contributions are a study of the properties of temporal networks and an evaluation of their feasibility for predicting the probability of future accident events in dynamic probabilistic risk assessment methods. Furthermore, an approach is proposed that uses a temporal network for probabilistic forecasting, i.e., predicting a probability distribution based on time series data. The output of this network could be used to inform and guide the dynamic risk assessment process.
