Behavior-based scenario discovery using time series clustering

Abstract

Scenario Discovery is a widely used method in model-based decision support for identifying common input space properties across ensembles of exploratory model runs. For model runs with behavior over time, these properties are identified by reducing each run to a single value, which obscures potentially decision-relevant dynamics. We address the problem of considering dynamics in Scenario Discovery by applying time series clustering to the ensemble of model runs, and then finding the common input properties for each cluster. This separates the input space into multiple scenarios, each corresponding to a distinct model dynamic. Policy interventions can be targeted at different scenarios by analyzing overlap of these subspaces. Our work expands Scenario Discovery by improving consideration of system behavior over time, which is highly relevant for the management of complex nonlinear systems such as ecosystems or technical infrastructure.