Towards an OpenModelica-based sensitivity analysis platform including optimization-driven strategies

Abstract

Parameter sensitivity analysis is a core activity to assess the robustness of dynamical models with regard to unreliable parameters. This becomes critical for nonlinear models with many parameters subject to large uncertainties. In such contexts too often numerical experimentation is required due to the lack of analytic expressions for the derivatives of state variables with respect to parameters. A naive sweeping of the full parameter space is usually not an option due to combinatorial explosion. In this work we present OMSens, an open platform to assess the sensitivity of Modelica models tailored to work with OpenModelica. OMSens uses different methods to sensitivity analysis including an approach based on derivate-free non-linear optimization. This is a new approach not previously used in Modelica tools which provides important advantages such as robustness and applicability to models for which the derivatives of state variables don't exist or are not available. We tested OMSens with a Modelica version of World3, a large nonlinear socio-economic model. OMSens was effective to pinpoint a nonintuitive subset of parameters that, when perturbed within small ranges, yield strong changes on key state variables.