Dimensionality reduction of the resilience model of a critical infrastructure network by means of elementary effects sensitivity analysis

International audienceModern Critical Infrastructures (CIs) are typically characterized by a large number of elements interconnected and interdependent. Their mathematical representation reflects these characteristics in models that typically turn out to be: 1) complex, since the relation between the variables can be nonlinear; 2) large, since a high number of variables is typically involved in the model; 3) dynamic, because the behavior of the system evolves in time. For this reason, the opportunities of exploring these models in order to extract information, such as identifying the most critical events, is conditioned by the computational cost of a simulation run and by the number of variables to explore. In this paper, we investigate the possibility of reducing the dimensionality of a model by identifying the variables that affect it most, by means of the Elementary Effects (EEs) method, which is a sensitivity analysis method capable of screening the input variables resorting to a limited number of model evaluations. Since the performance of the method relies on its settings, we analyze them proposing at the same time possible improvements. A hybrid network for gas and power distribution is considered as case study. The objective is to rank the importance of some uncertain parameters of the network (e.g., its failure and recovery characteristics) with respect to the system resilience properties (i.e., the capability of mitigating the effect of components failures and/or recovering its performance)