To better manage the risks of destructive natural disasters, impact models can be fed with simulations of extreme scenarios to study the sensitivity to temporal and spatial variability. We propose a semi-parametric stochastic framework that enables simulation of realistic spatio-temporal extreme felds using a moderate number of observed extreme space-time episodes to generate an unlimited number of extreme scenarios of any magnitude. Our framework draws sound theoretical justification from extreme value theory, building on generalized Pareto limit processes. For illustration on hourly gridded precipitation data in Mediterranean France, we calculate risk measures using extreme event simulations for yet unobserved magnitudes.