Whenever a flood happens, emergency services must deal with inhabitants that may agree or disagree the recommendations to evacuate. This paper investigates a model that aims to assess the number of unevacuated inhabitants in tall buildings located in urban areas when a flood occurs, considering different uncertainties that make the prediction difficult. The objective of this prediction is to apprehend the condition of evacuation of the population so that the emergency services can organize better. The main considered uncertainties are the flood hydrograph, the people propensity to evacuate by themselves, and the effects of different events associated to a given height of water (warnings, recommendations, power cut, etc). The model is analytical and inspired by multi agent systems. Different factors and interaction laws are introduced with their uncertainties in order to get the number of peoples unevacuated for each 6 hours step of a flood simulation. Based on Monte Carlo simulations, the confidence interval of the prediction may be assessed. The model has been developed to be able to predict the evacuation of 16 buildings of Parisian district.