Generating well-informed and reliable predictions for disaster evacuation is a large challenge. Crisis and disaster management policymakers have to deal with poor data quality, a limited understanding of households’ behaviour dynamics, and uncertainty regarding the effects of the various actions/measures in place. Agent-based simulation models are frequently used to support decisions when planning disaster evacuation procedures. However, one of the most important aspects of this issue, which is social influence, is not often considered. Most of existing evacuation models largely overlook the importance of the households’ behaviours and social influences, which leads to oversimplified models. Moreover, it is almost impossible to find models in the literature that focus on the extrinsic decision-making factors of some evacuees, such as compromised lifelines, in the case of catastrophic events. In contrast to the existing evacuation models, this paper suggests a probabilistic agent-based model that relies on the loss of different lifelines as factors affecting evacuees’ decision-making in addition to some intrinsic factors that are used to characterise the propensity of households to evacuate and explicitly allow for social contagion as well as uncertainties to be considered. This model, in which all the variables are considered uncertain and Monte Carlo Simulations are run to estimate the confidence range of the predictions, is tailored to estimate the potential number of inhabitants that have not been evacuated in high-rise buildings in the face of critical infrastructure failures induced by a slow-onset flood and/or the actions taken during the related crisis, considering different uncertainties that may affect the reliability of the prediction. The model has been specifically designed to predict the dynamics of households’ self-evacuations in fourteen residential high-rise buildings located in a flood-prone area in Paris. This paper describes the suggested model and also reports the results of an illustrative case study in which three scenarios are simulated to demonstrate the applicability of the model, to test its effectiveness and to explore the uncertainty regarding some modelling assumptions using sensitivity analysis