According to the Seveso Directives, the risk assessment is crucial for an effective control of major accident hazard. Nevertheless, the complexity of many Seveso sites, due to human, technical and organizational factors makes recognized common practices limited because of their intrinsic static nature. In this paper, a dynamic approach for risk assessment is proposed, which allows evaluating moment by moment the state of the system under analysis by Bayesian belief networks. A petrochemical coastal storage was selected as applicative case-study to verify the capability of the dynamic approach. Network training is performed by entering historical reliability data, near-miss and accidents data series collected on-site by periodical inspection plans on critical elements, as well as from the evidences of SMS reports. Upon proper refinement and further validation with reliable field data, the predictive approach may be used as a management decision-making tool.