With the high propotion integration of new energy sources, the risk assessment of cascading failure in the power grid has become increasingly crucial for its safe and stable operation. The stochastic variability of new energy sources and the characteristics of power electronics integration contribute to a heightened level of randomness and complexity in cascading faults. Traditional cascading fault models fail to effectively capture the disconnection of new energy integration units due to voltage fluctuations, and the risk assessment for large-scale power grids still grapples with the trade-off between accuracy and complexity. To address this challenge, this paper proposes a cascading fault risk assessment method based on a physics-informed neural network. Firstly, we establish uncertainty models for the random disconnection of new energy sources and line overload tripping. Simultaneously, we introduce a fault graph chain to describe the state transition characteristics of electrical quantities, topological relationships, and other physical features during cascading fault processes. Secondly, we present cascading fault risk indicators corresponding to the fault graph. Through physics-informed neural network, data-driven empirical replay calculation, the key physical information features of chain fault process (such as the state of electricity volume and topological relationship changes) in power system are embedded into the data-driven model power system into a neural network, aiming to approximate the nonlinear mapping relationship between fault graphs and cascading fault risk indicators, facilitating rapid risk assessment. Finally, the effectiveness of the proposed method is validated through case studies.