Urban power network (UPN) topologies are partially observable due to invisible underground cables, leading to incomplete information under malicious attacks. To perform vulnerability identification for UPN under such attacks, it is of great importance to consider topology completion uncertainties in attack and defense strategies. To this end, this paper first proposes a UPN topology completion method, where a series of possible complete topologies associated with their corresponding probabilities are determined by means of Monte Carlo simulation and a knowledge-assisted hierarchical network partitioning algorithm, i.e., the improved SHRINK (I-SHRINK) algorithm. Subsequently, a bi-level stochastic optimization model is established to determine attack strategies by considering topology completion uncertainties, further identifying UPN vulnerability. The upper-level problem aims to maximize the expected load shedding risk, while the goal of lower-level problem is to minimize the total load shedding. Case studies are performed in a real UPN to verify the effectiveness and superiority of the UPN topology completion and vulnerability identification methods.