Weapon-Target Assignment (WTA) is a core combinatorial optimization problem in the domain of air defense command decision-making. Existing studies largely rely on the complete-information assumption and adopt a single-solution paradigm, thereby overlooking the robustness differences arising from incomplete information and multimodal solutions, as well as the insufficient exploitation of knowledge across stages. To address this, this paper proposes a dynamic partially-observable WTA (D-POWTA) framework. First, a complete mathematical model is established to explicitly characterize information incompleteness and incorporate 13 classes of operational constraints. Second, a Robustness Metric Index based on the Hamming neighborhood (RMI) is introduced to quantitatively characterize the local quality-dominance margin of multimodal solutions within their encoding neighborhoods. Finally, a Knowledge-Guided Dynamic Rolling Evolutionary Algorithm (KG-DREA) is designed, achieving an organic unification of knowledge—multimodality—robustness through cross-stage knowledge reuse, Hamming-distance-based niche preservation, and RMI-driven robust selection. Experimental results show that, across 10 test instances with 24 to 500 targets, KG-DREA achieves the best performance in both solution quality and convergence stability, and ablation studies further verify the effectiveness and contribution of the three core mechanisms: cross-stage knowledge transfer, multimodal preservation, and robust selection.