Abstract: Typhoon-induced cascading power flow overloads expose the limitations of traditional emergency control methods, particularly under decision-making mechanisms based on static security criteria and discrete control actions, often resulting in insufficient predictive capability, low control precision, and poor economic efficiency. This paper proposes a data-driven and physics-constrained blocking control strategy for cascading power flow overloads in power grids. First, a spatio-temporal graph convolutional network (STGCN) is constructed to deeply explore the spatiotemporal correlation characteristics of typhoon evolution using ERA5 meteorological reanalysis data, enabling accurate prediction of overload propagation trends. Second, a physics-constrained generative adversarial network (GAN) is developed, conditioned on the spatiotemporal features extracted by STGCN, to generate time-series grid failure scenarios covering pre-disaster, disaster, and post-disaster phases. Physical constraints such as power flow balance and voltage security are explicitly embedded during training to ensure the engineering feasibility of the generated scenarios. Subsequently, a blocking control decision framework based on a parameterized deep Q-network (PDQN) is designed to achieve hybrid optimization and multi-stage coordinated control of generator tripping, load shedding, and phase-shifting transformer (PST) phase-angle adjustment, effectively blocking the propagation paths of cascading overloads. Finally, simulations on the IEEE 39-bus system demonstrate that the proposed method improves the overload elimination rate by 7% and reduces the total control cost by 23% compared to the proximal policy optimization (PPO) algorithm, validating its effectiveness and superiority in blocking cascading overloads.