The cooperative encirclement of an evading target by multiple Unmanned Surface Vehicles (USVs) is a critical focus in maritime operations. This paper presents MATD3-SHPER, a novel multi-agent cooperative encirclement approach for USVs, integrating an adaptive policy sharing mechanism with a hybrid prioritized experience replay strategy. To address the challenges of sparse rewards and low coordination efficiency in traditional multi-agent reinforcement learning (MARL), a phase-dependent reward mechanism is introduced, distinguishing between the approach and encirclement phases, thereby providing targeted behavioral guidance for the agents. Furthermore, inspired by federated learning (FL), a phased strategy aggregation and dynamic evaluation mechanism is proposed to enable policy fusion and recovery. To further improve sample efficiency and policy stability, a Hybrid Prioritized Experience Replay (HPER) strategy is employed, combining TD-error with the newly proposed Cooperative Encirclement Priority (CEP) metric, and adjusting dynamic weights to shift the training focus across stages. Experimental results demonstrate the effectiveness and stability of the proposed method in complex multi-agent cooperative tasks.