Abstract: With the high-penetration integration of distributed generation, zero-sequence protection faces the risk of malfunction due to the increased complexity of fault characteristics. To address this issue, this study proposes an adaptive protection method that integrates Bayesian optimization and random forest feedback collaboration. The method establishes a closed-loop architecture of "offline optimization–online decision-making": offline, Bayesian optimization is employed to generate a mapping set of optimal settings under multiple operating conditions; online, random forest is utilized for fault direction identification and dynamic fine-tuning of settings, with data feedback driving the updating of the mapping set. Simulation results based on Simulink demonstrate that the proposed method achieves a protection operation accuracy of over 96.7% under various scenarios, effectively resolving the adaptive protection challenge of zero-sequence protection in IIDG-integrated distribution networks.