Abstract: To address issues such as low load prediction accuracy, slow dynamic response, and equipment degradation caused by the high proportion of renewable energy grid connection in the energy internet（EI）, a two-layer energy management framework is proposed based on the collaborative optimization of the Kolmogorov-Arnold network（KAN） and model predictive control（MPC）. First, a KAN-driven load forecasting model is constructed to significantly improve forecasting accuracy. Second, an MPC rolling optimization model is designed to enhance the system’s dynamic response capability. Finally, a “prediction-decision-compensation” closed-loop architecture is designed to form a two-layer strategy for collaborative optimization. Case study analysis shows that the proposed method achieves superior load forecasting performance, with a root mean square error of 12.34 kW, a coefficient of determination of 0.995 6, and an average absolute percentage error of 1.82%; the power fluctuation suppression rate of control scheduling reaches 83.4%, with overshoot stabilized within 29.5 kW. Economic optimization reduces the average daily power purchase and sale cost to 358.6 RMB, lowers the battery wear cost ratio to 0.198, and increases the energy storage system’s cycle life by approximately 20%. The research findings significantly enhance the operational economic efficiency and robustness of EI in high-proportion renewable energy grid-connected environments. Through precise prediction, efficient scheduling, and equipment protection, they effectively reduce system operating costs, improve power supply reliability, and extend the lifespan of critical equipment. This provides an innovative solution for achieving more efficient, economical, and reliable operation of the energy internet, holding significant practical significance for promoting the safe and stable grid connection of high-proportion renewable energy and advancing energy transition.