Abstract: As a long-duration physical energy storage technology centered on thermal energy storage, Carnot battery （CB） can effectively mitigate grid fluctuations caused by high-penetration renewable energy. However, current research has focused predominantly on the heat storage side of CB, leaving a noticeable gap in the design of cold storage-type CB systems and the efficient optimization of their thermodynamic operating parameters. This study first proposed a novel CB system coupled with a liquefied natural gas （LNG） regasification process. By integrating LNG cold energy as the cryogenic heat sink for the Rankine cycle, the power-to-power （P2P） efficiency of the system was significantly enhanced. Furthermore, an artificial neural network-based surrogate model optimization strategy was developed. This approach substantially reduced the computational cost and iteration time associated with complex rigorous model optimization, while maintaining the quality of global optimization. The results indicated that the hyperparameter-optimized surrogate model could accurately predict system performance indicators and thermodynamic constraints. Specifically, the relative error between the predicted P2P efficiency and the results from the rigorous model was less than 1%, and the absolute error of the predicted pinch point temperature differences was within 1.5 ℃. When applied to particle swarm optimization, the average computation time was reduced from 430726.5 s to a mere 5387.1 s. Moreover, the optimized system achieved a P2P efficiency of 76.13%, representing an increase of 10.97 percentage points over the initial design. This confirms that the proposed strategy significantly improves computational efficiency while ensuring effective process optimization.