Objective The five-centered circular horseshoe shape has become the dominant cross-section shape for underground water-sealed caverns. Rapidly and accurately determining its optimal parameters is of great theoretical importance and engineering value for the safe and efficient construction of such projects.

Methods Based on the engineering geological characteristics of underground water-sealed caverns, the five-centered circular horseshoe-shaped cross-section was selected as the research focus. Under the condition of a fixed cavern cross-sectional area for oil and gas storage, the rise-span ratio was identified as the controlling parameter, and an intelligent prediction model for this ratio was proposed. The core of the model was constituted by an evolutionary neural network algorithm, in which the initial weights and thresholds of artificial neural networks were optimized using genetic algorithms. The lateral pressure coefficient at the project site, rock mass elasticity modulus, initial friction angle of discontinuities, and discontinuities combinations were adopted as prediction indicators. Through orthogonal experiments, combination schemes of the four prediction indicators were designed, and the optimal rise-span ratio for each scheme was determined using discrete element simulation. These indicator combinations and their corresponding optimal ratios were input into the evolutionary neural network algorithm as training or testing datasets, by which an intelligent prediction model for the cross-section’s rise-span ratio was established.

Results The prediction model was applied to an underground water-sealed cavern project in Zhejiang Province, where the predicted rise-span ratio of 0.642 closely matched the optimal 0.67 obtained from discrete element simulation and the final adopted ratio of 0.71, determined after considering factors including the blasting excavation scheme and construction convenience. Moreover, no major engineering issues occurred during construction, demonstrating the model’s high feasibility and reliability.

Conclusion As more underground water-sealed cavern projects succeed, increasingly reliable training datasets will be generated, enabling continuous updates to the evolutionary neural network model and creating a positive feedback loop with engineering practice. The proposed intelligent prediction model for the rise-span ratio of the five-centered circular horseshoe-shaped cavern cross-section offers a reliable design tool and valuable theoretical and technical guidance for similar projects.