Geometry prediction and design for energy storage salt caverns using artificial neural network

Salt cavern is one of the best storage for hydrogen, compressed air, and natural gas. However, the current physical/numerical simulation-based construction design cannot yield optimal solutions due to the complex correlation of multiple construction parameters. This paper proposes a novel machine-learning-enabled method for the geometry prediction and design optimization during salt cavern construction. A dataset of 1197 simulations of salt cavern construction is collected using our previously developed program. Construction parameters are included as inputs, and the simulated cavern geometries as outputs. A Gated Recurrent Unit model is selected and well-trained from 600 artificial neural network models. The model achieves a mean absolute error of 1.6m in the test dataset and of 2.83m for the geometric prediction of cavern JT52 in Jintan, meeting the field design requirements. Furthermore, an optimized construction design method is proposed by looping generating construction parameters, shape prediction, and deviation calculation until the deviation meets requirements. It succeed in designing an ideal ellipsoid cavern in approximately 51 min, with a capacity ratio fc 31 % larger than those of the field caverns. This approach demonstrates the potential for machine learning methods to serve as the third generation of construction design methods after physical and numerical modeling.