The caustic ratio is the key indicator of the low-temperature pipelined alumina dissolution process. It is defined as the molar ratio of caustic sodium oxide to alumina in the output solution, which is closely related to the dissolution efficiency. Currently, the caustic ratio can only be obtained via off-line analysis with a long time delay, leading to delayed control of the process and poor product quality. Therefore, it is imperative to establish an accurate model for online prediction of the caustic ratio. Based on the principle of the irreversible primary reaction of dissolution and process analysis, this paper proposes a mechanisyic predictive model of the caustic ratio. Then, an error compensation model based on the nuclear extreme value learning machine is established, further proposing the online predictive model of the caustic ratio in the low-temperature pipelined dissolution process. Finally, an adaptive updating strategy for the model is designed in this paper, developing an online adaptive predictive model of the caustic ratio. The accuracy and adaptability of the proposed model are demonstrated through verification experiments.