Marine Dissolved Oxygen Concentration (MDOC) is one of the key factors affecting marine aquaculture and marine ecosystems. Aiming at the problems of complexity and diversity of factors affecting MDOC and the difficulty of prediction, this paper proposes a research method of marine dissolved oxygen concentration prediction based on modal decomposition. Based on the MDOC data monitored by self-research sensors, this study selects the best data for subsequent analysis by pre-processing the raw data, analyzing the results of temporal resampling and the initial decomposition of Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). An adaptive Variable-division Modal Decomposition (VMD) decomposition method is proposed to decompose the high-frequency decomposition terms of the initial decomposition in a secondary decomposition to fully explore the internal features of the data. Prediction is performed by constructing a Convolutional Neural Networks (CNN)-Bidirectional Long Short-Term Memory (BiLSTM)-Attention Mechanism (attention). The experimental results show that the method proposed in this paper can fully extract many features from the data and achieve high prediction accuracy and effect, providing a new research idea for the prediction of MDOC.