An adaptive mQAM-OFDM wireless optical communication system for underwater dynamic-variable environments based on Meta-learning deep neural networks, is proposed in this paper. The receiver in this system firstly employs Meta-learning to train an initial model to obtain the general characteristics of different water quality types of optical channels, making it rapidly adaptive to unknown changes in underwater channel characteristics for optimal reception. It is shown by simulation experiment that the system has stronger learning capability, faster convergence and lower BER performance in than conventional LS-OFDM systems, machine learning-based C-DNN-and P-DNN-OFDM systems.