Because of the existence of high order dispersion, non-linear distortion of optical signals usually appears during the fiber propagation process, it results in the increase of bit error rate of received signals. In view of this, a prediction algorithm of optical signals based on deep neural networks is proposed. This algorithm summarizes the signal propagation model of fiber communication system first, then decides the parameters of input layer and output layer for the deep neural networks. Finally, based on Marine Predators Algorithm the algorithm searches the parameters of hidden layers of neural networks during the training phase. Experimental results show that the deep neural networks can predict the correct wave of received optical signals, and it can provide foundation for dispersion compensation of fibers.