As a common RF device in the microwave circuits, the power divider is an important building block for building the multiple-input multiple-output (MIMO) feed network in 5G communication systems. In order to optimize and quickly redesign the existing fixed frequency power divider structure so as to apply to any actually desired operating band, including the 5G operating band, in this paper, we propose a deep learning scheme based on a modified one-dimensional convolutional neural network taking a pre-designed dual-band power divider as the optimized design target. The one-dimensional convolutional neural network could predict that the geometric structure parameter of the power divider has good performance at other arbitrary double resonant frequencies. The self-organizing mapping neural network is used to select samples to improve the training efficiency of convolutional neural network. The predicted power divider performance is verified in the electromagnetic(EM) simulation software. The simulation results show that the return loss of the power divider is higher than 20 dB at the resonant frequency, the isolation is higher than 25 dB, the insertion loss is lower than 3.4 dB, and the working bandwidth is about 450~600 MHz, which proves that the optimized design of multi-parameter objective power divider is a fast and effective method.