Artificial Intelligence AI composition is one of the hot topics that have been debated in recent years. In this paper, we first extract monophonic and chordal features from MIDI digital music files. Then the WaveNet intelligent music generation model is used as a carrier to optimize its multilayer convolutional network structure. The audio files are fed into the optimized WaveNet model, and the final training parameters are obtained after several rounds of iterative training. After the model completes the training, music sequences are automatically generated. The results show that the optimized WaveNet model for training leads to a significantly higher accuracy rate in the validation set than before optimization. Compared to other models, the method in this paper generates music using a larger variety of notes, improving the quality of the music theory and chord aspects. Compared with the composite scores of human compositions, the percentage of WaveNet model compositions with scores of 4 and above is about 20.3%, and the percentage of scores of 3 and above is 30.5%. Therefore, the overall level of the compositions generated by the model in this paper is good.