Aiming at the problem that both accuracy and speed of the present three dimensional positioning techniques in the visible light communication system are still not good, an indoors positioning method for visible light communication system based on deep learning is proposed. Firstly, a neural network is designed to encode the fingerprint data to two dimensional array, and the convolutional neural network is utilized to learn the relationship between the fingerprint array and the target position; Then, the hyperparameters of the convolutional neural network are tuned automatically by the particle swarm optimization algorithm, thereby the training difficulty of the deep neural network is reduced. Besides, a method of dividing the training set、verification set and test set for the visible light positioning dataset is designed, it helps to mitigate the overfitting problem of the neural network, and improve the positioning accuracy. Simulation results show that, the average positioning error of the proposed method is 0.024m in 6×6×4m3 indoors environment, and the average positioning time is 0.478s.