In order to solve the problems of low reliability and user fairness in indoor non-orthogonal multiple access (NOMA) visible light communication(VLC) channel estimation methods, a method of using the Gray Wolf optimizer(GWO) algorithm to compensate and restore signals in deep neural network (DNN) in NOMA-VLC systems, that is GWO-DNN channel estimation method. This method introduces nonlinear convergence factors and two-dimensional chaotic mapping, effectively improving system reliability and multi-user transmission fairness. Experimental results show that, at a bit error rate (BER) of 104, the proposed method can achieve a maximum signal-to-noise ratio (SNR) gain of 4.3 dB compared to the minimum mean square error (MMSE) method, and the SNR difference between the two users is reduced from 3.2 dB to 0.9 dB. Under different orders of quadrature amplitude modulation, the performance of the proposed method is better than the MMSE method, and the higher the modulation order, the more significant the performance improvement.