A subcarrier-partition (SP) particle swarm optimization (SP-PSO) algorithm was proposed to search the optimal power allocation in discrete multi-tone signal for visible light communication (VLC) systems. Combining non-linear modeling of Volterra series, the relationship between the transmitted power and the effective signal-to-noise ratio of every subcarrier could be easily evaluated, which avoids the complicated deep neural network used in the iteration of the PSO. By SP and linearly interpolation, we further improved the performance of the PSO-based power allocation scheme. Simulation results show that only eight sub-channels are sufficient for the SP-PSO algorithm to search for the optimal power allocation. It has a gain of 0.03 to 0.2 bit / s / Hz when compared to uniform power allocation, water filling, and the pre-frequency domain equalization, which shows the great potential of increasing the channel capacity of the VLC systems.