Underwater visible light communication (UVLC) has the advantages of high speed, low latency, and high confidentiality. However, the signal transmission is susceptible to light-emitting diode (LED) modulation bandwidth, non-linear effects of LEDs, and underwater channels. Neural networks, capable of addressing complex nonlinear problems, are increasingly applied to signal equalization in visible light communication. It is found that multilayer perceptron (MLP) is capable of extracting spectral features and nonlinear relationships, and gated recurrent unit (GRU) is capable of handling timing correlation and channel fading problems. Therefore, we propose a GRU-MLP model as a post-equalizer for the UVLC system and experiment using orthogonal frequency division multiplexing modulated signals on a 60-cm underwater experimental platform. The results show that the GRU-MLP equalizer can extend the system transmission bandwidth by 47% higher than the bidirectional gated recurrent unit (BIGRU) and long short-term memory (LSTM) equalizer when only limited by the LED bandwidth; under the influence of the underwater optical channel, the performance of GRU-MLP is similar to that of BIGRU and LSTM. The bit error rate of the GRU-MLP algorithm is significantly lower than other algorithms under the combined effect of two factors. In summary, GRU-MLP demonstrates superior equalization performance in bandwidth-constrained complex channel environments.