A multi-transformer push-pull type photovoltaic inverter comprises a DSP, a DC/DC convertor, a DC/AC convertor and a filter, wherein the DC/DC convertor, the DC/AC convertor and the filter are sequentially connected, and a composite controller which is composed of a PI controller of the DC/AC convertor, an artificial neural network model and a feed-forward arithmetic unit is arranged in the DSP. The input end of the artificial neural network model is connected with the state variable of the DC/AC convertor in the kappath sampling period, a predicted value of the output voltage which is outputted by the artificial neural network model in the (kappa+1)th sampling period is connected with the negative input end of the feed-forward arithmetic unit, the positive input end of the feed-forward arithmetic unit is connected with a set value of voltage at the kappa moment, and an output signal of the feed-forward arithmetic unit is connected with an output signal of he PI controller and then controls the duty ratio of the DC/AC convertor. According to the multi-transformer push-pull type photovoltaic inverter, the output voltage of the inverter at the next moment is predicted through the neural network model, compensation for the duty ratio at the present moment is timely conducted through the feed-forward arithmetic unit, and therefore the hysteresis quality of the PI controller is overcome, and the robustness of the inverter system is improved.