Accurate photovoltaic ( PV) power output forecast is the basis of ensuring safe and economic operation of the grid with high ratio of PV power. To solve the problem that point forecast is difficult to express the probability credibility，an interval forecast method of PV power output based on conditional generative adversarial network was studied. The mapping relationship between the historical measurement data and the predicted PV power output was established，then the probable scenarios of PV power output were obtained. And on this basis， the bootstrap method and interval estimation were used to calculate upper and lower bounds of the confidence interval，so as to determine the forecast interval of PV power. Experimental results have shown that the forecast intervals of short-term PV power output obtained by the proposed method can cover the real measured power values， and provide valuable decision information for the safe and economic operation of power grid.