At present, the prediction models of power output of a grid-connected photovoltaic system are mainly divided into linear and nonlinear models. The linear model is simple, easy to implement but less flexible. Because of the strong randomness of the output power of photovoltaic power generation system, the nonlinear model often provides better prediction. Given that, the output power prediction method of a photovoltaic system based on Multiple Adaptive Regression Spline (MARS) is proposed. This paper first analyzes the principle of the algorithm, determines the flow of model analysis, and introduces the data source of the photovoltaic system analyzed in this paper. Secondly, taking temperature, sunshine time and other factors as independent variables, the MARS model is analyzed and studied, and the simulation model for photovoltaic power prediction is determined. Finally, the prediction method proposed in this paper is compared with the existing prediction method. RMSE, MAD and MAPE of various methods are compared and analyzed through training data and test data, and photovoltaic daily output power is predicted based on historical data. The comparison proves that the MARS model is more accurate in predicting the output power of a photovoltaic system. This work is supported by the National Key Research and Development Program of China (No. 2018YFB2100103).