The installed capacity of grid-connected photo-voltaic (PV) systems continues to grow. Due to variability in PV power production, accurate forecasts are essential to support power system operation. This paper presents a hybrid PV power forecasting method with parallel architecture, which combines a Seasonal Auto-Regressive Integrated Moving Average (SARIMA) model with an Artificial Neural Network (ANN) model using weighing factors computed periodically via a least squares problem. The method can be used to obtain short-term forecasts with prediction horizons from 15 minutes to 1 day or more, and is therefore well-suited for Model Predictive Control (MPC). We apply the method to forecast the power output of a rooftop PV system at the Lawrence Berkeley National Laboratory. Our analysis provides high-level suggestions to optimize the order and structure of SARIMA and ANN models. The results show that the hybrid method can reduce forecast error by 10% compared with using the individual models separately, while increasing resilience and redundancy due to its parallel architecture.