With the increasing penetration of photovoltaic generation (PV), its output has impacted the power grid significantly. However, due to complex weather factors, PV output is intermittent and uncertain, making it challenging to conduct power grid dispatching. Therefore, this paper presents a data-ahead PV output forecasting method based on deep learning autoencoder and long short-term memory neural network (DAE-LSTM). First, an extraction method is proposed to select the key weather factors based on the improved DAE. Then, the LSTM is applied to forecast PV output using key weather factors. Simulations for actual PV output data are conducted to verify the effectiveness of the proposed method.