The invention discloses a photovoltaic short-term output prediction method based on improved Gaussian process regression. A weather type index, a daily mean temperature and a humidity similarity are taken as references. Data is classified. Aiming at different weather types, different influence factors are selected. In a matlab, a corresponding weighted linear decreasing particle swarm Gaussian process regression LinWPSO-GPR prediction model is established. The method comprises the following steps of step1, carrying out data classification; step2, establishing a prediction model; step3, acquiring a prediction sample output expectation which is a power prediction value. In the prior art, an optimization searching process of the photovoltaic short-term output prediction method is easy to fall into a local optimization state, and an optimization effect and iteration convergence depend too much on an initial value. By using the method, the above problems in the prior art are solved.