In view of the shortcomings of the single variable prediction method of photovoltaic (PV) power, a new multivariable phase space reconstruction prediction method of PV power is designed. Firstly, based on the correlation analysis, the historical PV power and meteorological factors time series of the actual PV power plant are selected to form multivariate time series. Then, the multivariable phase space of PV power prediction is reconstructed by C-C method and false nearest neighbors ( FNN) method, and its chaotic characteristics are identified by small data method. Finally, combined with the powerful nonlinear fitting ability of radial basis function (RBF) neural network, a PV power prediction model based on multivariate phase space reconstruction and RBF neural network is established. The example analysis shows that the proposed multivariate phase space reconstruction prediction method has better performance than the single variable prediction method.