Abstract: In view of the traditional decomposition forecasting method neglecting the coupling relationship between multi-dimensional meteorological factors such as solar irradiance and photovoltaic power in time and frequency domains,as well as the low feature learning efficiency,slow training speed,over fitting and other problems in the training of depth neural network,a short-term photovoltaic power forecasting method based on multivariate variational mode decomposition（MVMD）and hybrid depth neural network is proposed.Firstly,MVMD is used to analyze the time-frequency synchronization of the photovoltaic power sequence and the multi-dimensional meteorological sequence,and decompose them into frequency-aligned multivariate intrinsic mode functions,thereby reducing the influence of nonlinearity and volatility in the sequence. Secondly,for the multivariate intrinsic mode functions,a forecasting model based on a hybrid deep neural network is established respectively. The model uses convolutional neural network and bidirectional long short-term memory neural network to extract the spatial correlation characteristics and temporal correlation characteristics of photovoltaic power and meteorological sequence,respectively,and uses attention mechanism to enhance the learning weight of important time point features. In addition,the residual connection is used to speed up the training speed of the network and alleviate the overfitting problem. The superiority of the proposed method in this paper is verified by the actual engineering experiment analysis.