The accurate forecast of photovoltaic power is of great significance for the security, stability and economic operation of the power grid. Therefore, a day-ahead photovoltaic power forecasting method is proposed. The method of wavelet transform （WT） is used to decompose numerical weather prediction （NWP） data and photovoltaic power data into frequency data with time information, eliminating the influence of randomness and volatility in data information on forecasting accuracy. Convolutional neural network （CNN） model is used to deeply excavate the seasonal characteristics and spatial correlation characteristics of input data, and bi-directional long-short term memory （BiLSTM） model is used to obtain the temporal correlation of input data series. A day-ahead photovoltaic power forecasting model based on WT-CNN-BiLSTM is constructed. Taking a certain photovoltaic power station as the calculation object, the forecasting results of WT-CNN-BiLSTM model, CNN-BiLSTM model, LSTM model, GRU model and PSO-BP model are compared and analyzed under different seasons and climatic conditions. The calculation results show that the forecasting accuracy of WT-CNN-BiLSTM model is higher than that of other models.