In recent years, the share of new energy generation in the total power output of electrical systems has been on a continuous rise. Accurate prediction of PV power, a vital component of new energy generation, is significant for the power grid's safe, stable, and cost-effective operation. It is becoming increasingly crucial to develop more accurate and reliable forecasting methods for new energy generation. This paper introduces a PV power forecasting method constructed on a Wavelet Transform-Convolutional Neural Network-Long Short-Term Memory-Attention Mechanism (WT-CNN-LSTM-Attention) model. In this proposed approach, wavelet transform (WT) is utilized to decompose the data, which includes both weather forecast information and operational data from the PV power plant, into frequency components with different time scales. The convolutional neural network (CNN) is utilized to capture spatial correlation features from the input data, while the long-short-term memory (LSTM) network captures the temporal relationships within the sequence of input data. Additionally, an attention mechanism is introduced to address the limitations of LSTM in retaining critical information when processing long sequences. By leveraging the complementary strengths of these various neural network architectures, the overall prediction accuracy is significantly enhanced. A comparative analysis of the proposed WT-CNN-LSTM-Attention model against traditional models, including LSTM, CNN-LSTM, and Gated Recurrent Unit (GRU), demonstrates that the proposed model delivers superior predictive performance.