Facing the global energy crisis and environmental pollution, photovoltaic (PV) power generation is highly valued due to its clean and renewable characteristics. However, the volatility of PV power poses challenges to grid stability. Accurate prediction of PV power output is crucial for optimizing power resources and ensuring grid stability. This study proposes a deep learning model called BiLSFormer, which integrates a bidirectional long short-term memory network (BiLSTM) with the Transformer architecture, specifically designed for long-term PV power forecasting tasks. The BiLSFormer model enhances its ability to capture key features in time series data by introducing adaptive spectral blocks (ASB) and interactive convolutional blocks (ICB). ASB highlights spectral features using Fourier analysis, while ICB refines temporal dynamic features; together, they improve the model's extraction capabilities of both local and global temporal features. Experimental results show that the BiLSFormer model demonstrates competitive predictive performance in long-term PV power generation forecasting, achieving an R2 value of 0.993, a mean absolute error (MAE) of 2.98, a mean squared error (MSE) of 28.35, and a root mean squared error (RMSE) of 5.32, all of which outperform existing models.