This study introduces a hybrid deep learning framework designed for photovoltaic power generation forecasting, which combines Long Short-Term Memory (LSTM) networks with a Transformer-based structure. The LSTM element is employed to model long-term temporal patterns in the time series, whereas the Transformer component increases computational efficiency and strengthens the model’s ability to capture global features via parallelized data processing. Experimental results indicate that the proposed LSTM-Transformer model attains higher prediction accuracy and enhanced robustness compared to baseline methods, capable of accurately representing both immediate variations and prolonged patterns in photovoltaic power output. The incorporation of multi-source data—including meteorological conditions and historical power records—further enhances forecasting performance. These findings highlight the model’s potential for improving the accuracy and stability of renewable energy predictions. Future work may extend this approach to other energy domains and explore the integration of additional external data sources to further boost predictive capabilities.