Hybrid CNN-LSTM modeling with attention mechanism for solar power forecasting

Accurate short-term forecasting of solar photovoltaic (PV) power is essential for improving power-electronics-based control and energy management in modern renewable energy systems. This study proposes a hybrid deep learning framework that integrates convolutional neural networks, long short-term memory units, and an attention mechanism to predict hourly PV power with high precision, enabling dynamic operation of dc-dc converters, inverters, and grid-connected PV units. To address the challenge of missing meteorological data in the first considered site, a Random Forest regression model is employed to reconstruct the incomplete solar radiation and cloudiness data. The framework was then pretrained on the PV data of a second site (Alice Springs, 2014–2017) and then fine-tuned using a five-month dataset collected from a PV plant near Urmia, Iran (Sep 2024–Feb 2025). Through numerical studies, it is shown that the attention layer of the proposed technique further enhances interpretability by highlighting time intervals that are most influential for power conditioning and converter response. Notably, training exclusively on the first site data outperformed the transfer learning setup, achieving a coefficient of determination ( R 2 ) of 0.9500, root mean squared error (RMSE) of 0.0688, mean absolute error (MAE) of 0.0321, and a peak error of 0.0398% under high-generation conditions. This level of performance supports more reliable short-term adjustment of converter setpoints and inverter dispatch, contributing directly to voltage stability and efficient energy management in grid-connected PV systems. Overall, the proposed approach reinforces the synergy between artificial intelligence driven forecasting and power electronics control, offering a scalable and interpretable solution that supports renewable energy integration and smart grid stability, particularly in data-limited environments.