A novel dual-stream attention-based hybrid network for solar power forecasting

Photovoltaic (PV) power forecasting is essential for providing accurate data on future power production, ensuring secure power grid operations, and reducing solar energy operation expenses. This research introduces a novel dual-steam hybrid model that uses Bidirectional Long-Short Term Memory (BiLSTM) and Convolutional Neural Networks (CNN) to predict PV power production. The proposed model employs a parallel processing technique, with BiLSTM and CNN analyzing input data independently to detect temporal and spatial features. These features are then combined and passed to the multihead attention layer to further identify the most desirable features for PV power forecasts. The proposed model's performance is thoroughly assessed by a series of experiments that include various window sizes, four seasons, and different weather conditions. Subsequently, the predictive accuracy of the developed model is compared with three single and five hybrid deep learning models. The findings show that the dual-stream attention-based hybrid network can precisely predict future PV production across various meteorological, seasonal, and climatic conditions.