A deep learning model based on multi-attention mechanism and gated recurrent unit network for photovoltaic power forecasting
Kuo Yang, Yanjie Cai, Jinrun Cheng
Abstract
Solar energy plays a crucial role in the power grid due to its clean, stable, and cost-effective nature, as well as its significant storage potential. Accurate short-term photovoltaic (PV) power forecasting is essential for effective grid management and dispatching decisions. This study introduces a hybrid deep learning model integrating multiple attention mechanisms and gated recurrent unit networks to forecast PV output power one day in advance. To address the impact of random weather variations and historical PV power data on forecasting accuracy, the model incorporates an input attention mechanism to process input features. Additionally, temporal and spatial attention mechanisms are embedded within the encoder-decoder framework to enhance prediction performance. These mechanisms effectively capture the relationships between historical PV power output and meteorological variables while identifying crucial time-dependent hidden states. The proposed model is validated on a real-world PV dataset, achieving a mean absolute error of 0.0903 under favorable weather conditions, demonstrating a 22.5 % improvement over traditional forecasting methods across various weather classifications. Comparative analyses with other state-of-the-art models confirm that the proposed approach offers superior predictive accuracy.