ATA: A transformer network for solar power prediction

Abstract With a global focus on sustainable energy, solar power is vital. However, its intermittency challenges the power grid’s stability. Conventional deterministic models fail to handle such variability. Thus, this study proposes a novel solar-power prediction method with uncertainty estimation, composed of a hybrid Transformer-LSTM architecture optimized by nature-inspired algorithms. Adaptive bandwidth kernel density estimation (ABKDE) is leveraged to capture prediction uncertainties. Experiments on real-world data show that the method improves R 2 by 20% and reduces RMSE by 25% versus conventional models. It not only provides accurate point predictions but also reliable prediction intervals (95% PICP at 0.95), delineating the stochasticity of solar power well and enhancing energy-system resilience.