Spatiotemporal Attention With Conditional Feature Modulation for Satellite-Based Solar Irradiance Prediction

Accurate short-term solar irradiance forecasting is critical for grid stability, yet the existing deep learning models often struggle to capture complex dynamics over longer horizons. In this letter, we propose the attention-contextual U-Net (AC U-Net), a new architecture for multistep solar irradiance map prediction using GK-2A satellite imagery over the Korean Peninsula. The proposed model enhances a standard U-Net by integrating spatiotemporal attention and a featurewise linear modulation (FiLM) layer. This layer injects high-level contextual information directly into the model's bottleneck, allowing for dynamic feature adaptation. Experiments demonstrate that AC U-Net outperforms not only the persistence baseline but also deep learning-based competitors, such as ConvLSTM and HRNet. Our work demonstrates that conditioning on contextual features is a powerful strategy for improving long-range forecasting accuracy.