Visual-Aware Solar Forecasting: Attention-Driven Spatiotemporal Learning from Sky Images for Grid-Responsive PV Prediction

Abstract

Solar energy plays a pivotal role in sustainable energy systems, yet its inherent variability due to weather dynamics poses significant challenges for grid integration and operational efficiency. This research addresses these challenges by proposing an attention-based sequence-to-sequence (Seq2Seq) deep learning model for short-term solar energy forecasting, leveraging the SKIPP’D dataset comprising sky images (64×64 resolution, 1-minute intervals) and synchronized photovoltaic (PV) generation data. The hybrid architecture integrates a convolutional neural network (CNN) encoder for spatial feature extraction from sky images and long short-term memory (LSTM) decoder enhanced with multi-head attention mechanisms to capture complex spatiotemporal dependencies. Rigorous preprocessing ensures data alignment, normalization, and partitioning into training, validation, and test sets stored in HDF5 format. Model performance is evaluated using metrics including root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²), with separate analyses for sunny and cloudy conditions. Results demonstrate robust performance on sunny days (RMSE: 2.156, MAE: 1.720, R²: 0.914) and moderate accuracy under cloudy skies (RMSE: 4.885, MAE: 3.818, R²: 0.579), achieving an overall RMSE of 3.779 and R² of 0.755. The model outperforms traditional baselines (e.g., ARIMA, LSTM) by 30% in RMSE during clear-sky periods, validating the efficacy of attention mechanisms in prioritizing critical temporal features. By addressing the limitations of conventional models in handling rapid cloud-induced fluctuations, this work contributes to enhanced grid stability, optimized energy storage dispatch, and scalable renewable energy integration.