Spatio-Temporal Sky Modeling for PV Output Nowcasting: A Cloud-Based Deep Learning Framework

The increasing integration of photovoltaic (PV) energy into power systems is challenged by the stochastic variability of solar irradiance, primarily driven by rapid cloud dynamics. This thesis proposes a modular Deep Learning-based framework for intra-hour solar nowcasting, leveraging ground-based fisheye sky images to predict PV power output. The research addresses the limitations of monolithic forecasting approaches by designing a decoupled pipeline consisting of domain-adaptive segmentation, spatio-temporal video prediction, and direct image-to-power regression. For cloud detection, a UNet++ architecture is implemented. To overcome the scarcity of labeled data, a domain adaptation strategy utilizing histogram matching in the CIELAB color space was developed, enabling the successful transfer of learned features between heterogeneous datasets. For cloud motion modeling, a comparative analysis of state-of-the-art architectures was conducted. While novel Transformer-based models (e.g., PredFormer) exhibited convergence issues, an adapted ConvLSTM baseline demonstrated stability and physical plausibility in capturing non-linear cloud advection. PV power forecasting is performed via a modified ResNet-18 regression network. A strong inductive bias was introduced by augmenting the input with a fourth channel explicitly encoding the sun's position. An "Oracle" benchmark analysis reveals that this mapping module achieves high performance on ground-truth images, isolating the video prediction uncertainty as the primary system bottleneck. Experimental results on the SKIPP’D dataset indicate that while the proposed pipeline struggles to outperform Smart Persistence at the 1-minute sampling frequency due to blurring artifacts, it exhibits a trend of skill recovery at longer horizons (10+ minutes). The findings demonstrate that a modular deep learning framework provides a scalable and robust foundation for industrial grid management and predictive control.