Modern power systems are being fundamentally reshaped by the increasing integration of renewable energy sources, such as solar and wind power generation. Their inherent variability poses significant challenges for real-time grid control. In particular, advanced control strategies, including model predictive control, require multi-step forecasts at temporal resolutions relevant to the specific control application, which in this work is on the order of seconds. However, commonly available meteorological datasets are often temporally sparse, creating a pronounced mismatch between data availability and control requirements. To address this challenge, this thesis investigates the use of physics-informed neural networks to forecast photovoltaic (PV) power generation in spatially distributed PV systems at temporal resolutions on the order of seconds under temporal data sparsity.

The proposed PINN incorporates partial physical knowledge of atmospheric processes through a spatiotemporal cloud motion equation, enabling physically plausible generalization to unseen time steps and conditions. To generate multi-step forecasts at temporal resolutions on the order of seconds, a recursive forecasting framework is developed. In addition, an improved training strategy is proposed in which the model is trained using its own recursively generated forecasts, thereby reducing the mismatch between training and inference. Finally, a novel correlation-based collocation point sampling strategy is developed to generate physically plausible and statistically representative collocation points, thereby supporting effective physics-based regularization.

The proposed methods are evaluated through two complementary case studies. In a case study based on a selected region in France, the PINN with the improved training strategy demonstrates a clear advantage under challenging atmospheric conditions compared to a persistence model, which assumes constant behavior over the forecast horizon. It also achieves superior data efficiency relative to purely data-driven neural networks and improved computational efficiency over a physics-only model during inference. The proposed collocation point sampling strategy also consistently outperforms uniform sampling.

In addition, a real-world case study based on a measurement site in Hawaii validates the forecasting framework under a realistic PV deployment scenario in which available data are not only temporally sparse but also satellite-derived. Because such data provide only an approximate, temporally and spatially smoothed representation of the true atmospheric conditions affecting the PV system, this setting introduces an additional data-reality mismatch. Nevertheless, the framework remains capable of producing meaningful forecasts at temporal resolutions on the order of seconds, establishing its practical applicability for PV systems worldwide without local measurement infrastructure.