RR-LTNet: Ramp-Rate-Centric Deep Learning Framework for Short-Horizon Photovoltaic Power Prediction

Reliable operation of photovoltaic (PV) fleets with high grid penetration demands prediction tools that translate plant-level irradiance variability into actionable intelligence for PV power management and balance-of-system design. Reliably managing high penetrations of solar PV within modern grids requires predictions that reflect the systemic consequences of fast renewable variability across cyber-physical energy infrastructure. This article proposes RR-LTNet, a ramp-rate (RR)-centric prediction architecture that elevates the RR as the core variability feature linking plant-level intermittency to grid-level operational risk. RR-LTNet first performs dynamic clustering of RR regimes to characterize rapid weather transitions and stability conditions. These regimes taken as feature for hybrid temporal learner that fuses recurrent memory with temporal convolution to capture multiscale dynamics. For predictive assessment, numerical experiments are conducted on one year solar power database of the Yulara Solar Project, Australia. The proposed RR-LTNet consistently outperforms other feature extraction methods by achieving up to 90% reductions in root mean square error (RMSE) for 5-min resolution. The reductions are significantly larger when baseline models are augmented with RR-aware features than when those features are absent. Cross-site validation on two additional PV plants with different capacity and data distribution confirms robustness and consistency required for system engineering deployment across fleets. By surfacing variability intelligence in real time, RR-LTNet supports PV-specific tasks central to reserve scheduling, advanced simulation of PV plant-grid interactions. This framework thus bridges PV monitoring analytics with system level reliability engineering, accelerating the integration of large-scale PV into modern power systems.