Skip to content
Open access

Physics-Anchored Ramp-Aware Multi-Horizon Learning for Reliable Photovoltaic Power Forecasting

2026 · Computer Modeling in Engineering & Sciences · 0 citations · 63 references

Abstract

: Ultra-short-term photovoltaic (PV) forecasting must preserve the deterministic solar trajectory that governs stable production while remaining responsive to cloud-induced ramps. This paper presents a physics-anchored ramp-aware multi-horizon forecasting framework in which clear-sky-index persistence remains the default and fallback forecast, while a validation-selected learned specialist is activated only when a causal gate identifies ramp-prone conditions. Learning is therefore used as a selective correction mechanism rather than as an unconditional replacement for the physical anchor. Separate horizon-specific systems are developed for 5, 15, 30, and 60 min forecasting. The primary evaluation uses Desert Knowledge Australia Solar Centre data from 2022–2024, with 2022–2023 used for model development and 2024 reserved for future-year testing. Additional in-year, leave-one-year-out, Alice Springs System 60, and Yulara experiments provide broader temporal and cross-system evidence. In the primary future-year test, the proposed framework achieves all-sample root mean squared error (RMSE) skill of 5.99%–8.54% and mean absolute error (MAE) skill of 3.13%–5.46% relative to clear-sky-index persistence. On realized ramp events, RMSE and MAE skill reach 8.10%–11.38% and 8.52%–13.53%, respectively, while the 95th-percentile absolute error decreases by 6.72%–10.95%. Dependent-sample statistical tests and moving-block bootstrap intervals support improvement over the physical anchor at every horizon. The broader benchmark characterizes an accuracy–adaptivity trade-off: the framework provides consistent anchor-relative skill, explicit physical fallback, interpretable regime-dependent routing, and a specialist duty cycle of approximately 20%.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.