Time-Series Foundation Models from Physics-Informed Synthetic Histories for Cold-Start PV Forecasting

arXiv·low signal

This paper tackles the cold-start problem — forecasting before any target-site observations exist — by pretraining time-series foundation models on physics-informed synthetic histories. The generalizable takeaway for practitioners is the recipe: synthesize physically-grounded pretraining data to bootstrap a foundation model when real data is unavailable, a pattern applicable well beyond photovoltaic forecasting.