Transfer Learning for PV Generation Forecasting under Data Scarcity: A Case Study in Brazil
Abstract
This study investigates the use of Transfer Learning (TL) techniques to improve photovoltaic (PV) power forecasting in Brazilian regions characterized by heterogeneous climatic conditions and limited data availability. An Encoder–Decoder Long Short-Term Memory (EDLSTM) model was initially trained using a multi-year dataset from São Paulo and subsequently adapted to three target sites, Tacaratu (PE), Itacarambi (MG), and Florianópolis (SC), through five distinct TL strategies. The experiments evaluated model performance under varying data availability scenarios (1 to 12 months) and were benchmarked using a persistence model. Among the tested configurations, TL1, which reuses pretrained weights without layer freezing, consistently achieved the lowest normalized Root Mean Square Error (pRMSE) and demonstrated superior stability across all climate contexts. These findings confirm that knowledge transfer from data-rich domains significantly enhances predictive performance in data-scarce regions, reducing both training effort and computational cost. The proposed approach provides a scalable and data-efficient framework for PV forecasting, contributing to the reliable integration of solar energy into the Brazilian power system.
Keywords
Transfer Learning, Photovoltaic Power Forecasting, LSTM, Encoder Decoder, Data Scarcity