Addressing photovoltaic (PV) forecasting challenges: Satellite-driven data models for predicting actual PV generation using hybrid (LSTM-GRU) model

This study proposes a robust approach for predicting actual PV generation in data-scarce regions using satellite-derived inputs, addressing key limitations in current forecasting models. Its novelty lies in applying a modified z-score transformation to bridge the distribution gap between satellite-derived and measured PV generation by introducing a clear and transparent empirical relationship between the two data sets. The effectiveness of the proposed approach is rigorously validated across a diverse set of well-established models (XGBoost, SARIMAX, CNN-LSTM, LSTM-GRU, and informer) through three distinct scenarios using 17 years of PVGIS satellite data and one year of measured PV generation data from Areza, Eritrea. The Informer model consistently outperformed others, underscoring its suitability for complex forecasting tasks, while traditional models showed low performance. The first scenario, which uses satellite-derived data for both training and testing, serves as a baseline to verify model performance and reliability under consistent conditions. In scenarios 2 and 3, actual PV generation was forecasted using models trained on satellite-derived data without and with modified z-score transformation, respectively. The transformed data (scenario 3) yielded promising accuracy, achieving an enhancement by up to 43 % in R² compared to the untransformed case (scenario-2). Furthermore, results showed that the prediction error difference between the first and third scenarios was only 0.69 %, indicating a nearly negligible disparity. Notably, data transformation improves forecasting accuracy across all models, demonstrating the approach’s robustness and effectiveness in data-scarce regions. The findings provide practical guidance for researchers, system operators, and policymakers aiming to scale PV integration in data-scarce regions.