Hybrid AI-Based Long-Term Solar Forecasting in Tunisia Using Transformer, Prophet, and Monte Carlo Simulation

Abstract

Accurate long-term solar forecasting is essential for optimizing Photovoltaic (PV) investments, especially in climate-sensitive regions like Tunisia where environmental variability introduces significant uncertainty. This paper proposes a hybrid machine learning framework that integrates the Transformer deep learning model and the Prophet statistical forecasting tool, each enhanced with Monte Carlo simulation to generate probabilistic Global Horizontal Irradiance (GHI) forecasts for the next 30 years from 2025 to 2055. The Transformer model captures short-term temporal dependencies through attention mechanisms, while Prophet excels in trend and seasonality decomposition. Climate variables, including temperature, humidity, and the clearness index, are used as input characteristics to reflect local environmental conditions. The experimental results demonstrate that Prophet achieves superior long-term accuracy (R2 = 0.978, MAPE = 3.65%), while Transformer remains effective for short-term responsiveness. Monte Carlo simulations further improve reliability by offering confidence intervals for future scenarios. The proposed hybrid approach offers a robust and interpretable solution for the long-term planning of solar energy under climate uncertainty.