Probabilistic long-term forecasting of solar irradiance under climate change using deep learning and Monte Carlo simulations

Abstract

While traditional forecasting models assume a stationary climate, the accelerating impact of climate change in arid zones like Southern Tunisia has introduced significant non-stationarity and variance in Global Horizontal Irradiance (GHI). The objective of this study is not to simulate specific IPCC climate scenarios, but rather to develop a robust hybrid deep learning and Monte Carlo framework capable of capturing the stochastic complexity and increased uncertainty inherent in recent historical records. By integrating LSTM and Informer architectures with probabilistic Monte Carlo simulations, this framework provides a distribution of potential outcomes, offering a resilient planning tool that accounts for the heightened climatic variability that characterizes the current transition period.

This study introduces a novel hybrid framework that combines the Informer and Long Short-Term Memory (LSTM) deep learning models with Monte Carlo simulations to generate probabilistic forecasts of Global Horizontal Irradiance (GHI) over a 30-year horizon (2023–2053) in southern Tunisia. Unlike conventional time series approaches, the Informer architecture is specifically designed to capture long-range temporal dependencies with high computational efficiency, making it well suited for modeling complex atmospheric dynamics. Multivariate climatic inputs including the clear sky index, ambient temperature, and relative humidity are used to enhance both temporal accuracy and climate sensitivity. Monte Carlo simulation integration enables rigorous uncertainty quantification by providing confidence intervals for predicted irradiance values. Notably, relative humidity, often neglected in desert climatology emerges as a key predictor statistically associated with irradiance variability, accounting for deviations of up to $\pm 5\text{\%}$ in post 2053 projections. By fusing deterministic deep learning with stochastic simulation, the proposed framework offers a scalable and reproducible methodology for long-term solar forecasting. The 30-year projections represent probabilistic risk envelopes rather than deterministic forecasts; high-confidence planning is supported primarily by the first decade (Reliability Zone: PINAW $<0.15$), while projections beyond 10 years constitute speculative scenarios suitable for stress-testing and strategic risk assessment. These findings provide actionable insights for infrastructure design and strategic planning, particularly in regions vulnerable to climate induced energy fluctuations.