Harnessing Prophet model with rigorous cross validation for long term solar insolation prediction

Accurate long-term solar irradiance forecasting is critical for renewable energy integration and grid management. This study introduces a novel framework combining the Prophet time series model with Mutual Information (MI) based feature selection for photovoltaic power prediction in Tunisia's Rjim Matouk site. Key contributions include: (1) A hybrid feature selection approach integrating MI and Pearson correlation, reducing MAPE by 46.7% compared to baseline Prophet; (2) Seasonal decomposition methodology achieving 50% error reduction in seasonal forecasts; (3) Comprehensive validation against 2024 state-of-the-art models (iTransformer, PatchTST, TimesFM, Chronos, SOFTS), demonstrating 6.5-23.6% lower MAPE with 96% lower training time; (4) Multi-horizon analysis (1 month to 10 years) with rigorous backtesting validation. Our Prophet+MI framework achieves MAPE of 5.12% (1-month), 7.62% (1-year), and 10.15% (10-year) horizons, outperforming recent transformer architectures while requiring only 0.35 hours training time versus 6-13 hours for competitors. The approach maintains 36× faster inference (12 ms vs. 156-507 ms) and 85% lower memory footprint, enabling deployment in resource-constrained environments. Seasonal models achieve MAPE of 2.50% (summer) to 16.48% (winter), with weighted aggregation yielding 8.38% overall seasonal MAPE. We acknowledge limitations including climate non-stationarity beyond validated horizons and geographical specificity to hot desert climates. Long-term projections (>10 years) should be interpreted as indicative trends requiring regular updates. This work advances interpretable, efficient solar forecasting while highlighting the importance of rigorous validation and uncertainty quantification for operational deployment. • A novel deep learning model improves long-term PV power forecasting accuracy. • Robust preprocessing with Mutual Information enhances model performance. • Prophet model effectively captures nonlinear and seasonal PV variations. • Seasonal trends are decomposed to reduce overfitting and boost flexibility. • MAPE error drops from 10.18% to 5.42%, outperforming traditional RNN models.