Abstract

Climate change poses significant challenges to renewable energy systems, particularly solar power, which is highly sensitive to meteorological variations. This study investigates the impact of climate change on solar power production using a hybrid CNN-LSTM deep learning model, designed to capture both spatial and temporal dependencies in historical and projected climate data. Leveraging a purely numerical dataset comprising variables such as temperature, solar irradiance, cloud cover, and aerosol concentrations, the model integrates convolutional neural networks (CNNs) to extract localized spatial features and long short-term memory (LSTM) networks to analyze temporal trends. Trained on multi-decadal climate records and validated against regional solar energy output data, the framework predicts future solar power generation under varying climate scenarios (e.g., RCP 4.5 and RCP 8.5). Results indicate a measurable decline in solar efficiency in regions experiencing increased cloud cover and particulate pollution, while arid zones show resilience due to stable irradiance levels. The CNN-LSTM model achieves robust performance of 8.2% RMSE and 6.5% MAE compared to standalone LSTM and CNN baselines, demonstrating its efficacy in handling complex spatiotemporal interactions. However, uncertainties persist due to variability in climate projections and data granularity. This work underscores the need for adaptive energy planning and highlights regions at risk of solar intermittency, offering actionable insights for policymakers and renewable energy stakeholders.

Key Contributions

• Introduced a novel CNN-LSTM hybrid architecture to capture spatiotemporal climate dependencies affecting solar production.
• Trained and validated on historical climate records across different Representative Concentration Pathways (RCPs).
• Achieved high prediction accuracy with RMSE of 8.2% and MAE of 6.5%, outperforming individual CNN and LSTM models.
• Integrated explainable AI tools like LIME for interpretable climate impact assessments.
• Highlighted at-risk regions and provided practical recommendations for adaptive solar infrastructure planning.

Conclusion

This study offers a significant step forward in understanding the future of solar energy under climate stress. The model's robustness and generalizability make it suitable for deployment in forecasting frameworks guiding sustainable energy policy and investment planning.

Limitation

While the CNN-LSTM model demonstrates high accuracy, its reliance on historical data from specific regions introduces geographic bias. Moreover, climate projection uncertainties and limitations in resolution may affect output consistency at the micro-regional level.

Future Work

• Incorporate satellite image data and dynamic meteorological inputs for real-time adaptation.
• Expand testing to tropical and sub-Saharan zones underrepresented in the current dataset.
• Deploy a cloud-based interactive dashboard for live climate impact simulation and forecasting.