Comparative analysis of deep learning architectures in solar power prediction

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Time-Series Forecasting

Background:

• Accurate solar power forecasting is crucial for integrating renewable energy into electricity grids.
• Existing forecasting methods face challenges in reliability and accuracy, especially with complex temporal patterns.
• Deep Learning (DL) offers potential solutions for improving solar power prediction.

Purpose of the Study:

• To conduct a comparative analysis of eight state-of-the-art Deep Learning architectures for solar power prediction.
• To evaluate the performance of DL models using key statistical metrics (RMSE, MAE, MAPE, R²).
• To identify the most effective DL models and provide a framework for real-world energy systems.

Main Methods:

• Utilized a dataset of 4,200 historical solar power records with 20 meteorological and astronomical features.
• Compared eight DL architectures: Autoencoder, LSTM, GRU, SimpleRNN, CNN, TCN, Transformer, and InformerLite.
• Assessed model performance on training, validation, and test datasets using RMSE, MAE, MAPE, and R² metrics.

Main Results:

• The Temporal Convolutional Network (TCN) demonstrated superior performance, achieving a test R² of 0.7786 and a balanced relative standard deviation of 0.6827.
• The Autoencoder model exhibited the greatest overall performance on the entire dataset, with a Whole R² of 0.8437.
• The Transformer model showed significantly poorer performance (Test R² = 0.0714), indicating limitations without modifications.

Conclusions:

• The TCN and Autoencoder are identified as the most effective Deep Learning models for solar power forecasting based on statistical metrics.
• The study provides a scalable, interpretable, and extensible forecasting framework for smart grid applications.
• Findings support the informed integration of DL for enhanced renewable energy management and future hybrid modeling.