Comparative analysis of deep learning architectures in solar power prediction

Integrating renewable energy sources into the electricity grid requires accurate forecasts of solar power production. With the aim of enhancing the accuracy and reliability of forecasts, this study presents a comprehensive comparative analysis of eight state-of-the-art Deep Learning (DL) architectures-Autoencoder, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Simple Recurrent Neural Network (SimpleRNN), Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), Transformer, and Lightweight Informer for Long Sequence Time-Series Forecasting (InformerLite)-applied to solar power prediction using a dataset with 4,200 historical records and 20 meteorological and astronomical features. A comprehensive assessment of Root Mean Squared Error [Formula: see text], Mean Absolute Error [Formula: see text], Mean Absolute Percentage Error [Formula: see text], and Coefficient of Determination [Formula: see text] metrics was performed on the training, validation, and test datasets. The TCN model had the greatest performance across all models, achieving a test R² of 0.7786, an [Formula: see text] of 429.4863, and a balanced relative standard deviation ([Formula: see text]) of 0.6827, so exhibiting an exceptional capacity to capture temporal patterns. The Autoencoder achieved a [Formula: see text] of 0.7648 and had the greatest overall performance on the entire dataset, resulting in a Whole [Formula: see text] of 0.8437. In contrast, the Transformer model demonstrated significantly poorer performance (Test [Formula: see text] = 0.0714), underscoring its limitations in this context without any architectural modifications. This study not only demonstrates the best DL models for solar power forecasting as qualified by useful statistical metrics, but also provides a scalable, interpretable, and extensible forecasting framework for real-world energy systems. The findings verify the informed DL integration to smart grid scenarios, laying the foundations for further developments in hybrid modeling, multi-horizon prediction, and deployment in resource-constrained environments with limited computational power and resources.