Abstract: Solar radiation forecasting is a key task for energy planning, grid management, and photovoltaic deployment, especially in tropical regions where weather variability reduces operational reliability. This work applies deep learning techniques to forecast hourly solar radiation in Mompox, Colombia, using Long Short-Term Memory (LSTM) neural networks. Three temporal windows were studied (5, 24, and 720 hours) to examine how sequence length affects prediction accuracy and model behavior. Hourly radiation data from 2021 to 2022 were used for training, and independent datasets from 2023 to 2024 were used for external validation to ensure long-term assessment and reproducibility. Most existing studies use short input windows designed for mid-latitude environments (5–24 hours), which do not capture multi-day tropical cloud persistence or sub-seasonal radiation variability. This gap limits forecasting accuracy and restricts practical use in tropical energy planning. To address this issue, this study introduces a long temporal input design that allows the model to learn month-scale variability more effectively. The three network configurations were trained under the same settings, allowing a direct comparison between short, daily, and long input memories. The LSTM-720 model performed best, achieving the lowest RMSE and the most stable predictions across all validation years, showing its ability to reconstruct both diurnal cycles and broader seasonal dynamics. Unlike most solar forecasting work, which treats window size as a tuning parameter, this study introduces a long-context LSTM design based on a 720-hour sequence. This allowed the model to learn intra-month atmospheric persistence—an essential tropical feature that short windows cannot represent—positioning the approach as a methodological contribution that expands the temporal learning paradigm rather than a configuration adjustment. Time-series comparisons revealed close agreement between measured and predicted radiation, particularly during stable climate periods. The proposed framework can support practical applications in solar plant design, renewable energy scheduling, and operational grid strategies in tropical regions. Future work will integrate satellite information and hybrid deep learning architectures to enhance spatial transferability and long-term forecasting accuracy.

Manuel Ospina, Gabriel Chanchí and Álvaro Realpe. “Deep Learning Approach for Solar Radiation Forecasting in a Tropical Region Using LSTM Networks”. International Journal of Advanced Computer Science and Applications (IJACSA) 16.12 (2025). http://dx.doi.org/10.14569/IJACSA.2025.0161263

@article{Ospina2025,
title = {Deep Learning Approach for Solar Radiation Forecasting in a Tropical Region Using LSTM Networks},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2025.0161263},
url = {http://dx.doi.org/10.14569/IJACSA.2025.0161263},
year = {2025},
publisher = {The Science and Information Organization},
volume = {16},
number = {12},
author = {Manuel Ospina and Gabriel Chanchí and Álvaro Realpe}
}