Speaker
Description
Abstract
Accurate solar irradiance forecasting is essential for the efficient integration of
photovoltaic systems and sustainable energy management. However, achieving reliable
predictions in data-scarce regions remains a significant challenge due to limited obser
vations and complex environmental variability. This study proposes a hybrid CNN
LSTMmodel for multi-horizon solar irradiance forecasting, aiming to improve predic
tive accuracy and generalization capability in such contexts. The proposed framework
integrates convolutional neural networks (CNN) for spatial feature extraction with long
short-term memory (LSTM) networks for temporal dependency modeling. Historical
meteorological data from Eritrea are used to evaluate the model across multiple fore
casting horizons ranging from 3 to 12 hours.
Comprehensive experiments are conducted to compare the proposed model with
baseline approaches, including persistence models, random forests, and conventional
LSTMarchitectures. The results demonstrate that the CNN-LSTM model consistently
outperforms the baselines, achieving significant improvements in RMSE and R² met
rics. Furthermore, cross-location validation experiments reveal strong generalization
performance when transferring the model across different geographic sites. Additional
analyses, including seasonal evaluation and interpretability assessment using SHAP,
provide insights into model behaviorandfeatureimportance. Alightweightdeployment
prototype is also developed to demonstrate the practical applicability of the proposed
approach.
The findings suggest that the proposed hybrid framework is a robust and effective
solution for solar irradiance forecasting in data-constrained environments, offering both
improved accuracy and enhanced interpretability.
Key Words: Solar Irradiance Forecasting, CNN–LSTM, LSTM, Multi-Horizon
Prediction, Time Series Forecasting, Renewable Energy