Application of hybrid deep learning models for seasonal solar irradiance forecasting in the city of Pala

Accurate solar irradiance forecasting is critical for optimizing the design, control, and monitoring of solar energy systems, yet its intrinsic intermittency remains a major challenge. This study proposes and evaluates three novel hybrid deep learning models AME-LSTM, AME-CNN, and AME-GRU all integrating Attentive Meteo-Embedding (AME) with recurrent or convolutional architectures and trained using the Adam optimizer. Using hourly meteorological data (temperature, relative humidity, wind direction) from 2015 to 2023 in Pala, Chad, the models achieve high predictive accuracy across seasons. The AME-GRU model delivers the best overall performance, with a mean absolute error (MAE) of 10.45 W/m 2 in summer and an R 2 of 0.997. AME-CNN also performs strongly in summer and autumn (R 2 = 0.997), while AME-LSTM shows higher errors in winter (MAE = 24.88 W/m 2 , R 2 = 0.984). These results demonstrate that the proposed hybrid models significantly outperform conventional approaches and provide a robust foundation for solar energy forecasting. This work directly supports the development of a hybrid power plant in Pala, contributing to Chad’s energy transition.