A comparative study of deep learning approaches for real-time solar irradiance forecasting

Abstract

Accurate forecasting of Global Horizontal Irradiance (GHI) is critical for enhancing both grid stability and the efficiency of solar energy systems. A comparative assessment of several deep learning models is presented in this study for real-time GHI forecasting, specifically Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and a hybrid LSTM-GRU architecture. Approach performance is evaluated using standard metrics, including MAE, RMSE, and the R². Findings indicate that while GRUs are computationally efficient, they struggle to maintain long-term temporal dependencies. In contrast, LSTMs effectively capture these dependencies, resulting in improved forecasting accuracy. Notably, the hybrid LSTM-GRU model outperforms the individual architectures, achieving the lowest MAE (12.931), RMSE (21.825), and the highest R² (0.996), thereby demonstrating superior predictive performance. These results highlight the potential of the hybrid model in real-time solar energy applications, improving forecast reliability and grid stability. This study advances solar irradiance forecasting methodologies, thereby facilitating the integration of renewable energy sources and improving the effectiveness and reliability of grid operations.