Abstract:

This paper proposes a hybrid deep neural model that combines Empirical Mode Decomposition (EMD), Bidirectional Gated Recurrent Units (BiGRU), and Multi-Layer Perceptron (MLP) for hourly forecasting of Global Horizontal Irradiance (GHI) in Tamanrasset, Algeria, a region characterized by highly stable solar conditions due to its arid desert climate. The proposed EMD-BiGRU-MLP model integrates two complementary modeling branches, each capturing distinct information from different input sources: measured GHI values and Numerical Weather Prediction (NWP) forecasts. The first branch processes measured GHI sequences using EMD, followed by a BiGRU model, while the second branch applies another BiGRU model directly to selected NWP variables identified through Pearson correlation coefficient analysis. The outputs from both branches are fused through an MLP network to produce the final forecast. Experimental results indicate that the NWP-based model achieves the best performance (R² = 0.99), while the EMD-BiGRU model and the full hybrid architecture perform comparably, with R² values of 0.938 and 0.94, respectively. These results highlight the effectiveness of NWP forecasts in stable environments. They also demonstrate that hybrid architectures, such as EMD-BiGRU, are valuable for capturing latent temporal structures. Furthermore, the findings emphasize the potential of integrating data-driven and physically modeled information to improve global solar irradiance forecasts of the proposed EMD-BiGRU-MLP model, particularly in stable regions like Tamanrasset. This underscores the importance of adapting model complexity to data characteristics, even when variability appears low.

Publisher's Version