GNN-enhanced temporal patch segmentation and frequency fusion model for robust solar energy production forecasting

Solar energy is a renewable resource with the potential to become the primary global energy source, effectively addressing rising energy demands while reducing reliance on fossil fuels. Accurate forecasting of solar energy production is essential for efficient energy management and for maximizing the utilization of renewable resources. However, traditional time series forecasting models face challenges in capturing both local temporal patterns and global dependencies within solar energy data. Furthermore, these models often overlook the spatial relationships between neighboring solar stations, limiting their ability to learn from interactions across nearby geographic locations. To address these challenges, we propose a novel hybrid model that integrates Graph Neural Networks (GNNs) with the Temporal Patch Segmentation and Frequency Fusion (TPS-FF) approach for robust energy forecasting. The GNN model is designed to learn spatial relationships between different solar stations, enabling the model to leverage interactions accross geographically dispersed locations. Whereas, the TPS-FF mechanism segmentes time-series data into patches to better capture local temporal patterns, while frequency disentangling methods decompose the signal into multiple frequency components, thereby capturing both rapid fluctuations and supporting multi-step ahead forecasting (i.e., 12–72 h). Experimental results demonstrate that the GNN-enhanced TPS-FF model significantly outperforms state-of-the-art forecasting architectures, achieving superior accuracy across various forecasting scenarios and substantially reducing both MAE and MSE metrics. Specifically, the model reduces MSE by 46 %, 34 %, 32.5 %, 30 %, and 21 % across prediction horizons of 12, 24, 36, 48, and 72 h, respectively, while MAE reductions over the same periods are 38 %, 25 %, 27 %, 23 %, and 20 %. • We integrate GNNs into the TPS-FF framework to capture spatial dependencies between solar stations, improving prediction by modeling inter-station interactions. • TPS divides time-series data into overlapping and non-overlapping segments—overlapping captures temporal continuity, while non-overlapping reduces computation and highlights distinct patterns. • Frequency disentangling per patch extracts short- and long-term trends. Multi-head attention then highlights key patterns and captures complex dependencies within each patch. • The enhanced TPS-FF patch aggregation layer captures global context, allowing information sharing across the series for improved long-term solar forecasting. • In our proposed approach, MSE is reduced by 46 %, 34 %, 32.5 %, 30 %, and 21 % across prediction horizons of 12 to 72 h, while MAE sees reductions of 38 %, 25 %, 27 %, 23 %, and 20 % over the same time spans.