A Hybrid Mamba–GNN Framework For Spatio-Temporal Solar Irradiance Forecasting Using Sky Camera And Satellite Data Fusion

Abstract

Solar irradiance forecasting is known to be one of the crucial issues in integrating the use of renewable energy sources into power grids. Solar energy is intermittent, and precise forecasts are crucial for grid stability and efficiency. At present, however, the methods used to forecast such events are often at odds with being able to capture the high-resolution local cloud dynamics and the synoptic scale atmospheric patterns that are present at the time of prediction, especially if computational efficiency is also important. To tackle these challenges, the proposed novel hybrid deep learning framework Mamba-GNN is introduced.For this purpose, a novel hybrid deep learning framework, named Mamba-GNN, is introduced to address these challenges. In this framework, a Vision State Space Model (Mamba) is used for processing high-frequency temporal sequences while a Spatio-Temporal Graph Neural Network (ST-GNN) is used for modeling cloud propagation over the region using the sky camera images from the ground. A dynamic cross-attention mechanism is embedded, which is able to adaptively fuse these heterogeneous features. As illustrated in this paper, the Mamba-GNN is consistently better than ConvLSTM for different forecasting horizons, achieving a forecast skill of 0.85 for 10 min forecast and 0.55 for 3 hrs forecast. Both RMSE and MAE values are significantly decreased in particular in longer durations. In addition, ablation studies validate the effectiveness of its constituent elements and the fusion process and computational study reveals its good accuracy-latency balance, making it suitable for real-time applications.