Graph state-space models for spatio-temporal renewable energy forecasting

Structured state-space models have shown promising results for temporal prediction of demand in renewable energy systems. However, such standard models consider each time-series individually, neglecting spatial and logical correlations between them. In real-world applications, forecasting models must also balance predictive accuracy with practical constraints such as computational cost, available resources, and deployment feasibility. As a solution, in this paper we propose graph-based extensions of structured state-space models for prediction tasks in renewable energy datasets, represented by solar and wind power plants. In order to explore the effectiveness of state-space mechanisms in this context, we adopt the Spatial-Temporal Graph Mamba model, which demonstrates excellent predictive performance but involves a relatively high computational cost due to its complexity. To address this limitation, we also introduce two lightweight and computationally efficient models, namely miniGConvLSTM and miniGConvGRU, which effectively combine the benefits of state-space modeling with graph-based processing. These models aim at striking a better balance between accuracy and resource efficiency, making them particularly suitable for real-world forecasting scenarios. Finally, we discuss the implications of these findings for future research and the potential for integrating graph-based state-space models into real-world energy management systems.