Material-specific photovoltaic systems integrated in smart grid using graph neural networks

The increasing penetration of photovoltaic (PV) systems in smart grids with varied types of loads necessitates advanced, data-driven methodologies for optimal integration and reliable grid operation. This paper proposes a novel Material-Aware Multi-Spatial Temporal Graph Attention Network (MAMSTGAT) that unifies load classification, PV sizing under grid-contractual constraints, and optimal PV siting based on Loss Sensitivity Factor analysis. Distinctively, the framework incorporates material performance specific to emerging PV technologies, such as perovskite and organic solar cells, thereby capturing their enhanced efficiency under low-irradiance conditions. By leveraging advanced graph-based spatial feature extraction and a recurrent attention mechanism for temporal dynamics, MAMSTGAT accurately models the complex spatio-temporal dependencies inherent in power distribution networks. The proposed method is validated on a 33-bus test system using multi-class load profiles and seasonal weather data representative of the United Kingdom. MAMSTGAT increases load-type classification accuracy to 99.7 %—a seven–percentage–point gain over the best competing spatio-temporal model. It also drives the composite loss down to 0.012, 93 % lower than the next-best method. Proposed hypothetical Organic PV systems require up to 16 % less installed capacity (2.79 MW vs 3.30 MW at low contractual limits) yet still deliver 32.3 MWh of annual energy, about 15 % more than their silicon counterparts; even under a stricter contractual limit a 0.868 MW proposed Organic PV array out-produces a 1.03 MW silicon system by 7 %. These findings underscore the potential of material-aware, graph-based approaches for efficient PV planning and integration in modern smart grids.