Geopolitical risks associated with transnational photovoltaic (PV) projects in Africa are dynamic, interconnected, and difficult to assess using conventional investment models. Specifically, existing approaches often struggle to characterize cross-border risk transmission and incorporate evolving geopolitical information into investment decisions. To address these challenges, this study proposes EG-GeoInvest, a framework that integrates generative artificial intelligence (AI) with spatiotemporal graph neural networks for geopolitical risk assessment and risk-aware investment decision-making. First, geopolitical event records, multilingual news, national governance indicators, and PV investment data are combined within an evidence-constrained, multisource heterogeneous information system to support the identification and semantic representation of risk events. Second, a Spatiotemporal Geopolitical Risk Encoding Module (ST-GREM) integrates a temporal convolutional network with relationship-aware graph convolution to represent the evolution of risks and their interdependencies across transnational investment, energy-trade, and financing networks. Third, cross-modal semantic projection and Low-Rank Adaptation (LoRA) enable graph-structured information and generative language models to contribute jointly to forecasts of geopolitical risks, project delays, and economic loss distributions. Finally, the predicted risk distributions inform a Conditional Value-at-Risk (CVaR) and distributionally robust optimization framework for investment allocation under return and tail-risk constraints. A proposed evaluation design encompassing comparative risk prediction, cross-country generalization, and investment scenario analysis provides a basis for assessing predictive performance and potential decision-making value.
