Accurate wind and solar power forecasting is crucial for grid stability and renewable energy integration, especially in island like Jeju Island, Republic of Korea. This paper tackles challenges such as nonlinear patterns, spatial interdependencies, and resource variability. Current models are limited because they are resource specific, treat spatial and temporal dependencies separately, and lack dynamic inter-site relationship modeling, which are essential in dispersed systems. To overcome these limitations, we propose a spatio-temporal hybrid graph-Transformer (ST-HGT) model for high-fidelity renewable energy forecasting， which combines graph attention networks (GATs), temporal convolutional networks (TCNs), neural basis expansion analysis for interpretable time-series forecasting (N-BEATS), Transformer， and Informer, within a modular architecture. The proposed ST-HGT model captures spatio-temporal dynamics by learning inter-site correlations from weather and power data via a dynamic graph and employing complementary temporal modules to model local features, long-range dependencies, and complex patterns. The proposed ST-HGT model is trained and evaluated independently for wind and solar power forecasting, using real-world datasets from multiple stations across Jeju Island. The results show that the proposed ST-HGT model markedly surpasses state-of-the-art baselines, including Informer, temporal fusion Transformer (TFT), N-BEATS, and recurrent networks. For solar power forecasting, the proposed ST-HGT model achieves a root mean square error (RMSE) of 21.71 kWh and a coefficient of determination R² of 0.999, with an improvement of more than 96.6% in RMSE relative to the best baseline model. For wind power forecasting, the proposed ST-HGT model achieves an RMSE of 853 kWh and an R² of 0.99, outperforming all baselines, which have RMSEs exceeding 65000 kWh.