GridFusionX: Network-Aware Probabilistic Forecasting for Multi-Regional Power Systems

Smart grid networks exhibit complex spatial–temporal dependencies where regional nodes are interconnected through electrical transmission, economic coupling, and shared meteorological patterns. Traditional forecasting methods model regions independently, neglecting network topology and spatial correlations that govern system-level behavior. Graph neural networks capture spatial structure but provide only deterministic forecasts, while probabilistic transformers quantify uncertainty yet ignore network topology, preventing risk-aware coordination across interconnected systems. This paper presents GridFusionX, a multimodal transformer model that resolves this tension through theoretically grounded uncertainty quantification in graph-structured systems. We introduce dual-head transformer encoders with asymptotic calibration guarantees, precision-weighted fusion that adapts to spatially varying data reliability, and tight bounds for uncertainty propagation across networked sequences. Unlike prior approaches that sacrifice either spatial awareness or probabilistic rigor, GridFusionX achieves both through convergence-rate guarantees for network-aware uncertainty estimation, optimal precision weighting under modality independence conditions, and calibrated confidence intervals for operational decision-making. Experiments on ten interconnected European regions demonstrate 4.8–55.9% accuracy improvements and 39.5–66.1% reductions in reserve sizing compared to deterministic graph and topology-agnostic probabilistic baselines, achieving 98.4±0.2% reliability with 90.0±1.6% prediction interval coverage. GridFusionX establishes a theoretically grounded framework for probabilistic forecasting in networked infrastructure.