Probabilistic spatio-temporal machine learning methods for distribution system state forecasting

This paper investigates probabilistic, data-driven methods for short- to mid-term distribution system state forecasting to support proactive grid operation. We compare multiple forecasting methods that integrate spatial and temporal learning, ranging from baseline models based on standard load profiles to fully integrated spatio-temporal neural networks. In particular, we evaluate the A3T-GCN architecture, which combines graph convolutions, gated recurrent units, and temporal attention for multi-horizon state prediction. Using both synthetic SimBench and real low-voltage grids, we assess deterministic and probabilistic forecast accuracy as well as the capability to anticipate critical operating conditions. Our results show that purely temporal models already achieve competitive forecasting performance, capturing a large share of system dynamics. Explicitly modeling spatial–temporal dependencies improves robustness across networks. Among the evaluated methods, the A3T-GCN achieves the most stable performance across grids and forecasting horizons, demonstrating its potential as a transferable framework for reliable, uncertainty-aware state forecasting in future distribution systems.