Grid-aware spatio-temporal graph neural networks for multi-horizon load forecasting

Modern power systems are intricate webs of interconnected components that must operate in harmony to ensure optimal grid performance. As renewable energy penetration increases, accurate short-term load forecasting becomes ever more critical for maintaining reliability. We introduce a hybrid Space-Then-Time spatio-temporal graph neural network that separates spatial and temporal learning into distinct stages. First, a multiscale graph attention network encoder transforms the grid topology and operational constraints such as line capacity, efficiency, length, and carrier type into rich spatial embeddings. These embeddings, combined with dynamic load and exogenous features, are then fed into temporal models to predict 1-, 6-, and 24-hour horizons. This modular design ensures that the temporal models operate on context-aware representations of the system state. We evaluated our method using three years of Brazilian state-level electricity data and benchmark it against state-of-the-art temporal and joint Space-And-Time baselines. Across all horizons, our approach achieves lower error metrics while matching or surpassing the baselines in runtime, memory, and parameter efficiency. The results show that decoupling spatial and temporal learning, combined with grid-aware modeling, improves accuracy and robustness—emphasizing that load forecasting is more than just a time series problem.