GIST-Grid: Graph-Imputed Spatiotemporal Transformer for Short-Term Load Forecasting on Data-Scarce Nigerian Distribution Networks

Accurate short-term load forecasting is critical for Nigerian power networks but severely hindered by missing data, irregular sampling, and incomplete topology. We propose GIST-Grid (Graph-Imputed Spatiotemporal Transformer), a framework explicitly designed for data-scarce infrastructure that integrates a hybrid spatial graph fusing partial topology with geographic proximity and historical correlation, self-supervised masked load modelling for robust learning under missing data, continuous-time encoding for irregular timestamps, and Monte Carlo dropout for calibrated prediction intervals. Evaluated under 40-60% data sparsity, GIST-Grid reduces MAE by 27-42% and RMSE by 30-44% over LSTM and GCN baselines at 50% sparsity, with error varying under 6% across sparsity levels compared to LSTM's over 30% fluctuation; ablations identify masked pretraining as the dominant driver of this robustness. A distillation-and-quantization pipeline targets deployment on edge-grade substation hardware, demonstrating that efficient spatiotemporal deep learning can deliver reliable AI for resource-constrained African infrastructure.