Transformers Excel in Multi-Level Electrical Load Forecasting Benchmark

Accurate electrical load forecasting is critical for the future of smart grids, spanning from aggregated control area forecasts to individual end-consumer predictions. This research presents a comprehensive benchmark evaluating ten short-term load forecasting methods across three distinct grid levels: a transmission system operator (TSO) control area, low-voltage grid feeders, and individual end consumers. The study utilized three diverse datasets to represent these levels. The findings demonstrate that Transformer-based approaches consistently deliver superior performance, reducing forecast error by a significant 6.6% to 10.7% compared to established methods. While a flexible Transformer architecture, YAformer, was introduced and optimized, the standard Transformer model ultimately achieved the best results, suggesting that complex architectural modifications may not always be necessary. The benchmark also evaluated Chronos-2, a Transformer-based time-series foundation model, which showed competitive zero-shot performance on two datasets but struggled with special events in TSO data. Detailed analyses underscore the critical role of long input contexts, relevant covariates, and continuous retraining for achieving high-accuracy load forecasts, aspects often overlooked in time-series forecasting literature.