Work

Forecasting Electricity Demand Under Severe Load Shedding: An Explainable and Statistically Validated Deep Learning Benchmark

Year: 2026
Type: preprint
Abstract: Electricity demand forecasting is harder on grids where supply often falls short, because load shedding changes how demand behaves. This study benchmarks one-hour-ahead forecasting on the national grid of Bangladesh from April 2015 to March 2026, a period in which load shedding rose from less than 1% of hours to 44% of hours in 2022 to 2025. Eleven models were compared under a single leak-free protocol: three naive baselines, two linear models, Random Forest, XGBoost and four deep learning architectures (PatchTST, iTransformer, CNN-BiLSTM and BiGRU). All were scored on the same 14,142 test hours from July 2024 to March 2026, 41% of which had load shedding. PatchTST was the most accurate model, with an RMSE of 203.8 MW and a MAPE of 1.26%. Its RMSE was 34% lower than that of a tuned XGBoost and 58% lower than persistence. Diebold-Mariano tests with autocorrelation-robust variance and paired block-bootstrap intervals confirmed that it was significantly better than every other model. The advantage came mainly from load-shedding hours, where the error of XGBoost rose by 36% while that of every deep learning model changed by about 5% or less. Integrated Gradients and TreeSHAP showed that the transformers used the full weekly demand cycle and the recent shedding history, whereas XGBoost could not predict demand above its training range. BatchNorm improved the transformers but not the recurrent models. Forecasting models for energy-constrained grids should therefore be evaluated separately for hours with and without load shedding.
Language: English
Open Access status: green
OpenAlex ID: W7234792006
Cites: 58
Cited by:
Related to: 10