Forecasting Daily Household Electricity Demand in a Hot-Arid City: A Comparative Benchmark of Local, Global, Deep and Foundation Models on ESHED-38
Abstract
This study benchmarks 24 forecasting configurations on ESHED-38, daily smart-meter consumption for 38 homes in Erbil, Iraq, collected from May 2020 to June 2025 with station weather and a household survey. The configurations span five base families, plus two cross-family ensembles: naive baselines, per-house statistical models, global gradient-boosted trees, global deep networks, and zero-shot time-series foundation models. All are evaluated on the same 54,720 forecasts from 48 weekly origins over a 359-day held-out period, at lead times up to 30 days, using scaled absolute error, the continuous ranked probability score, calibration, and Friedman, Nemenyi and Holm-corrected Wilcoxon tests. The best configurations cut the scaled error of a seasonal-naive forecast by 20 to 25 percent, and the leading models beat it in all 38 houses. Chronos-2, used without training, is significantly the most accurate model one day ahead; at one week it is statistically tied with TimesFM-2.5, Chronos-Bolt and a pre-specified ensemble, and the two ensembles lead at 30 days. Deep networks trained from scratch do not outperform exponential smoothing on average, and the household survey adds negligible skill. Metered demand peaks in winter and falls as temperature rises. Recorded weather did not improve any model run with and without it. A quality screen helps explain why: on 68 percent of dates a temperature value recurs exactly on other dates, in three annual blocks covering almost all of the test period, suggesting that much of the weather record reuses one annual sequence. The value of weather information is therefore treated as unresolved, and the protocol is released as a baseline.
Related
- TRACE: A Reproducible Benchmark for Electricity Price Forecasting with Official Operational Text
- Online Conditional Vine Copulas: Forecasting Electricity Demand
- Combining Weather Forecast Aggregation and State-Space Models for Adaptive Probabilistic Electricity Load Forecasting
- Bi-Mamba-Load: A bidirectional structured state space model with sequence relationship adaptive perception for electricity load forecasting
- Day-Ahead Electricity Price Forecasting Using a Multivariate Group Lasso Method
BibTeX
@misc{neupane2026forecasting,
title = {{Forecasting Daily Household Electricity Demand in a Hot-Arid City: A Comparative Benchmark of Local, Global, Deep and Foundation Models on ESHED-38}},
author = {Bisheshwor Neupane},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.14535},
eprint = {2611.14535},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.14535},
}
RIS
TY - UNPB
TI - Forecasting Daily Household Electricity Demand in a Hot-Arid City: A Comparative Benchmark of Local, Global, Deep and Foundation Models on ESHED-38
AU - Bisheshwor Neupane
PY - 2026
DA - 2026/11/01
PB - arXiv
DO - 10.48550/arxiv.2611.14535
UR - https://arxiv.org/abs/2611.14535
AB - This study benchmarks 24 forecasting configurations on ESHED-38, daily smart-meter consumption for 38 homes in Erbil, Iraq, collected from May 2020 to June 2025 with station weather and a household survey. The configurations span five base families, plus two cross-family ensembles: naive baselines, per-house statistical models, global gradient-boosted trees, global deep networks, and zero-shot time-series foundation models. All are evaluated on the same 54,720 forecasts from 48 weekly origins over a 359-day held-out period, at lead times up to 30 days, using scaled absolute error, the continuous ranked probability score, calibration, and Friedman, Nemenyi and Holm-corrected Wilcoxon tests. The best configurations cut the scaled error of a seasonal-naive forecast by 20 to 25 percent, and the leading models beat it in all 38 houses. Chronos-2, used without training, is significantly the most accurate model one day ahead; at one week it is statistically tied with TimesFM-2.5, Chronos-Bolt and a pre-specified ensemble, and the two ensembles lead at 30 days. Deep networks trained from scratch do not outperform exponential smoothing on average, and the household survey adds negligible skill. Metered demand peaks in winter and falls as temperature rises. Recorded weather did not improve any model run with and without it. A quality screen helps explain why: on 68 percent of dates a temperature value recurs exactly on other dates, in three annual blocks covering almost all of the test period, suggesting that much of the weather record reuses one annual sequence. The value of weather information is therefore treated as unresolved, and the protocol is released as a baseline.
ER -