Scenario-Contextualized Short-Horizon Electricity Demand Forecasting for Istanbul: A Multi-Model Assessment for Emergency Power-System Planning

Major earthquakes can simultaneously interrupt routine electricity use and concentrate emergency-service requirements. This study reports a 22-method retrospective comparison of short-horizon hourly electricity-demand forecasting for Istanbul in a hypothetical Mw 7.5 disaster-information context. The Istanbul hourly electricity-demand series was obtained from EPİAŞ for the period 1 January 2020–1 January 2025 and is retained in the author-supplied workbook, which contains 43,968 records, including 96 repeated timestamps. The historical demand observations were used directly; no national-to-Istanbul scaling by 0.15 was applied. No earthquake-induced load reduction or critical-load increment is independently documented in the target. For the 24 h of 1 January 2025, the archival LSTM predictions yield MAE 79.89 MWh, RMSE 104.01 MWh, R2 approximately 0.97 and MAPE 1.65%. A separate information-removal experiment has a full-feature reference MAE of 666.2882 MWh; its RMSE–R2 relationship is compatible with an 8794 h historical holdout rather than the 24 h local comparison. Nine California/CAISO forecast series were reconciled on 24 common target observations; Random Forest had the lowest error (MAE 661.89 MWh; R2 0.7220), whereas the TensorFlow neural additive method was less accurate. An independently developed chronology-controlled demand-only study is also reported separately. The analyses quantify forecast errors, not observed earthquake effects, network adequacy, or validated operational restoration.