Accurate peak demand forecasting is critical for power planning as underestimation leads to blackouts and load shedding, while overestimation results in stranded capital and inflated tariffs. Yet common approaches, deterministic extrapolation, machine learning, and time-series models, are ill-suited to small, non-stationary, data-constrained settings typical of developing economies. This paper develops a non-stationary Generalized Extreme Value (GEV) framework for forecasting annual peak electricity demand, in which the location parameter evolves linearly with time to capture systematic trends in extremes. To address the well-documented instability of shape-parameter estimation in small samples, we adopt a hybrid two-step strategy: the shape parameter is estimated via L-moments for robustness, while trend and scale parameters are estimated via maximum likelihood. Projection uncertainty is quantified through non-parametric bootstrap resampling. Applied to 31 years of annual peak demand data from Bangladesh, the model identifies a statistically significant upward trend and generates T-year return level projections through 2050, with bootstrap-derived 95 per cent confidence intervals. The resulting central estimates are substantially lower than existing official projections, which rely on deterministic extrapolation without uncertainty bounds. This study offers the first application of non-stationary extreme value theory to long-run electricity demand planning, providing a statistically defensible, transferable, and uncertainty-aware methodology for capacity planning in small-sample, data-constrained developing-country contexts.

Authors: Atikuzzaman Shazeed, Abrar Ahammed Bhuiyan, Khondaker Golam Moazzem
Publication Period: July 2026