Temperature-indexed extreme demand protocols for developing-economy power systems: An EVT-LSTM decision-support framework for GRIDCo Ghana

Developing-economy power systems face rising exposure to extreme electricity demand as load growth, limited reserve margins, and climate-sensitive cooling demand interact. This paper develops a temperature-indexed EVT-LSTM decision-support framework for translating extreme-demand forecasts into operational reserve-activation protocols for Ghana Grid Company. Using 78,912 hourly electricity-demand observations for Ghana from 2016 to 2024, the framework combines Peaks-over-Threshold extreme value modelling, LSTM-based 24-hour-ahead forecasting, calibrated exceedance probabilities, and operational alert rules. The resulting protocol links three temperature triggers to demand-at-risk levels, return-period benchmarks, and probability overrides. Back-testing shows that the probability override increases the Level 2 alert hit rate from 76.4% to 88.7%, while the full protocol generates an estimated USD 46.9 million net present value over ten years under the reported cost-benefit assumptions. The results show how probabilistic machine-learning forecasts can be converted into auditable grid-management actions for reserve procurement, outage prevention, and resilience planning in data-constrained power systems.