What Predicts Dumsor? A Modeling Analysis of Satellite-Based Power Outage Detection in Accra, Ghana

Abstract

Unreliable electricity supply, known locally as ”dumsor,” remains a persistent chal- lenge in Ghana. The Electricity Company of Ghana (ECG) maintains internal fault reporting systems, but these records are fragmented and sparse which make it difficult to map outage patterns at scale. This thesis investigates whether integrating meteoro- logical data and historical trends into a satellite-based detection pipeline can improve the identification of power outages in Accra. Building on the framework of Shah et al. (2022), nightly radiance observations from NASA’s VIIRS satellite is paired with ground-truth outage labels from 147 GridWatch sensors across the Greater Accra re- gion spanning 2019 through mid-2023. Spatially aggregating individual sensor sites into 0.01-degree patches produced the single largest improvement in detection per- formance. Nine models consisting of logistic regression, random forest, and XGBoost were then trained, benchmarked against a z-score threshold baseline, with weather features time-matched to the satellite overpass at roughly 1 AM local time. Weather variables contributed modest but consistent gains across all algorithms and historical outage rate emerged as the most valuable individual predictor. An STL decomposition approach confirmed that weather features already encode much of the seasonal variation STL targets. The best-performing model was a random forest combining weather, temporal, and historical features with an AUC of 0.8346. Beyond detection, the cross-model comparison serves as a cross-sectional study of how different feature types contribute to satellite-based outage classification. Overall, this thesis demonstrates that meteorological covariates can support outage detection in data-scarce power systems and highlights the potential for similar approaches in other regions with limited grid monitoring infrastructure.