Dry Reservoirs, Dark Cities

Image Credit: Courtney Lindeque

In 2024 Lake Kariba dropped to 15 percent of its usable storage and Zambia went to 12-hour daily power cuts, because the dam's turbines had too little water to run. The size of the lake can be measured from satellite images every month, going back to 1984. We will build a model that uses the lake's area and the rainfall over its basin in recent months to predict how much electricity the Kariba plant generates, and test how many months ahead a shortfall shows up. We will check the forecast against the plant's reported output and against nighttime satellite images of Lusaka, which dim when the power goes out.

Research questions

1. How well does the lake's area, together with rainfall over the basin in the months before, predict the plant's monthly output one, three and six months ahead?

2. How many months before the 2024 shortfall would the model have raised a warning, and how often would it have raised one when nothing happened?

Datasets

• JRC Global Surface Water monthly history (lake area, 1984 to 2021) and Sentinel-2 and Sentinel-1 images for 2022 onward

• CHIRPS rainfall over the Zambezi basin above the dam

• Monthly generation figures for Kariba North Bank from Zambezi River Authority and ZESCO reports

• NASA Black Marble monthly nighttime lights for Lusaka

• OpenStreetMap power lines and substations, WorldPop population, hospital locations

Machine learning

• A gradient boosting regression model that predicts monthly generation from lake area and rainfall over the previous one to twelve months, tested by holding out whole years

• A simple classifier on the same features that flags the start of a shortfall, with hit and false-alarm rates reported by lead time

GIS

• Outline the reservoir and the drainage basin above the dam (HydroSHEDS), and average each month's rainfall over that basin

• Map the grid area the plant serves from OpenStreetMap lines and substations, and count the people and hospitals inside it

• Pull Lusaka's nighttime-light values for the same months to compare with the forecast

References

• Retief, Kayathri, Vigneswaran, Ghosh, Garcia Andarcia and Dickens (2025). Satellite-surface-area machine-learning models for reservoir storage estimation: regime-sensitive evaluation and operational deployment at Loskop Dam, South Africa. arXiv.

• Shah, Hu, Edelenbosch and van Vliet (2024). Global loss in global hydropower supply under droughts using a hybrid model. EGU General Assembly 2024.

• Zambezi River Authority (March 2024). Press statement on the hydrological situation at Kariba Dam. (LINK ERROR)