Machine Learning Maps Climate Risks to West Africa’s Hydropower Future

West Africa’s dams are the quiet engines of the region’s electricity supply, turning the rainfall that sweeps across the Volta, Niger, and Senegal river basins into power for millions of homes and industries. But that dependence on water makes hydropower one of the most climate-sensitive parts of the region’s energy system, and a new study published in PLOS Climate suggests the risks ahead are both larger and more unevenly distributed than many planners have assumed. Using an ensemble of machine learning models, researchers have projected how changing rainfall and rising temperatures could reshape reservoir inflows and electricity generation at seven major dam basins across West Africa through the end of the century.

The research team, led by Franck Hervé Akaffou and colleagues, set out to solve a persistent problem in climate impact studies: single models tend to capture only part of the complex, lagged relationships between weather and river behavior. Rain that falls in a basin today may take weeks or months to reach a reservoir, and temperature drives evaporation losses that further complicate the picture. To handle these dynamics, the team built a five-step framework that begins with a broad pool of fifteen machine learning algorithms and progressively narrows and combines them into stronger predictive systems.

Central to the approach is what the authors call a multi-lag integration of precipitation and temperature. Rather than feeding a model only the current month’s weather, the framework incorporates rainfall and temperature from preceding periods, allowing the algorithms to learn how antecedent conditions propagate through a watershed into reservoir inflow. This matters enormously in West African basins, where seasonal monsoon dynamics and long hydrological memory mean that a wet season’s signature can appear in dam inflows long after the rains have ended.

The selection process was deliberately competitive. All fifteen candidate models were first evaluated on their ability to reproduce historical inflow and generation patterns, and only the top performers advanced to the next stage. Through iterative ensemble stacking, in which the outputs of strong models are themselves combined and retrained, and through the systematic elimination of weak learners that added noise rather than signal, the team refined its predictions layer by layer. The result was a marked improvement in accuracy and efficiency across the modeling chain, with the coefficient of determination and the Nash-Sutcliffe efficiency, two standard measures of hydrological model skill, both exceeding 0.6 for all inflow simulations and for the energy simulations at the Bagre, Nangbeto, and Taabo dams.

To train and validate the models, the researchers drew on the CHIRPS and CHIRTS datasets, which provide high-resolution satellite-based and station-calibrated records of precipitation and temperature from 1983 to 2014. This three-decade historical window gave the algorithms a rich sample of West Africa’s climate variability, including major droughts and wet years, against which to learn the relationships between climate and dam performance. Once calibrated, the models were turned toward the future using projections from twelve bias-adjusted CMIP6 climate models, the latest generation of global climate simulations, together with their ensemble mean.

The future scenarios spanned three emissions pathways: SSP1-2.6, a low-emissions world consistent with ambitious mitigation; SSP2-4.5, an intermediate pathway; and SSP5-8.5, a high-emissions trajectory with continued reliance on fossil fuels. Projections were made for two future periods, a near future from 2036 to 2067 and a far future from 2068 to 2099, allowing the team to distinguish changes likely to emerge within the working lifetime of today’s dams from those that may arrive by century’s end.

The climate projections themselves are stark. Under the highest emissions scenario, temperatures across the study basins could rise by as much as 4.5 degrees Celsius, a level of warming that would sharply increase evaporation from reservoir surfaces and alter the intensity and timing of monsoon rainfall. Precipitation changes, by contrast, are spatially heterogeneous: some basins are projected to become wetter while others dry out, and the direction and magnitude of change vary across scenarios and time horizons. This patchiness is precisely why basin-by-basin machine learning assessments are valuable, since a single regional average would obscure the divergent fates of individual dams.

For reservoir inflows, the projections reveal a region divided. Buyo faces the steepest declines, with inflows projected to fall by up to 24 percent, while Nangbeto could see reductions of up to 13 percent. Reduced inflow means less water passing through turbines, and the knock-on effects for electricity generation are severe: hydropower output at Nangbeto may decline by up to 19 percent, and at Taabo the projected drop reaches a remarkable 58 percent, a loss that would ripple through national grids that rely on these plants for a substantial share of their supply. Yet the picture is not uniformly grim. Manantali and Taabo are projected to experience inflow increases, and Bagre may actually see energy gains of up to 42 percent under the high-emissions scenario, illustrating how climate change can create winners as well as losers even within a single region.

These contrasting outcomes carry important implications for how West African countries plan their energy futures. A dam projected to gain generation capacity may be able to shoulder more of the regional load, but planners cannot simply reallocate risk from one basin to another, because the projections are scenario-dependent and the high-emissions pathway that produces the largest gains at Bagre also produces the most dangerous warming elsewhere. The authors argue that the findings underscore an urgent need for adaptive management strategies: strengthening the resilience of hydropower systems themselves, diversifying national and regional energy portfolios, and integrating additional renewable sources such as solar, which is abundant across the Sahel and largely uncorrelated with hydrological variability.

For hydropower managers and policymakers, the message is that proactive measures must begin now, well before the mid-century projections materialize. Reservoir operating rules calibrated to a stable twentieth-century climate may become maladaptive as inflow patterns shift, and the long lead times for infrastructure investment mean that decisions made this decade will determine whether the region’s power systems bend or break under climate stress. By demonstrating that machine learning ensembles can skillfully link climate projections to dam-level outcomes, the study offers a practical template for the kind of granular, basin-specific risk assessment that energy security in a changing climate will increasingly demand.

Subject of Research: Climate change impacts on reservoir inflow and hydropower generation in West African dam basins assessed with ensemble machine learning

Article Title: Integrating machine learning to assess climate risks on reservoir’s inflow and hydropower generation across West African basins

Article References: Akaffou, F. H., Obahoundje, S., Diedhiou, A., Kouassi, K. L., Diallo, D., Amoussou, E., Yamegueu, D., & Ofosu, E. A. (2026). Integrating machine learning to assess climate risks on reservoir’s inflow and hydropower generation across West African basins. PLOS Climate, 5(9), e0000986. https://doi.org/10.1371/journal.pclm.0000986

Image Credits: AI Generated

DOI: 10.1371/journal.pclm.0000986

Keywords: hydropower, machine learning, climate change, West Africa, CMIP6, reservoir inflow, ensemble modeling, SSP scenarios, CHIRPS, energy security, adaptive management, PLOS Climate