From reservoir inflow prediction to increasing hydropower generation: a Machine Learning-based FIRO strategy in Southern Africa

Seasonal streamflow forecasting is vital for climate-resilient water management, particularly in hydroclimatically volatile regions like southern Africa, where global teleconnections and local conditions strongly influence hydrologic variability. However, existing forecast systems often struggle to capture the evolving and compound impacts of these climate drivers at the catchment scale, ultimately producing forecast products with limited skill at the sub-seasonal to seasonal time scale. Here, we address this gap by using Machine Learning (ML) methods to enhance seasonal streamflow predictability and to translate forecast skill into tangible operational value. Our forecast model builds on the Nino Index Phase Analysis for extracting preseason sea surface temperature anomalies linked to the most influential climate teleconnections. It also incorporates a Neural Granger Causal model to select causally relevant predictors and generate seasonal inflow forecasts for Lake Kariba in the Zambezi Watercourse. Using El Nino Southern Oscillation and Pacific Decadal Oscillation as the dominant sources of predictability, our ML-based model largely outperforms the skill of two state-of-the-art global forecasting systems (GloFAS and WW-HYPE), particularly in capturing interseasonal variability and hydrological extremes. Finally, all forecasts are used to design forecast-informed reservoir operation (FIRO) policies targeting the maximization of average and firm hydropower generation at Kariba. Using our ML-based forecasts, hydropower production at Kariba increases by 4 GWh/year—an economic gain of $320,000 annually—demonstrating the practical benefits of integrating advanced ML tools into water systems management.