Stochastic Planning for African Power Grids Cuts Projected Blackouts by 78 TWh — If You Model Uncertainty Instead of Ignoring It

SDDP model for Ethiopia and Zimbabwe shows deterministic planning massively underestimates blackout risk. Carbon pricing adds only 3-5% to costs when paired with concessional solar finance.

Imagine you're planning a road trip across a region where gas stations sometimes run dry and weather reroutes you without warning. You could plan the cheapest route assuming everything goes perfectly — or you could build in fallback options at each decision point, accepting slightly higher baseline cost for dramatically lower odds of being stranded. This paper argues sub-Saharan Africa's power planners are doing the former when they should be doing the latter. The committed claim: a multistage stochastic generation expansion model solved with SDDP, applied to real Ethiopian and Zimbabwean power systems over 2025-2050, reduces cumulative expected unserved energy by roughly 78 TWh compared to the standard deterministic least-cost plan. That is not a marginal improvement — it is the difference between chronic rolling blackouts and a system that can absorb demand shocks and drought years. The model explicitly represents two sources of uncertainty that deterministic planners hand-wave away: stochastic electricity demand growth and country-specific Markov-chain hydrological variability (how much water is behind the dam in any given year). The architecture sits in the SDDP family — stochastic dual dynamic programming, a decomposition technique for multistage linear stochastic programs with a large number of scenarios. SDDP avoids the curse of dimensionality that makes brute-force scenario trees intractable by building piecewise-linear approximations of the cost-to-go function through iterative forward and backward passes. The model adds technology-specific construction lead times (you cannot decide to build a dam this year and have it online next year) and endogenous reliability constraints. This is computationally mature tooling — SDDP has been used in hydrothermal scheduling in Brazil since the 1990s — but its application to long-horizon generation expansion planning in sub-Saharan Africa with coupled demand and hydro uncertainty is where the novelty sits. The country results tell different structural stories. Ethiopia's least-cost pathway is hydro-dominant with wind and solar providing increasing diversification over time — the country is sitting on enormous untapped hydro potential but is correspondingly exposed to hydrological risk. Zimbabwe's optimal path expands solar, hydro, and gas while inherited coal capacity naturally retires — a brownfield transition rather than a greenfield build. The paper then layers carbon pricing on both systems: a carbon tax of unspecified level increases resource costs by only 3-5%, and when combined with concessional solar financing, actually delivers the lowest average annual emissions while keeping total resource costs below the no-policy baseline. That last result is the policy punchline — concessional finance plus carbon pricing is cheaper than doing nothing. The integrity picture is mixed. Both country applications use real system data, but validation is entirely model-on-model: the stochastic plan is compared to the authors' own deterministic expected-value plan, not to an independent planning authority's actual investment schedule or a competing commercial model (like PLEXOS or OSeMOSYS runs by other teams). The 78 TWh unserved-energy reduction is a model output, not an observed reduction. Benchmark selection appears reasonable but not pre-registered, and code availability is not stated. The milestone question for this line of work is adoption. The technical contribution is solid but incremental within the SDDP literature — the real value unlocks when African power utilities and development finance institutions actually use stochastic rather than deterministic planning tools for investment decisions. The gap between 'published model' and 'embedded in procurement processes at Ethiopian Electric Power or ZESA' is measured in institutional capacity, not algorithmic innovation. The obvious next experiment the authors did not run: applying the model to a multi-country interconnected system (the Eastern Africa Power Pool or Southern African Power Pool) where cross-border trade provides an additional hedge against country-specific hydro risk. Likely reason: data access and model complexity — interconnected multi-country stochastic expansion planning is a substantially harder problem and probably their next paper.