Explainable AI for Carbon Accountability in Decentralized Energy Systems across Sub-Saharan Africa

Sub-Saharan Africa contributes roughly 3% of global carbon emissions but absorbs a disproportionate share of the consequences. The continent's path out of this injustice runs through decentralized renewable energy: minigrids, solar home systems, and peer-to-peer trading networks that can electrify the 600 million people still without reliable power without locking in decades of fossil fuel dependence. The AI systems currently managing these grids are opaque, single-objective, and built for temperate infrastructure assumptions that do not hold in East Africa's climate and connectivity conditions.

This paper proposes an explainable AI framework combining Long Short-Term Memory models with SHapley Additive exPlanations and Multi-Objective Reinforcement Learning to optimize battery health, energy routing, and carbon accountability simultaneously in decentralized East African energy systems. LSTM-SHAP provides interpretable predictions of battery degradation calibrated for high-ambient-temperature conditions and the irregular charging patterns of pay-as-you-go solar systems. MORL manages the cost-carbon tradeoff across P2P minigrid nodes, optimizing energy flow without black-box decisions that community operators cannot interrogate or regulators cannot audit.

Validated against operational data from Kenyan minigrids and PAYG solar deployments across East Africa, the framework targets a 30% reduction in carbon intensity relative to diesel-hybrid baselines. The system runs on edge hardware deployable at minigrid control nodes without continuous cloud connectivity, and produces explanations designed to be legible to community energy cooperative operators, not only data scientists. Open-source, licensed for adaptation across the continent.

There is a specific irony in the East African energy picture that I keep coming back to. The Lake Turkana Wind Power project, 310 megawatts cutting across Kenya's Turkana corridor, is the largest wind farm in Africa. It pumps clean electricity into the national grid, helping Kenya achieve one of the highest renewable energy shares of any country in the world, around 90% of grid generation. Yet 200 kilometres from those turbines, communities in Lodwar run petrol generators for four hours a night and pay roughly ten times the grid tariff for the privilege. The grid exists. The renewable capacity exists. The gap is in the last mile, and it is measured in diesel fumes.

East Africa's answer to the last-mile problem is the minigrid. Kenya, Tanzania, Uganda, and Ethiopia collectively operate more than 3,000 minigrids, with the Africa Mini-grid Developers Association projecting that 90% of new rural electrification across the continent will come from decentralized sources by 2030. These systems pair solar generation with battery storage, often supplemented by diesel backup, and increasingly route surplus energy across peer-to-peer networks between households and small businesses. The economics are improving. The carbon accounting, for the most part, does not exist.

The AI optimization tools that exist for decentralized energy grids were built for European and North American contexts: stable grid connections, mild ambient temperatures, consumers who interact with energy apps on smartphones, and regulators who think in megawatt-hours not kilowatt-hours. Applied to a rural Kenyan minigrid, those assumptions break in sequence. Battery degradation in 35-degree heat follows a different curve than at 20 degrees. Pay-as-you-go charging patterns, where households top up daily credit via M-Pesa rather than paying a monthly bill, create irregular load profiles that standard demand forecasting handles poorly. And the energy cooperative that manages the local minigrid, often run by a committee of farmers and small traders, cannot act on an AI recommendation it cannot understand.

Africa's carbon accounting problem adds another layer. The continent's Nationally Determined Contributions under the Paris Agreement are among the most ambitious in proportional terms: Kenya has committed to a 32% emissions reduction by 2030. The Africa Carbon Markets Initiative, launched at COP27, targets 300 million verified carbon credits from African projects by 2030. But the minigrid systems generating real emission reductions from displacing diesel have no reliable mechanism to measure, attribute, and verify those reductions in a form that satisfies carbon standard auditors. What is being burned, when, and by how much is either unrecorded or stored in disconnected spreadsheets that cannot be audited at scale.

This project builds the missing layer: an interpretable AI framework that simultaneously optimizes battery longevity, minimizes carbon intensity, and produces auditable explanations of every routing and dispatch decision, designed from the ground up for East Africa's conditions rather than retrofitted from a European smart grid context.

Four phases over 24 months, sequenced around data access realities. East African minigrid operational data is available but fragmented across operators, so Phase 1 requires active partnership development, not just API calls.

The standard argument for applying AI to energy systems is efficiency: smarter dispatch, lower cost, flatter demand curves. That argument works fine in a European smart grid context where efficiency is the binding constraint. In Sub-Saharan Africa, the binding constraints are different, and they make the XAI problem both harder and more consequential.

Start with batteries. A lithium iron phosphate battery rated for 2,000 cycles at 25 degrees Celsius will deliver something closer to 1,200 cycles if it spends its life at 35 degrees, which is the average ambient temperature in parts of northern Kenya and the Rift Valley lowlands. Every early battery replacement adds cost that either bankrupts a minigrid operator or gets passed to households paying 50 to 80 KES per kilowatt-hour for electricity they desperately need. Getting battery health prediction wrong in this environment is not an inconvenience. It determines whether a minigrid survives its first five years.

Now add trust. East Africa's minigrid ownership model is often cooperative. A group of households buys into shared infrastructure, elects a committee, and that committee makes decisions about energy allocation that affect livelihoods. Maize grinding mills, water pumps, refrigeration for small traders, phone charging businesses. If the dispatch algorithm shifts load away from grinding machines at peak morning hours to conserve battery capacity for the evening, that committee needs to be able to explain that decision to the members who couldn't grind their maize. A black-box RL agent that produces no audit trail and no explanation cannot function in that governance structure. The community will override it or stop using it.

The carbon accounting angle is equally specific to Africa. Western XAI energy research is largely about helping utilities meet net-zero commitments in systems that are already majority-renewable. In East Africa, the marginal unit of electricity is a diesel generator. Every kilowatt-hour cleanly dispatched from solar plus battery displaces real diesel combustion with real emissions. The Africa Carbon Markets Initiative needs exactly this kind of verifiable, attributable carbon reduction at scale, and minigrids are where most of it will come from. But you cannot sell a carbon credit without an audit trail, and you cannot build an audit trail from an opaque reinforcement learning policy. XAI is not optional here. It is the mechanism by which AI-optimized minigrids access the financial system that funds their expansion.

The technical research problems that this context creates, hot-climate battery degradation modeling, MORL for high fuel-price-volatility environments, SHAP interfaces for low-literacy users, edge deployment without cloud dependency, are genuinely novel. They are not edge cases of the European smart grid literature. Solving them for East Africa produces insights that apply across the continent and, eventually, to any decentralized energy context where trust, cost, and carbon accountability are all binding constraints at once.