Explainable AI model finds Kenyan prepaid power demand is mostly persistence

The research, published open access in Energy Informatics on 5 October 2026, built an explainable forecasting framework for the Kenyan prepaid electricity sector and asked a deliberately provocative question: does the complexity of a modern gradient-boosting model actually buy a meaningful forecasting advantage over far simpler benchmarks?

The answer, on most measures, is only barely.

On the primary test period, XGBoost achieved a mean absolute error of 19.78 kWh and a mean absolute percentage error of 52.31%. OLS came in close behind at 20.03 kWh and 57.07%. Notably, the simpler model actually beat the machine-learning approach on root mean squared error, scoring 37.69 kWh against XGBoost’s 38.35 kWh. Both fitted models did clearly outperform the naïve Lag-1 benchmark, which recorded an MAE of 23.11 kWh, so the statistical models added real value over pure persistence, but the margin separating sophisticated machine learning from a straightforward linear regression was thin.

The explainability layer of the study proved the most revealing. Using SHAP values to interpret the XGBoost predictions, the researchers found that recent consumption history dominated the model’s output. The second-most-recent month (Lag 2) produced a mean absolute SHAP value of 33.16, the most recent month (Lag 1) scored 30.70, and the same month a year earlier (Lag 12) scored 7.84. By contrast, cooling degree days, a meteorological measure of heat-driven energy demand, registered just 0.93, and each tariff variable contributed less than 0.40.

Those numbers point to a striking conclusion: prepaid residential electricity demand in this setting is primarily persistence-driven. Households’ recent consumption patterns, not weather or price signals, carry almost all of the predictive information. That behavioural persistence is precisely why the authors caution that forecasting approaches developed in other electricity-market settings may not transfer directly to prepaid systems, where demand exhibits strong temporal persistence and distinct consumption characteristics shaped by the pay-as-you-go metering model.

The study’s broader lesson concerns validation discipline rather than algorithmic novelty. XGBoost’s improvement over OLS, while statistically supported, did not hold across all metrics, and it also failed to carry through to an April 2026 sensitivity period. A model that looks superior on one error metric or one evaluation window can lose that edge on another, and the authors argue that this fragility underscores the value of rigorous temporal validation and transparent benchmarking alongside model complexity.

The findings land at a time when utilities in rapidly urbanising regions are leaning more heavily on prepaid metering for residential electricity service delivery, making accurate demand forecasting increasingly important for utility planning and energy management.

The study also illustrates why explainable AI tools such as SHAP matter in applied energy forecasting. By quantifying how much each input actually contributes to predictions, they expose when a complex model is essentially rediscovering persistence under a more sophisticated guise. For the Kenyan prepaid grid, the message is clear: the strongest signal in the data is yesterday’s consumption, and any model that cannot beat that simple truth by a wide, robust margin should be treated with healthy scepticism.

Subject of Research: Explainable machine-learning forecasting of urban residential electricity consumption on the Kenyan prepaid grid

Article Title: Predictive modeling of urban residential power consumption: an explainable AI approach to the Kenyan prepaid grid

Article References: Obasanjo, P., Mutinda, A., & Awuor, A. (2026). Predictive modeling of urban residential power consumption: an explainable AI approach to the Kenyan prepaid grid. Energy Informatics. https://doi.org/10.1186/s42162-026-00696-9

Image Credits: AI Generated

DOI: 10.1186/s42162-026-00696-9

Keywords: electricity demand forecasting, Kenya, prepaid metering, XGBoost, SHAP, machine learning, OLS regression, energy informatics, temporal validation, residential power consumption, Predictive, modeling