Locally Trained Interpretable Regression Models for Hourly Wind Speed Forecasting in Nigeria: A 3-to 20-Fold Accuracy Improvement over Global Reanalysis

Accurate short-term wind speed forecasting is a major barrier to cost-effective wind and hybrid renewable energy deployment across sub-Saharan Africa. Leveraging 15 years (2008-2022) of hourly 10-m ground-measured wind speed data from the Nigerian Meteorological Agency (NiMET) at stations representing all six geopolitical zones, we develop four transparent regression-based models. The best-performing robust linear regression model (Model 3) yields normalized mean absolute errors (nMABE) of 2.6-10.9% and Nash-Sutcliffe efficiencies (NSE) of 0.84-0.99 on fully independent 2018-2022 test data. In direct comparison with NASA's widely used MERRA-2 reanalysis (at the same 10-m height), Model 3 reduces nMABE by factors of 3.1-19.7 (median 8.2×) and corrects the systematic underestimation seen in MERRA-2 (median bias-49%, negative NSE in every zone). The resulting explicit forecasting equation requires only latitude, longitude, day-of-year, and the previous year's hourly record; inputs already accessible to rural mini-grid operators and planners. These computationally lightweight, fully interpretable models offer an immediately deployable solution for precise wind resource assessment, turbine siting, hybrid system design, and real-time energy management in data-scarce regions throughout Africa.