victorkashumoru/nigeria-solar-forecasting

Solar radiation forecasting for Nigeria using 4 ML models (Holt-Winters, SARIMAX, CatBoost, Ensemble) with NASA POWER data (1984–2022) and spatial error maps. # Nigeria Solar Radiation Forecasting – Full Analysis **Authors:** [Your Name] **Based on:** Furiati et al. (2026), *Ecological Informatics* **Data:** NASA POWER (1984–2022, monthly) **Locations:** 72 points (37 NIMET stations + 32 grid points + 3 supplementary Sahel points) ## Overview This repository contains the complete code, data, and results for a solar radiation forecasting study over Nigeria. We compare four time‑series models: - **Holt‑Winters** (exponential smoothing) - **SARIMAX** (seasonal ARIMA with exogenous variables) - **CatBoost** (gradient boosting on decision trees) - **Ensemble** (simple average of the above three) Two cross‑validation methods are evaluated: **Fixed Start** (expanding window) and **Rolling Window** (fixed 5‑year window). Performance metrics: MAE, RMSE, MAPE. ## Key Results - **SARIMAX** achieved the lowest overall MAE (0.190 kWh/m²/day) under Fixed Start CV, followed closely by Ensemble (0.190) and Holt‑Winters (0.196). - **Ensemble** was the most consistent, ranking first under Rolling Window CV. - **CatBoost** performed well in the Sahel but had higher errors in southern zones. - The Rolling Window method was computationally twice as fast while producing statistically equivalent errors (Wilcoxon test, p > 0.05). - Spatial MAE maps show lowest errors in the Guinea Savanna (central belt) and highest in the Sahel (north) due to greater solar variability. ## Repository Structure