kudumens/capstone-solar-ghana

Forecasting of Solar Irradiance and Reliability-Constrained Off-Grid Solar Sizing across Ghana’s Climatic Zones using Machine Learning # Machine-Learning Forecasting of Solar Irradiance and Reliability-Constrained Off-Grid Solar Sizing Across Ghana's Climatic Zones QM 640 Data Analytics Capstone — Walsh College / Great Learning (MSc) Author: Richard Adjei-Mensah · Term 3 - 2026 ## Overview This project (1) develops and statistically compares classical (persistence, day-of-year climatology, and SARIMA-class dynamic harmonic regression) and machine-learning (random forest, XGBoost, LSTM) models forecasting daily global horizontal irradiance (GHI) at five sites spanning Ghana's coastal, forest, and savannah climatic zones, and (2) converts the same twenty-year irradiance record into a reliability-constrained off-grid sizing model that computes, per climatic zone, the PV array and battery capacity required to serve a household load at target loss-of-load probabilities (LOLP). ### Research questions - **RQ1** — How accurately can ML models (random forest, XGBoost, LSTM) forecast daily GHI across Ghana's climatic zones, and do they significantly outperform classical approaches (persistence, day-of-year climatology, dynamic harmonic regression) at the 1-day and 7-day horizons? - **RQ2** — Which meteorological variables are most strongly associated with daily solar-irradiance variation, and which contribute most to predictive performance? - **RQ3** — How does forecasting accuracy differ between the wet and dry seasons (including the Harmattan) and across the north–south climatic gradient? - **RQ4** — What PV array + battery capacity does an off-grid household need per climatic zone to serve 100/200/300 kWh/month at 1–5% LOLP? ## Key results (test window 2023–2024, single evaluation) Best ML model per site (frozen on 2021–2022 validation) vs. the strongest classical baseline (DHR), 1-day horizon: | Site | Best ML | Test RMSE (kWh/m²/day) | Skill vs. DHR | Holm p (one-sided HLN DM) | | --- | --- | --- | --- | --- | | Navrongo | LSTM | 0.845 | +0.4% | .365 | | Tamale | Random forest | 0.828 | +2.7% | …