martin-kagya/uniSolar

Solar power prediction for sub-saharan africa # UniSolar — Bankable Solar Resource Assessment for West Africa Lender-grade solar yield assessment that pairs **NASA POWER satellite irradiance** with a **machine-learning layer that brackets a deterministic physics engine** — correcting the inputs it can genuinely improve, and quantifying the **P90 risk** a lender underwrites on. Part of work on solar resource assessment in data-sparse regions (ZINDI Solar Challenge lineage). --- ## Table of Contents 1. Executive Summary 2. The Core Finding 3. ML-A — Fixing the Irradiance Decomposition 4. Validation — Energy vs Measured Ground Truth 5. ML-B — Calibrated Uncertainty (P50/P90/P99) 6. The 6-Layer Pipeline 7. Data Sources 8. Reproducing the Results 9. Honest Caveats & Limitations 10. What Changed & Why 11. Way Forward --- ## Executive Summary UniSolar evaluates satellite-derived irradiance for West African sites and turns it into **lender-ready energy and financial projections**. The ML is deliberately placed where it adds *defensible* value — not where it can only fit noise: - **NASA POWER GHI is already accurate** — on Tier-1 reference pyranometers its bias is **+1.1 W/m²** and essentially **0% of the error is correctable**. So the ML does **not** chase GHI point-corrections (they only fit sensor bias). - **ML-A fixes the DNI/DHI decomposition.** NASA POWER's three irradiance components are mutually inconsistent — they understate **plane-of-array (POA)** irradiance by **~9%**. A learned separation model restores a physically consistent split and is the only method that improves POA on utility-scale **single-axis trackers**. - **Validated against measured ground truth**, the corrected pipeline lands **within ±2%** of annual energy — after we found and fixed a pre-existing bug that was understating yield by **~45–50%**. - **ML-B delivers a calibrated P90.** Regime-conditional conformal uncertainty is **empirically calibrated out-of-station (P90 coverage 90.6%)** — the exceedance probability lenders size debt …