PROFBEN10/Physics-Informed-Neural-Network-for-Solar-Irradiance-Forecasting-in-Nigeria

Embedding thermodynamic laws into a neural network's loss function to produce accurate and physically consistent solar irradiance forecasts across Nigeria's five climatic zones. # Physics-Informed-Neural-Network-for-Solar-Irradiance-Forecasting-in-Nigeria Embedding thermodynamic laws into a neural network's loss function to produce accurate and physically consistent solar irradiance forecasts across Nigeria's five climatic zones. --- ## 📌 Table of Contents - Overview - Problem Statement - Scientific Background - Project Architecture - Dataset - Methodology - Results - Repository Structure - Getting Started - Usage - Key Visualisations - Relevant Research & Institutions - Future Work - References - Author --- ## 🔭 Overview Nigeria sits within the African Sun Belt and receives some of the highest solar irradiance globally — an estimated **427,000 TWh/year** of theoretical solar potential. Yet solar energy accounts for less than **1% of Nigeria's national grid**, partly because existing forecast systems are unreliable and physically inconsistent. This project develops a **Physics-Informed Neural Network (PINN)** that embeds the **Bird Clear-Sky Model** — derived from atmospheric radiative transfer physics — directly into the model's architecture and loss function. The result is a solar irradiance forecasting system that is simultaneously: - ✅ **Accurate** — competitive RMSE and MAE against a pure data-driven baseline - ✅ **Physically consistent** — zero predictions violating the clear-sky upper bound - ✅ **Geographically comprehensive** — evaluated across Nigeria's five distinct climatic zones - ✅ **Operationally deployable** — a 24-hour ahead forecast dashboard with uncertainty quantification --- ## ❗ Problem Statement > *Can embedding atmospheric physics equations directly into a neural network's loss function produce more accurate and physically consistent solar irradiance forecasts for Nigeria's five climatic zones?* Purely data-driven models (standard LSTMs, GBTs, MLPs) suffer from a fundamental flaw: they are **physically unaware**. They can predict: - Solar irradiance at midnight ❌ - Irradiance values **exc …