Predictive Urban Scale Digital Twins For Optimising Solar Panel Energy Yield

The adoption of smart, renewable energy systems in developing regions is often challenging due to the cost associated with obtaining proprietary spatial data and dependence on forecasting models that do not adequately capture localised microclimate phenomena. We present a physics driven AI digital twin to optimise solar energy infrastructure planning using an educational institution as a case study for the pilot application in this research work. Current solar assessment methodologies often isolate spatial mapping from temporal forecasting. They also utilise machine learning models that neglect thermodynamic laws and in some instances, they ignore region specific energy drawbacks such as periodic harmattan dust soiling, and extreme thermal degradation. We resolve these limitations by integrating open source geospatial intelligence (OpenStreetMap) and physics constrained machine learning. The methodology utilises a modular physics engine to calculate metrics such as building level shading factors and baseline thermal degradation. These physical representations serve as the features for an optimised LightGBM gradient boosting predictive model trained on hourly ERA5 meteorological data, to accurately forecast 8,760 hour temporal energy yields. We also deploy a spatial nearest neighbour algorithm to triangulate unnamed structures with local road networks/landmarks to generate an automated operational directory. The resulting framework is an interactive 3D WebGL digital twin which includes financial and environmental dashboards that can be localised based on the region. The model quantifies the exact energy losses attributed to seasonal dust or heat and the deployment strategy targets the highest yield, lowest shading facilities.