Shadow-aware rooftop PV model

Shadow-aware rooftop PV model

Spatiotemporal deep learning that predicts shading and rooftop PV potential about 158 times faster than simulation.

Model structure

The model takes a height map of the surrounding buildings and the sun position, and predicts the shading on each roof for every hour of the year. A temporal branch then turns hourly shading and weather into PV output.

Why it is fast

Ray-traced shading has to be recomputed for every hour and every building. The learned model replaces that step, which is what makes city-wide runs about 158 times faster than simulation.

Inputs and outputs

  • Inputs: building heights, roof geometry, sun position, hourly weather.
  • Outputs: hourly shading masks and rooftop PV potential per building.
  • Model
  • Rooftop PV
  • Deep learning

Author

DONG Kechuan

PhD Student (Year 1)