Skip to main navigation menu Skip to main content Skip to site footer

Vol. 10 No. 1 (2024)
Special Issue: Proceedings of CVIS 2024

Articles

Perceiver Model Ensemble for Solar Power Forecasting: A Winning Solution for ClimateHack.AI 2023-2024

https://doi.org/10.15353/jcvis.v10i1.10018

Published 2025-01-31

  • Trevor Yu
    • ,
  • Carter Demars
    • ,
  • Areel Khan

Abstract

In this paper, we present Team Waterloo's winning approach for solar power forecasting in ClimateHack.AI 2023-2024, an international machine learning competition. Our model leverages Numerical Weather Prediction (NWP), high-resolution visible (HRV) satellite imagery, and solar panel site metadata to predict photovoltaic (PV) power output over a 4-hour window. Our solution was an ensemble of Perceiver models that used spatial semantic pointers for spatial-temporal encoding, dynamic cropping, and efficient data handling. Our model can provide low-latency, high-accuracy forecasts and achieved a mean absolute error of 0.081 on the competition test set.
PDF