PiggyCast - Improving Weather Prediction Accuracy through a Stacking-Based Ensemble AI Approach. (Tutorials Track) Spotlight
Josiah Kimani (African Institute for Mathematical Sciences (AIMS) - South Africa)); Oliver Angélil (Ishango.ai); Chris Toumping (Inshango.ai); Steffen Knoblauch (University of Heidelberg)
Abstract
Accurate weather prediction is fundamental to understanding and adapting to the impacts of climate change. As our climate shifts, the frequency and intensity of extreme weather events are changing, making reliable forecasts more critical than ever. This tutorial introduces PiggyCast, an ensemble machine learning model designed to improve weather prediction accuracy by stacking forecasts from various numerical, AI-based, and hybrid weather prediction models. We demonstrate how a combined approach, using gradient-boosted decision trees (XGBoost), can surpass the predictive performance of individual base models. The main contributions of this tutorial are: 1. Developing and assessing an ensemble model:** We build and evaluate PiggyCast, a stacking-based ensemble model that leverages forecasts from state-of-the-art weather prediction models (IFS HRES, GraphCast, Pangu Weather, and NeuralGCM) and trains an XGBoost regressor on top of them to produce more accurate predictions of geopotential height at 500 hPa. 2. Investigating feature importance: We use SHAP (SHapley Additive exPlanations) values to analyze the contribution of each base model's forecast and geographic coordinates to PiggyCast's predictions, providing insights into the model's decision-making process.