Spatially Aware Calibration of NWP and AI Precipitation Forecasts
Belinda Trotta, Esteban Abellan
Abstract Rainfall is often highly localized and its location is difficult to predict exactly with a numerical weather prediction (NWP) model. Previous research has shown that this problem can be mitigated by spatially aware calibration methods which incorporate forecast information from neighboring grid cells. In this work we investigate whether artificial intelligence (AI) weather models similarly benefit from spatially aware calibration. We evaluate pointwise and spatially aware calibration approaches for the physics‐based HRES and artificial intelligence AIFS models from ECMWF. For each forecast, we test 3 types of neural network calibration approach: a simple pointwise method using a fully connected architecture, a convolutional neural network (CNN) using a small neighborhood of each point, and an intermediate approach, which flattens the points of the neighborhood into a single dimension and feeds it into a fully connected architecture. We find that for HRES the neighborhood approaches generally outperform the pointwise approach on site metrics, with the CNN being the best model. However for AIFS all methods perform similarly. This suggests that spatial uncertainty is already modeled adequately by the AI model, so simpler calibration approaches suffice. Additionally, despite its lower resolution, AIFS is generally more accurate than HRES: whether comparing the raw forecasts, or the calibrated outputs of the two models, AIFS scores better on key metrics.