Coordinate-RNN for error correction on numerical weather prediction

In this work, we present a coordinate-based Recurrent Neural Networks (RNN) for error correction on the Numerical Weather Prediction (NWP) model. We show that the output errors on NWP have spatial and temporal properties, which is collinear with meteorological data. The correction model reflects these characteristics by encompassing the latitude and longitude coordinates as direct inputs to RNN. Examined with the NWP data in Korea, the proposed RNN-based correction reduces the humidity prediction errors by 4.8% and 4.2% compared to the predictions without correction and with simple linear correction, respectively. The overall result highlights the promise of our approach.