EVALUATING THE PRECIPITATION PREDICTION SKILL BY GRAPHCAST IN THE JAPAN REGION

Abstract

 GraphCast, a recently developed machine learning-based weather forecasting model, can predict global weather with a spatial resolution of 0.25° for the next 10 days. This study evaluated the precipitation prediction skill of GraphCast in the Japan region by comparing it with the results of ECMWF’s deterministic model (HRES) and ensemble mean (ENS). GraphCast exhibited the lowest mean absolute error in precipitation predictions for most lead times, demonstrating superior performance over the three methods in overall precipitation predictiong. Additionally, the prediction skill for heavy rainfalls exceeding 100 (mm/24h) was evaluated using the Threat Score. The results indicated that HRES performed best up to a lead time of 3 days, after which GraphCast excelled. Furthermore, it was revealed that GraphCast's skill in predicting heavy rainfall for lead times of 4 to 10 days was comparable to that of HRES at a lead time of 4 days.