MIT Uses AI to Improve Two-to-Six-Week Weather Forecasts
MIT researchers added a machine-learning error-correction mechanism to an existing weather forecasting model rather than building a new model. According to TechNews, the method uses past forecasts and observations to learn condition-specific errors and produced twice the forecast skill of the original approach for two-to-six-week predictions in an ECMWF competition. The team also reported improved detection of extreme conditions such as unusually cold weather and floods, and the study was published in Nature.
TechNews says the research could help aviation, energy and agriculture identify weather risks and allocate resources weeks earlier, while the original code was released without requiring a supercomputer.


現今的天氣預報能提供未來幾天的氣象資訊,但若想提前一個月安排旅行、戶外婚禮或露天活動,預報往往仍不夠準確。因為兩週以後,大氣變化的不確定性增加,預測難度也隨之提高。最近,麻省理工學院(MIT)利用人工智慧改善天氣預報模型,提升兩至六週內的預測度,並在國際預報競賽贏得冠軍。成果近日刊登於《Nature》期刊,可望讓人們更早掌握天氣趨勢,不僅有助安排出遊,也能協助航空、能源與農業等產業提前調度資源,降低天氣變化帶來的損失。