Artificial intelligence for multi-scale meteorological forecasting in French Polynesia

Meteorolgical forecasting plays a crucial role in decision making across all level of society. French Polynesia, a vast marine territory the size of Europe in the South Pacific, with 118 island and atolls, is particularly sensitive to meteorological extremes. Hence, accurate meteorological forecasts are of the utmost importance to minimize the impact of severe weather events, such as tropical cyclones, not only in the short term but also 2 to 3 weeks in advance. Sub-seasonnal to Seasonnal (S2S) meteorological forecasts, i.e 15 – 45 days lead time, has long been a gap in operational weather forecasting. Precise S2S predictions are especially difficult to implement due to the dampening of the predictibility in the atmosphere beyond the 2-week limit. However, the study of the Madden-Julian Oscillation will help as it represents the most important source of predictibility on a sub-seasonnal timescale.

Here, we explore the sub-seasonnal forecast capabilities of three pre-trained models, namely FourCastNet, GraphCastNet and Pangu weather, respectively implemented by Nvidia, Google DeepMind and Huawai. The data-driven forecasts are compared with S2S predictions from various meteorological data centers. Transfer learning and fine-tuning of the different machine learning models is implemented to forecast precipitations and a heat index (bioclimatic comfort) on a sub-seasonnal timescale in the South Pacific, with a focus on French Polynesia. The performance of the different models is also explored to forecast the occurrence, amplitude and trajectory of tropical cyclones in S2S forecasts. Finally, we will discuss our forecasting results for the potentially unsual 2023 – 2024 South Pacific season associated with the ongoing El Niño.