In this article, the authors propose a new machine learning model called FuXi-S2S for global subseasonal forecasts. Subseasonal forecasting refers to predicting weather patterns over a period of several weeks or months ahead. This is crucial for industries such as agriculture and water management, which rely heavily on accurate weather predictions.
The authors explain that traditional machine learning models struggle with long-term subseasonal forecasts due to the complex nature of atmospheric dynamics. To address this challenge, they develop an autoregressive, multi-step loss function, similar to FuXi, to mitigate cumulative errors over long lead times. The training process follows an autoregressive training regime and a curriculum training schedule, incrementally increasing the number of autoregressive steps from 1 to 17. Each autoregressive step undergoes 1000 iterations of training with a batch size of 1 on each GPU. The optimization is performed using AdamW with an initial learning rate of 2.5 x 10^-4 and a weight decay coefficient of 0.1.
The authors then describe the evaluation method used to assess the performance of FuXi-S2S. They use the Climate Forecast System (CFS) database, which contains atmospheric circulation data with a lead time of up to 6 months. The model is evaluated based on its ability to predict temperature and precipitation patterns during the warm season in the North Hemisphere.
The authors highlight several key findings from their evaluation:
- FuXi-S2S outperforms other machine learning models, including the traditional multi-step forecasting method, in terms of prediction accuracy.
- The model demonstrates improved performance when using a larger number of autoregressive steps, indicating that it can capture complex patterns in the data.
- The authors find that the curriculum training schedule improves the model’s performance, as it allows for gradual increases in the number of autoregressive steps and batch size.
The authors conclude by emphasizing the potential of FuXi-S2S to improve subseasonal forecasting accuracy for various applications. They note that the model’s ability to capture complex patterns over long lead times makes it a valuable tool for industries such as agriculture, water management, and disaster risk reduction.
In summary, this article presents FuXi-S2S, a machine learning model that can accurately predict weather patterns over several weeks or months ahead. By developing an autoregressive, multi-step loss function and using a curriculum training schedule, the authors demonstrate how FuXi-S2S can capture complex patterns in atmospheric dynamics and outperform other machine learning models. The potential applications of this model are vast, ranging from agriculture to disaster risk reduction, making it an important contribution to the field of subseasonal forecasting.
