Efficient Subseasonal Weather Forecast Using Teleconnection-informed Transformers

Subseasonal forecasting, which is pivotal for agriculture, water resourcemanagement, and early warning of disasters, faces challenges due to the chaoticnature of the atmosphere. Recent advances in machine learning (ML) haverevolutionized weather forecasting by achieving competitive predictive skillsto numerical models. However, training such foundation models requiresthousands of GPU days, which causes substantial carbon emissions and limitstheir broader applicability. Moreover, ML models tend to fool the pixel-wiseerror scores by producing smoothed results which lack physical consistency andmeteorological meaning. To deal with the aforementioned problems, we propose ateleconnection-informed transformer. Our architecture leverages the pretrainedPangu model to achieve good initial weights and integrates ateleconnection-informed temporal module to improve predictability in anextended temporal range. Remarkably, by adjusting 1.1parameters, our method enhances predictability on four surface and fiveupper-level atmospheric variables at a two-week lead time. Furthermore, theteleconnection-filtered features improve the spatial granularity of outputssignificantly, indicating their potential physical consistency. Our researchunderscores the importance of atmospheric and oceanic teleconnections indriving future weather conditions. Besides, it presents a resource-efficientpathway for researchers to leverage existing foundation models on versatiledownstream tasks.