435132 Improving the Subseasonal Forecasting of the Western U.S. Precipitation with A Causality-Based Statistical Model

Wednesday, 31 January 2024: 12:00 AM
345/346 (The Baltimore Convention Center)
Jiwang Ma, The Artificial Intelligence Group, Southern Marine Laboratory, Zhuhai, Guangdong, China; and X. S. Liang

Subseasonal prediction of precipitation is critical to agriculture, hydrology, health, energy, and other sectors of our society. It is, however, a notorious and continuing challenge due to the chaotic nature of the atmosphere. In 2017, the U.S. Bureau of Reclamation and the National Oceanic and Atmospheric Administration (NOAA) launched a national competition, the Subseasonal Climate Forecast Rodeo, for a one-year real-time subseasonal forecasting of precipitation over the Western U.S. at a lead time of 2-6 weeks. Hwang et al. (2018) then built a dataset for this competition, which allows us to evaluate the ability of our own model and compare the results to the benchmark they built on this dataset.

Using a recently developed quantitative causality analysis, namely, the Liang-Kleeman information flow (IF) analysis, a causal inference method rigorously derived from first principles and, particularly, born from atmosphere-ocean science, the causal relations of different variables with precipitation are analyzed. The causalities from relative humidity, sea surface temperature and geopotential height at 10 hPa are found to be the top 3 strongest; these three variables are hence selected as input variables to build the model. As a starting step, a simple linear model architecture is adopted. Remarkably, the forecast of the week 3-4 precipitation during the Rodeo period reaches a skill score as high as 0.459, much higher than the benchmark skill score, 0.2364. For the weeks 5-6, our model reaches a score of 0.437, compared to the benchmark score 0.2315. Multi-year average skills (from 2011-2017) are also evaluated and compared to the benchmark. For the forecasts for weeks 3-4 and weeks 5-6, our skill scores are both 0.41, compared to the respective benchmark scores 0.1968 and 0.1857.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner