Weather forecasting has improved significantly in recent decades. Thanks to advances in monitoring and computing technology, today’s 5-day forecasts are as accurate as 1-day forecasts were in 1980. Artificial intelligence could revolutionize weather forecasts again. In a new study, Arcomano et al. present a machine learning model that forecasts weather in the same format as classic numerical weather prediction models.
Previously, the team developed an efficient machine learning algorithm for the prediction of large, chaotic systems and demonstrated how to incorporate the algorithm into a hybrid numerical machine learning model for dynamical systems like atmospheric conditions. In the new proof-of-concept study, the researchers build on their previous work by using a reservoir computing–based model, rather than a deep learning model, to reduce the training time requirements for their machine learning technique.
The researchers trained their model using data from the European Centre for Medium-Range Weather Forecasts and prepared 171 separate 20-day forecasts, each of which took just 1 minute to prepare. They compared the machine learning forecasts to three benchmark forecasts: daily climatology; a persistence model, which assumes that the atmospheric state will remain constant throughout the forecast; and the Simplified Parameterizations, Privitive-Equation Dynamics (SPEEDY) model, a low-resolution version of numerical weather prediction models.
They found that the machine learning model typically forecast the global atmospheric state with skill 3 days out. It outperformed both daily climatology and the persistence model in the extratropics, though not in the tropics, and bested the SPEEDY model in predicting temperature in the tropics and specific humidity at the surface in both the tropics and the extratropics. However, the SPEEDY model still outperformed the machine learning model for wind forecasts more than 24 hours out. The authors note that overall, the reservoir computing–based machine learning model is highly efficient and may be useful in rapid and short-term weather forecasts. (Geophysical Research Letters, https://doi.org/10.1029/2020GL087776, 2020)
—Kate Wheeling, Freelance Writer
Source: https://eos.org/research-spotlights/improving-atmospheric-forecasts-with-machine-learning
