Applying artificial intelligence to better predict real weather
Peer-Reviewed PublicationThe morning’s weather report called for sunshine, but rain is pouring down — why is weather difficult to predict and can it be improved? An international team of researchers turned to artificial intelligence to increase the accuracy of weather forecasts for the next 24 hours to 10 days out.
They published their results in Advances in Atmospheric Sciences.
“Accurate weather forecasting is critical in protecting socioeconomic conditions in many areas such as transportation, agriculture and water resources management,” said first author Lei Han, a professor of Information Science and Engineering, Ocean University of China, and affiliate scientist with the Institute of Urban Meteorology, China Meteorological Administration. At present, operational weather forecasts rely on numerical weather prediction (NWP) models, which have undergone significant improvement over the past four decades. Despite improvement, inevitable errors still arise from how atmospheric dynamics and physics are represented in NWP models, which lead to either over- or under-forecasting. “Developing efficient and effective methods to reduce forecast bias is essential to improving the forecasting skill of numerical models,” Han added.
Correction methods have been previously proposed, yet almost all tend to focus on correcting localized forecast data at individual weather observation stations rather than a broader approach.
The researchers used deep learning, a type of artificial intelligence, to train a network to better understand the uncertainty involved in the weather forecast products. The network learned to better predict four weather variables — temperature, humidity, wind speed and direction — from twice-a-day forecasts issued by the European Centre for Medium-range Weather Forecasts, or ECMWF.
The researchers were inspired by a deep learning network called U-net, which segments biomedical images to yield more precise and detailed information. They called their method correction U-net, or CU-net. They transformed the forecast correction problem into an image-to-image translation problem in deep learning under the CU-net architecture first, then CU-net compares data and finds similarities and differences, learning how they might influence each other, and it complies the data in convolutional layers, pooling more information as it learns. CU-net has many intermediate representations, whose benefit is that different convolutional layers can detect features at different spatial resolutions. Eventually, it gathers enough information to predict what a grid block should look like based on the data from the blocks surrounding it, as well as how it may change in the coming days.
“Our ultimate goal is to help improve NWP models to further reduce deviation and improve forecast performance,” said corresponding author Mingxuan Chen, a senior scientist with the Institute of Urban Meteorology, China Meteorological Administration. The researchers used over 10-yr data (2005-2016) to train their AI model, and independent data from 2017 and 2018 to test and quantify their model performance. They found that the network improved forecasts for all four weather variables. “The confidence gained from this research can directly lead to innovative diagnostics for weather and climate variability, as we have never been able to imagine before the era of AI,” said Haonan Chen, a professor of electrical and computer engineering at Colorado State University and coauthor of this research article. “We’ll be seeing AI as the game changer in Earth system research,” added Chen.
###
Other contributors include Kangkai Chen and Yanbiao Zhang, College of Information Science and Engineering, Ocean University of China; Bing Lu, Linye Song and Rui Qin, Institute of Urban Meteorology, China Meteorological Administration.
JOURNAL
Advances in Atmospheric Sciences
DOI
10.1007/s00376-021-0215-y
ARTICLE TITLE
A Deep Learning Method for Bias Correction of ECMWF 24–240 h Forecasts
ARTICLE PUBLICATION DATE
9-Jul-2021