Last year’s hurricane Milton was flooding in Pentagorda, Florida. (Blanol/Shutter stock)
According to a research published, a new device of artificial intelligence can make flood prediction more accurate and viable AGU develops. Researchers found that the machine learning model, combined with the US National Maritime and Environmental Administration (NOAA), has significantly improved the accuracy of the national flood predictions on the national scale.
The system, called the error cast net, was developed by a team headed by a scientist and a scientist at Michigan University. Based on the nervous network, the error casts acts as a correctional layer for the NOAA process -based model, which identifies and learns from organized errors in past forecasts. According to the researchers, when connected with the National Water Model, the hybrid approach increased the accuracy of the element of four to six in the lead hours of one to ten days. These results show that, instead of changing physics -based systems with AI, scientists are now looking for ways to combine both powers.
Make a pair of physics with machine learning
Noaa’s national water model is imitated and predicted for rivers and rivers in the United States, which provides updates many times daily. This model pulls data from about 11,000 operational water gaugees that measure rain, flow and river flow. This model is also a factors in the POW plant, urban development, and drainage samples of more hydrophrological accuracy.
Prediction errors occur when this complex interaction is incomplete modeling or when there are local data differences. Lead researcher Tran and his colleagues trained their nervous network for years NAUR network, including rain and flood records, to help identify where and why the similarities occurred.
Interactive Noaa Map of National Water Model. Each color has a dot water gauge that has information about the flood capacity. (Credit: Noaa)
“So especially for flooding, the performance of the pure AI model is quite bad.” Agony Release “The advantage of AI models is that they are very easy. You just need to train the model and use data to provide predictions, but what we need to worry about is to ensure the accuracy of flood events that can cause significant loss.”
The role of the error cast net is to analyze the historical performance of the national water model, to determine what kind of errors can be corrected, and the results can be improved. Some errors, such as physical limits or lost data, cannot be determined, but they can increase system training.
“You can’t throw physics,” said Valerie Ivanov, a physiological hydrophologist at Michigan University and author of the study. “This is just in accordance with the definition you cannot do. You have to understand that the system is different. The scenes are different. You have to calculate the dominant physical process in your prediction model.”
Keeping this philosophy in mind, researchers turned to machine learning to see how far they could move existing models without rejecting basic physics. This system uses a focused long -term short -term memory (LSTM) network, designed to handle a long series of input data such as a seasonal time series, a type of repeated nerve network has been used. The “Attention” method helps the model focusing on highly relevant patterns in these streams, which improves the ability to find out when and where the past predictions have gone wrong.
To calculate uncertainty, researchers used a technique called Monte Carlo Dropout, which operates thousands of times the network with small random changes. The result of the forecasts not only provides an estimate of a series of flows, but also a range of potential outcomes and their possibilities are known as the forecast. The model’s performance was evaluated using the established hydroology matrix, which includes the performance of the killing shopping, which measures overall prediction skills, as well as peak error and timely peak error, which captures that the model predicts the height of the flood and the height of the flood.
The result is a hybrid framework that takes advantage of AI to correct the leading prejudice in the model while preserving the physical realism needed for hydrophrologic predictions, which can serve as a map for how the AI can enhance the scientific modeling system.
Medium range (10 days) reviews the flu flu forecast framework. (a) NWM integration with ECN. The ECN is trained using NWM Hindus, forcible climate, and USGS flow observations to estimate a pair of errors, which is combined with NWM forecasts to predict the possibility. (B) Prediction framework that uses 1 to 10 days error predictions: observation (gray) and predictions (colorful) weather, observed flow, and NWM predictions. “+” Indicates a combination of data. N is the Lickback window for observed data. (Credit: study authors)
Faster, more reliable flood warning towards
According to the study, the error cast net can produce national -scale flood predictions in minutes using your computer effective framework. Researchers believe that this approach can eventually enable the detailed flood predictions several days before an incident, while promoting the reliability of early warning systems around the world, especially in areas where flood monitoring infrastructure is limited. The authors say that although the existing system has been trained on the US data of the US, it can be molded for other regions using local hydrological information.
The accuracy, reliability and economic value of the continental flood predictions titled “AI” has been improved. “ AGU develops. The co-authors include Michigan University, Pacific North West National Laboratory, NASA Goddard Space Flight Center, Virginia University, Wisconsin University-Medicine, and Scientists at Alsis University in South Korea.
Relevant
