Presentation
(PhD08) A Machine Learning Workflow for Hurricane Prediction
SessionPhD Forum
Event Type
PhD Forum
AI/Machine Learning/Deep Learning
Big Data Analytics
TimeMonday, June 25th1:29pm - 1:33pm
LocationAnalog 1, 2
DescriptionThe Atlantic hurricane season runs from June 1st to November causing massive destruction and loss of life. In 2017, 17 named storms hit the Atlantic causing destruction worth an estimated $316 million and at least 464 fatalities. Meteorologists, by studying previous weather data, predict the expected number of hurricanes in the season. These predictions help authorities prepare for disasters and over the years, better predictions have minimized loss of life and property. However, these predictions rely on human expertise and are often extremely complex due to the thousands of parameters involved and the chaotic nature of weather.
The aim of this study is to introduce Machine Learning to hurricane prediction. Recent scientific advancements have seen geostationary satellites capable of collecting tens of Terabytes of daily data of the weather. On the other side, machine learning models propose efficient techniques to analyze such large data and extract meaningful information. With this large amount of data and the power of high performance computing, machine learning could be an alternative tool for climate study.
Machine learning has the ability to understand complex models and relationships in data. Recent developments in deep learning models such as Deep Neural Networks (DNN) have led to significant achievements in accuracy. We introduce machine learning model to hurricane prediction to explore the complex relationship between multiple factors such as sea surface temperature, sea level pressure, sea ice cover and wind patterns. We aim to apply a deep learning model to understand the effect of these parameters in the hurricane season and the number of hurricanes. Such insights could significantly improve disaster preparedness and give authorities a better picture of what to expect in the hurricane season.
Using historical data, we train a DNN to classify the hurricane season based on the number of hurricanes likely to occur. Our initial experiments are aimed at finding the most significant factors in storm formation. Preliminary results show that Sea surface temperature has the highest impact on the prediction of the number of storms. Furthermore, given the average sea surface temperatures in a month, our DNN model is able to predict the number of hurricanes with about 60% of accuracy.
Our future goal is to develop a complete end to end workflow to continuously learn weather patterns that affect the hurricane season and accordingly, to make predictions. We plan to implement distributed learning using PyCOMPSs (a programming model and runtime which aims to ease the development of parallel applications for distributed infrastructures, such as Clusters and Clouds) to reduce or eliminate the need to expensive computational infrastructure in climate science.
The aim of this study is to introduce Machine Learning to hurricane prediction. Recent scientific advancements have seen geostationary satellites capable of collecting tens of Terabytes of daily data of the weather. On the other side, machine learning models propose efficient techniques to analyze such large data and extract meaningful information. With this large amount of data and the power of high performance computing, machine learning could be an alternative tool for climate study.
Machine learning has the ability to understand complex models and relationships in data. Recent developments in deep learning models such as Deep Neural Networks (DNN) have led to significant achievements in accuracy. We introduce machine learning model to hurricane prediction to explore the complex relationship between multiple factors such as sea surface temperature, sea level pressure, sea ice cover and wind patterns. We aim to apply a deep learning model to understand the effect of these parameters in the hurricane season and the number of hurricanes. Such insights could significantly improve disaster preparedness and give authorities a better picture of what to expect in the hurricane season.
Using historical data, we train a DNN to classify the hurricane season based on the number of hurricanes likely to occur. Our initial experiments are aimed at finding the most significant factors in storm formation. Preliminary results show that Sea surface temperature has the highest impact on the prediction of the number of storms. Furthermore, given the average sea surface temperatures in a month, our DNN model is able to predict the number of hurricanes with about 60% of accuracy.
Our future goal is to develop a complete end to end workflow to continuously learn weather patterns that affect the hurricane season and accordingly, to make predictions. We plan to implement distributed learning using PyCOMPSs (a programming model and runtime which aims to ease the development of parallel applications for distributed infrastructures, such as Clusters and Clouds) to reduce or eliminate the need to expensive computational infrastructure in climate science.
Poster PDF
