A multitasking learning approach to precipitation extremes forecasting

• Grant number: 17/19397-9
• Support Opportunities: Research Grants - Research Partnership for Technological Innovation - PITE
• Duration: December 01, 2018 - May 31, 2020
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Convênio/Acordo: IBM Brasil
• Principal Investigator: Fernando José von Zuben
• Grantee: Fernando José von Zuben
• Host Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Host Company: IBM Brasil - Indústria, Máquinas e Serviços Ltda
• City: Rio de JaneiroCampinas
• Associated researchers: André Ricardo Gonçalves
• Associated scholarship(s): 18/09887-1 - A multitask approach to the forecasting of precipitation extremes, BP.MS

Abstract

Abnormal climate events, such as precipitation extremes, are becoming more frequent, widespread, and intense, imposing huge challenges to society, even in urban or rural areas. Therefore, we are facing an increasing demand for high-performance information forecasting services. In fact, the effectiveness of economical and safety policies is directly promoted by our ability to manage the risks and estimate the impacts associated with extreme events. Agricultural productivity is closely linked to climatic conditions, thus expanding the demand for reliable forecasting systems. The purpose is to avoid huge losses in production or even operational issues in transportation and storage. Heavy rainfall events tend to foment flood scenarios, responsible for wiping out entire crops over wide areas. Excess of water may also cause other negative impacts, such as waterlogging, anaerobicity and reduced plant growth. Earth system models (ESMs) trust on physical principles that govern climate behavior, and can be used to perform projections of future climate conditions. Many ESMs have been proposed and their response may differ significantly. That is why researchers usually perform climate projections based on ensembles of ESMs, aiming at better accuracy and reduced uncertainty. Here, ensembles will follow a multitask learning (MTL) perspective, where MTL looks for an improved generalization performance by simultaneously learning related tasks. A structural representation of the information sharing among tasks is going to be properly explored to make more robust and accurate precipitation extremes forecasting, also including spatial gridding and specific statistical tools devoted to extreme events. This multitask learning approach for ensembles of ESMs has already been applied with success in temperature and humidity forecasting, by the same research group that are proposing this project, and the intention is to extend the approach to precipitation extremes. (AU)