Full text
| |
| Author(s): |
Nicole Santos e Aguiar
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação |
| Defense date: | 2019-12-11 |
| Examining board members: |
Fernando José Von Zuben;
Mateus Giesbrecht;
Guilherme Palermo Coelho
|
| Advisor: | Fernando José Von Zuben |
| Abstract | |
The main objective of this work is to present a multitask, efficient and automatic approach to estimate thresholds for a generalized Pareto distribution, aiming at high-performance prediction of extremes in multiple precipitation time series. Based on Extreme Value Theory, the only information used to model the heavy tail distribution by maximum likelihood estimation is given by the samples of the time series exceeding a user-defined threshold. This approach suffers from two fundamental drawbacks: (1) the subjectivity of the threshold definition, even when resorting to some graphical guidance, (2) the inherent sparse nature of the above-threshold samples, which, by definition, belong to the tail of the distribution. The proposal presented here for multitask learning automatically creates a hierarchical relationship among the prediction tasks and uses a nested cross-validation to automatize the choice of the optimal thresholds. Given the obtained hierarchical relationship among the prediction tasks, the multitask learning explores data from multiple related prediction tasks toward a more robust maximum likelihood estimation of the parameters that characterize the generalized Pareto distribution. The proposed methodology was applied to precipitation time series of South America and its performance was compared to a single-task learning method and to the traditional graphical approach, indicating a consistent performance improvement. Another advantage of the approach is the possibility of performing a qualitative interpretation of the obtained hierarchical relationship among the tasks, when associated with the geographical locations of the precipitation time series (AU) | |
| FAPESP's process: | 18/09887-1 - A multitask approach to the forecasting of precipitation extremes |
| Grantee: | Nicole Santos e Aguiar |
| Support Opportunities: | Scholarships in Brazil - Master |