Full text
| |
| Author(s): |
Vinicius Araujo
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Geociências |
| Defense date: | 2022-10-20 |
| Examining board members: |
Ana Elisa Silva de Abreu;
Filipe Antonio Marques Falcetta;
Paula Dornhofer Paro Costa
|
| Advisor: | Ana Elisa Silva de Abreu |
| Abstract | |
Several flood management measures have been taken to mitigate the damages of flash-flood events worldwide. Among them we can mention the generation of data-driven hydrological models to support early warning systems, since such models do not have the need to characterize the several physical parameters involved in the modelling and require a smaller amount of data when compared to more traditional models. In this context, the present work aims at developing a data-driven model to predict future water levels in the Proença watershed, located in Campinas city, São Paulo state, Brazil, which historically presents floods that pose risks to the population and damage to property. To this end, sub-hourly water level data and weather radar data were used for the period from November 2014 to June 2019, in addition to flood occurrence records collected in digital media and in the Civil Defense database. From the analysis of the generated cotagram, three thresholds were applied (2.30m, 1.96m and 1.58m), based on the level exceeded or equalled in 3, 6 and 10% of the analysed time, respectively. The analysis of the 145 events selected by the most comprehensive threshold (1.58 meters) recorded the presence of 17 flood events in the watershed, in addition to verifying that the levels of 2.30 meters and 3.00 meters can be used as levels of attention and alert for the basin, respectively. For the development of the research, a multimodal dataset was organized from real data, which was made available at Unicamp's research data repository. XGBoost algorithm was used to model water levels with 30 minutes lead time using this dataset. This model presents for the test data an RMSE of 0.284 meters and an NSE of 0.614 and can be classified as a satisfactory model. For the 3 flood events belonging to the test data, the model was able to predict levels above the attention level for 2 events, but none above the alert level. Modeling other lead times revealed that as the lead time is decreased, the prediction quality improves. Analyzing the lag-time between precipitation and water level for the 145 events mentioned above, the presence of two distinct groups of values emerged: one with average lag time of 30 minutes and the other with average lag time of 70 minutes. The first one is related to events of greater rainfall intensity, while the second one is related to events of lower rainfall intensity. This result and the performance of the data-driven model, suggests that there is high variability intrinsic to the problem and that it has been only partiallycaptured by the data-driven model developed in the present research (AU) | |
| FAPESP's process: | 20/00058-2 - Stream water level forecasting in a built-up basin in Campinas, SP, using machine learning and aiming at structuring a flash flood early warning system |
| Grantee: | Vinicius Araujo |
| Support Opportunities: | Scholarships in Brazil - Master |