Banca de DEFESA: LUIS GUSTAVO DE MOURA REIS

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : LUIS GUSTAVO DE MOURA REIS
DATE: 27/07/2022
TIME: 14:00
LOCAL: Programa de Pós-Graduação em Engenharia Civil
TITLE:

Machine Learning Modeling for Seasonal Low-Flow Predictions Using Climate and Satellite Derived Drought Indexes

KEY WORDS:

low-flow prediction, long-term streamflow prediction, machine learning, statistical learning, SVM, RF, XGBoost

PAGES: 130
BIG AREA: Engenharias
AREA: Engenharia Civil
SUMMARY:

Long-term low-flow predictions are of great importance in a scenario marked by changes in precipitation patterns and increasing drought frequency, severity, and duration. In the presence of significative hydrological trends, water permit limitations based on permanent low flow quantiles may not work out, leading to incompatible restrictions during the wet or dry season. On one side is the increasing availability of climatic data on a global scale and satellite-derived data, with high spatial and temporal resolution. On another, data-driven models derived from machine learning algorithms have shown powerful skill in dealing with loads of data and delivering accurate results in prediction tasks. Nevertheless, its application in hydrology sciences is still underused. This work employed machine learning algorithms associated with satellite-derived and reanalysis data to forecast low streamflows in the dry season with 6- and 9-months lead time. Global drought indexes: Temperature Condition Index - TCI and Vegetation Condition Index -VCI provided from NOAA/AVHRR-VIIRS, and Palmer Severity Drouth Index – PDSI were employed as predictors. Climate hazards infrared precipitation with stations – CHIRPS was considered as input. We also investigate the skill PDSI as a proxy to observed streamflows. The skill of machine learning algorithms: SVM (support vector machine), RF (random forest), and XGBoost (Extreme Gradient Boosting) were also investigated.The experiment was carried out in five watersheds located in Brazilian Atlantic Forest and Cerrado Biome(Brazilian Savana). Algorithms’ performance, generalization ability, and skill of predictors were tested during the most severe drought period. Influences of spatial variability were investigated by the extraction of covariation modes and EOFs. Our experiments showed that an inter-annual low-frequency signal controls the predictability of low streamflows. The lower the annual rainfall, the higher interannual variability. Machine Learning models derived from the SVM-Linear algorithm drastically outperformed RF and XGBoost and showed great skill in predicting low flows with a six-month lead time. SVM showed great generalization skills when the models were submitted to a proxy basin test. Observed streamflows play an essential role in the prediction, when associated with smoothing and differentiation technics can drastically improve algorithm’s learning about hydrologic water balance.

BANKING MEMBERS:
Externo à Instituição - ANDRÉ RICARDO BACKES - UFU
Externo à Instituição - BENEDITO CLÁUDIO DA SILVA - UNIFEI - UNI
Externo à Instituição - ADRIANO ROLIM DA PAZ - UFPB
Presidente - 1688881 - ALFREDO RIBEIRO NETO
Externa ao Programa - 1960878 - DORIS REGINA AIRES VELEDA