Banca de DEFESA: Thiago Lappicy Lemos Gomes

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : Thiago Lappicy Lemos Gomes
DATE: 24/07/2023
TIME: 08:30
LOCAL: Sala de Informática do PPGTARH com Transmissão pela Plataforma MS Teams
TITLE:

MONTHLY FORECASTS FOR THE RESERVOIRS BELONGING TO THE BRAZILIAN HYDROPOWER SYSTEM USING STOCHASTIC METHODS AND CLIMATIC INFORMATION

KEY WORDS:

Streamflow forecasting; SIN Hydropower plants; Stochastic methods;Climate information.

PAGES: 146
BIG AREA: Engenharias
AREA: Engenharia Sanitária
SUBÁREA: Recursos Hídricos
SUMMARY:

A better monthly streamflow forecast implies in an extremely impactful gain for the Brazilian electric sector. The national operator for such system (ONS) currently uses stochastic methods (PAR) for the main part of such forecasts, including generation of future scenarios. Even though these methods are well structured in the existing literature and they are easily interpretable, they are still subject to improvements.In the existing literature there are two main innovation possibilities. One is the use of climatic indicators to aid in the forecasts. The most commonly found are the sea surface temperature (SST) in different places, the wind at a certain altitude (pressure) and the inclusion of rainfall forecast to aid the streamflow forecast. In this study is it used the trade winds (U1) and the SST in two spots (SST2 at the Atlantic and NINO3 at the Pacific Ocean). One way to integrate such variables is through an autoregressive exogenous model (PARX). The exact relation between climate and streamflow are still unknown and the currently knowledge surrounding them is limited to knowing they are complex and non-linear.The other possibility for innovation is directed towards the existing non linearity, being possible through machine learning techniques. The gain from using such methods comes at a cost of a lack of interpretability. The L2 regularization, also known as ridge regression, is a machine learning technique that is able to maintain this interpretability – therefore being the third choice for evaluated models.This study shows a systematic gain with the use of climatic information when comparing both PARX and RIDGE to PAR. This improvement was evaluated by the NSE and KGE metrics with their individual components (α, r, β_NSE and β_KGE) for different months and lead times (1 to 6 months ahead). The NSE, KGE, α, r, β_NSE and β_KGE showed gains of up to, respectively, 68%, 79%, 90%, 71%, 81% and 82% when grouping the hydropower plants by each month. Spatially, a bigger gain was seen in the North and South Regions of Brazil, showing a difficulty in the forecasts of more central areas and in the Northeast of Brazil.The gains are smaller when grouping the individual hydropower plants into equivalent energy reservoirs (REE) and subsystems, but they are still present. With such grouping, the effect of each climatic indicator in the forecasts are easier to observe. The same patterns are seen (North and South with better forecasts), even though that the “worst” subsystem, the Northeastern, still shows a gain in 49% of the scenarios (72 scenarios were made from the 12 months and 6 lead times) – which is notable. The “best” gain was for the North, where 78% of the scenarios showed an improvement on KGE metrics when using the proposed climatic indicators.

COMMITTEE MEMBERS:
Interno - 1646016 - ALEXANDRE KEPLER SOARES
Presidente - 2764698 - CARLOS HENRIQUE RIBEIRO LIMA
Interno - 1809020 - DIRCEU SILVEIRA REIS JUNIOR
Externo à Instituição - WILSON DOS SANTOS FERNANDES