Electricity generation in Brazil heavily depends on hydroelectric power, making it vulnerable to fluctuations due to its reliance on weather patterns. Accurately forecasting water inflow into hydroelectric plants is vital for the National Electric System Operator to make decisions regarding the monthly scheduling and operation of the power system. In this paper, an evaluation of predicted flows for a 14-day horizon are evaluated for the Tucuruíhydroelectric plant, located in the Tocantins river in the North of Brazil. The temporal fusion transformer (TFT), long short-term memory (LSTM), and temporal convolutional networks (TCN) are compared. The findings demonstrate that the TFT is a more suitable alternative than LSTM and TCN models for predicting inflows for the next 14 days. The TFT model is hypertuned by Optuna to achieve an optimized structure (h-TFT). The h-TFT had a mean absolute percentage error of 13.1 and a Nash–Sutcliffe of 0.96, outperforming its initial version and even the bidirectional LSTM model in benchmarking. Based on the error results, h-TFT showed promise for flow forecasting and provides insights into decision-making processes in the Brazilian electricity sector.
Enhancing hydroelectric inflow prediction in the Brazilian power system: A comparative analysis of machine learning models and hyperparameter optimization for decision support
Frizzo Stefenon S.
2024-01-01
Abstract
Electricity generation in Brazil heavily depends on hydroelectric power, making it vulnerable to fluctuations due to its reliance on weather patterns. Accurately forecasting water inflow into hydroelectric plants is vital for the National Electric System Operator to make decisions regarding the monthly scheduling and operation of the power system. In this paper, an evaluation of predicted flows for a 14-day horizon are evaluated for the Tucuruíhydroelectric plant, located in the Tocantins river in the North of Brazil. The temporal fusion transformer (TFT), long short-term memory (LSTM), and temporal convolutional networks (TCN) are compared. The findings demonstrate that the TFT is a more suitable alternative than LSTM and TCN models for predicting inflows for the next 14 days. The TFT model is hypertuned by Optuna to achieve an optimized structure (h-TFT). The h-TFT had a mean absolute percentage error of 13.1 and a Nash–Sutcliffe of 0.96, outperforming its initial version and even the bidirectional LSTM model in benchmarking. Based on the error results, h-TFT showed promise for flow forecasting and provides insights into decision-making processes in the Brazilian electricity sector.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
