Estudo comparativo entre algoritmos de previsão de cheias sazonais usando rede neural artificial e método de aprendizado baseado em comitê

Abstract

The occurrence of seasonal floods of rivers affects, mainly, the riverside population. It is identified in bibliographic databases the accomplishment of several researches in the subject prediction of occurrence of floods. The results of these surveys constitute important contributions to public policies, since flood forecasting tools can enable preventive actions, which minimize the damage caused to the riverine populations. In this work, methods capable of predicting the peak of the river flood were evaluated. The methods developed were evaluated in the flood forecast of the Negro River, the largest tributary of the left bank of the Amazon River and the main river that passes by the city of Manaus. The predictors implemented were: Artificial Neural Networks (ANN) and Learning methods based on Committee. The following input data were used for the period 1951-2017: climatic indexes and the level of the river itself. These data were later subjected to a process of selection of characteristics. For the predictor using ANN, three architectures, differentiated by the number of neurons in the hidden layers, were evaluated: 6, 8 and 10, which were trained using the following generalization methods: L2 regularization and early stopping. The forecast period was varied from 1 to 4 months in advance of the occurrence of the maximum flood peak in the region. Additionally, it was proposed to predict floods in four categories: high, medium-high, medium-low and low. For the predictor using Committee-Based Learning methods, the bagging and boosting algorithms were used to create the regression committee. The efficiency of the predictors was evaluated through the Pearson Correlation Coefficient (PCC) and the accuracy of the flood categorization. The best Pearson coefficient result for the predictor using ANN with early stop was rp = 0,9592, while the best committee result was obtained using the bagging method, rp = 0,9374. The classification of floods into categories presented an accuracy of 85,07% for the predictor ANN (early stop and bootstrap) and with the Method based on committee was of 82,09% (bagging).