Redes neurais aplicadas na previsão das taxas de afretamento por tempo de Platform Supply Vessels no Brasil

Abstract

During the exploration and production process of an offshore oil basin, in addition to the movement of oil and its derivatives, are moved the most different supplies used to maintain these activities. To supply this demand, the majority of the Brazilian fleet of support maritime units is make up of Platform Supply Vessels (PSV). Due to high costs associated with the charter rates of these vessels, oil companies must consider contracts strategies to become flexible to variations in market scenarios, in order to achieve a competitive advantage. Therefore, the future rates forecasting for charter contracts has a fundamental role. This study presents a comparison between different Artificial Neural Networks performances applied in the prediction of Term Charter Rates for PSV contracts in Brazil, considering factors associated with their volatility. In this master thesis, 53 independent variables associated with the offshore market were used in the charter rate forecasts, and the results were compared with forecasts using a reduced number of variables through the method of variable selection with Partial Least Squares Regression (PLS). With the selection of the most important variables, it was possible to reduce the factors' number to 10 in the two study models, after comparing the results with three different neural networks. As a result, better performance and greater adherence to forecasts were obtained using Neural Networks with Long Short-Term Memory (LSTM) in all models, when compared to traditional neural networks in the literature, such as Radial Base Function and Multilayer Perceptron. It was also possible to verify that in four of the six evaluated prediction models, the use of techniques of variables selection of greater importance presented not only a reduction in the model complexity, but also a reduction in the Mean Absolute Percentage Error. In the charter rates forecasts using LSTM networks, errors of less than 3% were obtained for the two classifications of PSV vessel capacity studied, with a coefficient of determination of 84% in the best case [resumo fornecido pelo autor].