Comparative analysis of machine learning methods for the prediction of brake power and rate of revolution for bulk carriers

In the preliminary ship design process, key aspects such as machinery and powering must be specified, which involves estimating the brake power and rate of revolution of the main engine. Traditionally, these parameters are derived from existing ship databases; however, conventional estimation methods are often limited by outdated models, inadequate noise handling, and restricted capabilities for capturing nonlinear relationships, leading to reduced accuracy and generalization. This study employs a range of machine learning approaches to develop predictive models for estimating the brake power and rate of revolution of bulk carrier main engines. Special emphasis is placed on mitigating noise in both input and output data through the application of a spline smoothing technique. Accordingly, a data preprocessing workflow is proposed, incorporating spline smoothing to enhance the generalization potential of the machine learning models. A comprehensive comparative analysis is conducted across various methods, including four linear regression models, three regression trees, four Gaussian process regression models, two tree ensemble methods, and five neural network models. The performance of the employed machine learning models, evaluated using both raw and smoothed data, is compared in terms of accuracy and generalization capabilities. Results obtained using the smoothed data indicate that the hypertuned Gaussian process regression model exhibits superior accuracy in both validation and testing phases. Furthermore, linear regression models based on smoothed data demonstrated sufficient accuracy for practical implementation, leading to the development of simple predictive formulae for brake power and rate of revolution that are applicable in early-stage ship design.