Application of Machine Learning and Mathematical Programming in the Optimization of the Energy Management System for Hybrid-Electric Vessels Having Cyclic Operations

Shipping contributes today to 2.1% of global anthropogenic greenhouse gas emissions and its share is expected to grow in the coming years. At the same time, fuel prices are increasing and companies of the related increase in operational costs. This demands for higher efficiency in ship operations. In these regards, battery-powered vessels are often regarded as a promising solution. The existence of an energy storage element in the system, however, introduces additional challenges in its efficient control. This paper presents the application of machine learning and mathematical programming to the optimization of the energy management system of Diesel-electric vessels with an energy storage system operating according to a cyclical operational profile. The proposed energy management system uses unsupervised exclusive machine learning algorithms, k-means or k-medoids, to learn from prior operations. Then mathematical programming based on mixed-integer linear programming is used to address the problem of the optimal unit commitment by means of optimizing the system’s operations for minimizing fuel consumption. The calculated optimal state of charge of the energy storage system is used as the reference value for a proportional-integral controller during the real-time operations. The proposed energy management system is evaluated through its application to a case study corresponding to a hybrid-electric ferry operating in a urban area having cyclic operations through several stations. The results show that the efficiency of the control action is high with an accuracy ranging between 87% and 99%, when compared to an ideal controller, even in presence of large variations in the operational profile and the charging stations. Between the two tested clustering algorithms, k-means showed higher efficiency in the reduction of fuel consumption in presence of charging stations, while in absence of these, k-medoids showed to provide a better performance.