Power Demand Forecasting for a Hybrid Marine Energy System with Shallow and Deep Learning

To ensure that future autonomous surface ships sail in the most sustainable way, it is crucial to optimize the per-formance of the Energy and Power Management (EPM) system. However, marine EPM systems are complex and often coordinate various distributed energy resources, energy storage systems, and power grids to ensure reliable and safe power delivery. Traditional control methods for marine EPM systems are limited by evaluating processes using simplified component models over a short time horizon, or relying on historical insights gained from earlier journeys, and are not the optimal approach for complex hybrid marine EPM systems. Advanced control strategies, such as Model Predictive Control (MPC), offer a promising control method that considers predicted future system responses over an extended time horizon to determine the best control input, making them an effective strategy for optimizing the performance of hybrid marine EPM systems. However, to learn the onboard energy profiles based on component behavior in a hybrid system from past experiences is not a trivial task, and one of the primary barriers to implementing MPC for marine EPM control. For this reason, in this work, we address the challenge of learning energy profiles for a marine EPM system by utilizing shallow and deep machine learning for total power demand forecasting. The forecast is an essential reference for an MPC-based controller and will enable this control strategy to provide reliable and safe power delivery for hybrid marine EPM systems. The proposed approach compares state-of-the-art machine learning models to identify the best-performing algorithm, considering accuracy and computational requirements. We illustrate the potential of the proposed approach by using real world operational data from a vessel with a hybrid marine EPM system. Results indicate that shallow models, trained on engineered features handcrafted with classical signal processing techniques, allow forecasting the total power demand up to a horizon of 5min with minimal loss in accuracy and a negligible computational burden.