Implementation of Data- and Simulation-Driven Models for Ship Performance and its Application in Weather Routing

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a comparative analysis of model fidelity in weather routing, useful for researchers working on ship energy efficiency and operational optimization.

🏢実務担当者:Shipping companies can adopt the grey-box model correction approach to enhance existing performance models and achieve fuel savings.

🏛政策担当者:IMO regulators and national maritime authorities can reference the quantified fuel savings to support policies promoting operational efficiency measures.

As the global shipping industry seeks to reduce greenhouse gas (GHG) emissions in order to comply with regulations coming into force from the International Maritime Organisation (IMO), different approaches to achieve this reduction are being investigated. To support this endeavour, the German R&D project MariData developed a decision-support system (DSS) for the energy management of ships which takes simulation data, measurements on board and geo-information into account. The two main goals of this DSS are to provide better routing recommendations and to give an insight into which resistance components influence the ship performance and thus the GHG emissions. This information can then be used to decide on different retro-fitting options. This study has been conducted using E.U. Copernicus Marine Service Information, NOAA environmental data and ship operational data (e.g. speed through water, main engine power and speed, draught etc.) collected within the scope of MariData. The resulting dataset covering roughly one year was processed by using threshold filters for the standard deviations of main engine speed, load and speed through water in one-hour long sliding windows with a threshold value of 1 % of their total range. This resulted in a dataset of around 900 hours with relatively steady ship operation, from now on referred to as the filtered dataset, which was used to validate the white-box model for ship hydro- and aerodynamics developed in the same scope. The prediction accuracy and its relevance for routing has been analysed in Marzi et al. (2024). For two specific routes a fuel saving potential of around 7 % was discovered. The numerical predictions were added to the filtered dataset. This contribution provides insight into how the fidelity of the underlying consumption model affects weather routing with regard to ship performance. Four different approaches were evaluated and compared: a simple white-box model which predicts the machine load based on empirical formulas and only considers the added resistance due to wind, a sophisticated white-box model which was the result of the MariData project and is based on potential flow and Reynolds-averaged Navier-Stokes (RANS) simulations for different resistance components (e.g., calm water resistance, added wave resistance) and propulsive performance, a grey-box model which consists of the prediction of the sophisticated white-box model (b) plus a correction model trained on the difference between the measured machine load on board of the test ship and the predicted machine load of the sophisticated white-box model and, finally, a black-box model which was trained on the measured machine load data. For the grey- and black-box models (c) and (d), multiple hyperparameter optimizations were conducted and Gaussian process (GP) models as well as neural networks (NN) were tested for the training. The Weather Rooting Tool package (https://github.com/52North/WeatherRoutingTool, Marzi et al., 2024) was applied for the weather routing. In particular, a genetic algorithm based on the Python library pymoo (Deb et al., 2002) was utilised. First, the accuracy of the different models was analysed by comparing their predictions to the measured machine loads for fixed routes and travel time. The sophisticated white-box model (b) showed to predict the percentage of the nominal machine load with an mean absolute error (MAE) of 4.32 %. For (c) and (d) the Gaussian process models trained on the filtered dataset performed best with a MAE of 0.31 % and 0.17 % while the neural networks trained on the complete dataset outperformed the NNs on the filtered dataset. The best MAE for the grey-box model (c) of a NN was 0.42 % and for the black-box model 0.41 %. The second analysis looked at the performance of the four approaches (a-d) when used as ingredients for the weather routing in three weather scenarios in the Bay of Biscay: one artificial scenario for average weather conditions, one artificial scenario for rough weather conditions and one real-weather scenario for rough weather. The artificial scenarios enabled to investigate the four models in a controlled environment for weather conditions that are typical for the Bay of Biscay as well as for weather conditions that maximise potential savings if the ship deviates from the shortest route. The real-weather scenario, however, showed to what extent the significance of model precision decreases due to the complexity of real weather conditions. As a result of the second analysis, the spatial difference of the routes for the four consumption models and the three weather scenarios will be summarized as well as the results for the fuel consumption in dependence of the travel distance. Based on the behaviour of the models for the three weather scenarios, conclusions will be drawn for the necessary model precision in average and rough weather conditions.  

• semanticscholar https://upcommons.upc.edu/bitstreams/bf313eaf-4f8d-4c7b-9e07-cee08d552fdd/downloadfirst seen 2026-06-29 06:22:07 · last seen 2026-07-05 05:46:03