A benchmark study on ship speed prediction models in arctic conditions: Machine learning, process-based and hybrid approaches

Abstract

Ship speed forecasting in harsh and uncertain ice conditions is a core technology enabling Arctic fleet planning at strategical, tactical, and operational levels of maritime logistics. In this benchmark study, we examine five alternative models that differ by their nature (process-oriented, data-driven, and hybrid approaches) and set of predictors (combinations of AIS and ice parameters). To develop these models, AIS data on YamalMax LNG carriers and ice data from diagnostic charts were merged in a joint dataset covering the period 2017–2022. A detailed description of this dataset is given in the article, including spatiotemporal variability of ice parameters and ship speed. Actual and predicted speeds are compared by five statistical indicators not only in the case of speed estimation at a point, but also when calculating average speed over long-distance route segments. The article analyses the difference in accuracy of speed predictions in various geographical areas of the Arctic and during different seasons. The degree of influence of model parameters is estimated using a SHAP approach. Several conclusions about the architecture of forecasting models and their predictors are drawn as a result.