Machine Learning Based Moored Ship Movement Prediction

Abstract

Several port authorities are involved in the R+D+i projects for developing port management decision-making tools. We recorded the movements of 46 ships in the Outer Port of Punta Langosteira (A Coruña, Spain) from 2015 until 2020. Using this data, we created neural networks and gradient boosting models that predict the six degrees of freedom of a moored vessel from ocean-meteorological data and ship characteristics. The best models achieve, for the surge, sway, heave, roll, pitch and yaw movements, a 0.99, 0.99, 0.95, 0.99, 0.98 and 0.98 R2 in training and have a 0.10 m, 0.11 m, 0.09 m, 0.9°, 0.11° and 0.15° RMSE in testing, all below 10% of the corresponding movement range. Using these models with forecast data for the weather conditions and sea state and the ship characteristics and berthing location, we can predict the ship movements several days in advance. These results are good enough to reliably compare the models' predictions with the limiting motion criteria for safe working conditions of ship (un) loading operations, helping us decide the best location for operation and when to stop operations more precisely, thus minimizing the economic impact of cargo ships unable to operate.

Multi-direction prediction based on SALSTM model for ship motion

“…Q(query) is a query matrix obtained by multiplying an input vector by a weight matrix; K(key) is a key matrix obtained by multiplying an input vector by a weight matrix; V(value) is a value matrix obtained by multiplying an input vector by a weight matrix. The self-attention mechanism can be expressed by formulas (9) and (10).…”

“…The four evaluation metrics are Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R 2 ). Alvarellos et al [9] suggested that RMSE alone cannot measure the quality of the model without comparing it with the range of motion predicted by the model, and that a good model should have an RMSE that is less than or equal to 10% of its predicted range of motion. Chicco et al [10] argued that R 2 is more informative and reliable than other evaluation metrics, so this paper also includes R 2 to evaluate the model's performance.…”

Vessel Velocity Estimation and Docking Analysis: A Computer Vision Approach

“…Regarding the docking problem, Alvarellos et al [5] presented a new computer vision technique for monitoring docked ships. The proposed technique uses visual feature correlation of ship images to estimate its movement over time.…”

“…Currently, in SUAPE, port management relies on the labor of employees to identify the docking and undocking. It is important to note that in several ports, the criteria of docking and undocking are established based on subjective observations of operators and shipmasters, without particularized studies to support them [5].…”

Ship Attitude Prediction Model Based on Cross-Parallel Algorithm Optimized Neural Network

“…The prediction methods of the ship motion and attitude are mainly divided into three categories: mathematical model, statistical model, and machine learning model [5]. Ghadimi P. et al [6] proposed a mathematical model which used the asymmetric 2D + T theory to predict the attitude of the planning craft.…”