Abstract
Ocean’s surface currents, although not easily measured by ship’s sensors, affect the movements of any vessel at sea. Especially when dealing with Turret-Moored FPSOs, due to its weathervaning property, there is an intrinsic relationship between the Platform’s motion and environmental conditions. In this sense, we propose using Machine Learning regression algorithms to estimate the surface currents that affect a Turret-Moored FPSO based on data commonly measured on board. These data are expressed by wind speed and direction (measured via anemometer and anemoscope); Platform’s heading (obtained by GPS or Magnetic/Gyro compass); and FPSO’s oscillating motion that is given by the standard deviation of pitch, roll, heave and yaw (measured by MRU sensors and considered as proxy variables of first-order waves’ forces). The prior dataset was composed by local environmental conditions at a specific offshore Basin (in Brazil), observed over ten years at a 3-hour time stamp. This corresponds to approximately 30,000 conditions, each used as input into numerical simulations for a partially loaded FPSO (length between perpendiculars: 257 m, beam: 52 m, draught: 15.6 m). Simulation results provide the Platform’s motion time-series used to generate the final dataset previously mentioned. After dividing this final dataset for train/validation/test into 70/20/10 proportion, a K-means algorithm was fitted to the training data, which grouped it into 3 clusters in which, for each one, 2 ‘specialized’ MultiLayer Perceptron (MLP) Neural Network (one for current’s velocity and another for direction) were implemented for. The mean measured value of current’s velocity of the test dataset is 0.33 m/s, whereas the mean prediction with the method proposed is 0.32 m/s. In current’s direction, the mean measured value is 220°, while the mean prediction is 225°.