A deep learning model for predicting mechanical behaviors of dynamic power cable of offshore floating wind turbine

Abstract

Dynamic power cable acts as both electricity current channel and information channel between offshore floating wind turbine and substation. Key role as it plays, the mechanical behaviors of dynamic power cable in operation is complicated, owing to the multiple internal loads involved in the cross section and the drastic stress distribution along the cable. In this paper, a multi-task integrated model based on LSTM is proposed to realize both the tension and the bending moment prediction at several fatigue-prone locations. Regarding the parameter combination in high dimension and time-consuming search for the optimum, halving grid search algorithm is applied to conduct hyper-parameters optimization with higher efficiency over wider range. Additionally, due to the motion effects brought by the buoyancy section, the prediction accuracy of the model at locations other than the hang-off point is lower, which is resolved by introducing an subsea sensor at the hog bend to provide additional 3-DOF motion inputs. The improvement brought by the additional inputs at the other points than the hang-off point can be up to 10%. The reliability of the multi-task integrated model is evaluated in a stochastic irregular wave generated by a different random seed as well as wave parameters scaled at different ratios, and a satisfying consistency is observed. Furthermore, the proposed model is applied in extreme sea state, in which the model exhibits comparable performance with its performance in operational sea state. The desirable performance in both the operational and extreme sea states demonstrates the robustness of the model, and implies its potential in more various sea states.