Abstract
The prediction of wave energy converter (WEC) performance is crucial for guiding optimal design and active control. The traditional model-based prediction method requires strict establishment the system dynamics model, which can be challenging to achieve accurately when the system exhibits inherent nonlinear or artificially introduced nonlinear characteristics. A novel prediction method called multiple-input operator network (MIONet) is proposed for nonlinear WECs to address the low accuracy issued caused by the initial value sensitivity of nonlinear systems. The MIONet introduced in this research integrates wave height and initial response conditions into various branch nets. Time information is inputted into the trunk net, and the final nonlinear WEC response is calculated by performing element-wise product and sum operations on the outputs of the branch nets and the trunk net. A bistable wave energy converter (B-WEC) is used as a typical nonlinear system to apply the new forecasting method. Comparative analysis of the prediction results of other neural networks in both linear WEC and B-WEC reveals that the MIONet exhibits significantly superior predictive accuracy and generalization ability in B-WEC, especially in the intra-well motions. The MIONet provide substantial support for the future development of digital twin models for WECs.