A 2D-CNNpred-MIRB framework for cross-condition response prediction of floating offshore wind turbine platforms

Real-time motion forecasting of floating offshore wind turbine platforms is challenging due to continuously varying wind–wave excitations and strongly nonlinear, non-stationary, cross-condition responses. Most existing data-driven methods treat responses as one-dimensional series and require condition-specific training, limiting unified prediction under diverse sea states. This study proposes a 2D-CNNpred-MIRB framework for cross-condition motion forecasting of a 15 MW floating platform. Wind, wave, and historical response sequences are transformed into a 2D tensor, allowing joint capture of local temporal features and condition-dependent evolution. A 2D convolutional neural network extracts high-level tensor features, integrated with a multi-input recursive BiLSTM for stepwise multi-horizon prediction. Input-feature enhancement and hyperparameter optimization further improve accuracy–efficiency balance. For the typhoon cases, the same model architecture is retained, with minor hyperparameter re-tuning to account for their distinct response characteristics. Validation under normal and typhoon sea states shows average R 2 above 0.972 across conditions, outperforming representative baselines. Prediction is 649× faster than high-fidelity OpenFAST simulations. Embedded into a 15 MW turbine's PI pitch-control loop, the model maintains feasible online performance, supporting its potential for real-time motion forecasting and control-oriented applications within the considered sea-state range.