Multi-response prediction and tension warning for a dual-rotor TLP wind turbine

To improve short-term multi-response prediction for a dual-rotor tension-leg-platform wind turbine, multi-source time-series data were generated using STAR-CCM + under regular-wave, irregular-wave, and an additional representative wind–wave–current condition. A parallel feature-fusion PF-TCN-GRU model was developed to jointly predict heave, pitch, and TLP2 tension by combining multi-scale local features extracted by TCN with temporal dependencies learned by GRU. The model achieved R 2 values above 0.997 under regular waves and average R 2 values of 0.9297 and 0.9239 under the two baseline irregular-wave cases. Under the additional wind–wave–current condition, the average R 2 remained 0.9284. Comparative and repeated-run tests showed that PF-TCN-GRU remained competitive with the closest PF-CNN-GRU baseline, although the difference was not statistically significant. Prediction accuracy decreased as the horizon increased from 1 to 4 s, with pitch showing the greatest sensitivity. The TLP2-based P95 warning achieved a precision of 1.000 and a recall of 0.314. The results demonstrate the potential of PF-TCN-GRU for short-term coupled-response prediction and statistical tension-anomaly identification, while further experimental or field validation is still required.