Ultra-short-term mooring forces forecasting for floating wind turbines with response-frequency-informed deep learning and on-site data

Accurate dynamic response forecasting is crucial for the operational monitoring, maintenance, and dynamic control of floating wind turbines (FWT). In this study, an ultra-short-term forecasting model of mooring line tension for a full-size FWT is developed by combining a Long Short-Term Memory (LSTM) encoder-decoder network with frequency decomposition (FD), i.e., the LSTM-FD method. After presenting the principles of the LSTM-FD-based ultra-short-term forecasting model, full-scaled measurement data from the Hywind Scotland wind farm is used to validate and demonstrate the accuracy of the proposed model. The result shows that the LSTM-FD method has good consistency between different datasets, and higher accuracy than the LSTM without frequency decomposition. For instance, achieving a 10% enhancement in the accuracy of maximum forecasting for Line 1 Bridle 1 over a 60-second horizon. More importantly, various effects of mooring tension across the entire frequency range can be captured through LSTM-FD. In summary, the proposed method can facilitate more precise and timely maintenance scheduling, reduce operational costs, and enhance the overall safety of FWT operations by mitigating the risk of mooring line failures.