Improved robust control of semi-submersible floating offshore wind turbines via Mamba and contrastive learning

Semi-submersible floating offshore wind turbines (FOWTs) exhibit notable potential for harnessing offshore renewable energy resources. To effectively tackle harsh winds and waves, high-quality control systems are imperative. Physical models of FOWTs are first revisited. By integrating innovative tensor product attention with Mamba networks, a data-driven modelling method is presented to establish a comprehensive model bank (CMB) for an FOWT during the offline stage, covering all typical operating conditions. For real-time control design, a contrastive learning (CL)-based model selection mechanism is proposed to simplify the CMB. To quantify model similarity more effectively, instantaneous information of wind and waves extracted via the Hilbert transform is fused with model parameters and operating variables into an augmented feature vector. A joint CL loss function is presented by integrating the conventional CL loss function with the maximum mean discrepancy (MMD) to promote the contrast between the models. The μ-synthesis-based control strategy is proposed using the generated simplified model bank, in which both power generation and vibration suppression are controlled. Simulation results verify the effectiveness of the proposed control scheme. The undesirable power variation is decreased considerably, and fore-and-aft tower motions are also alleviated.