A Novel State-Attention-Based Hybrid RNN Controller for Dual Active Bridge Converter With Multiple Degrees of Freedom Modulation

Due to the exceptional adaptability and capacity to accurately capture complex dynamics, artificial intelligence (AI) has emerged as a promising approach in the controller design of dual active bridge (DAB) converter. However, the existing AI-based control strategies of DAB converter typically suffer from heavy computational burden, limited robustness, and lack of interpretability. Besides, owing to the coupling effects of various modulation angles on state variables, DAB systems with multiple degrees of freedom (DOF) modulation exhibit strong nonlinearity. Therefore, their transient control methods are almost exclusively limited to conventional linear controllers, resulting in poor transient performance and stability. To tackle these challenges, a novel state-attention-based hybrid recurrent neural network (RNN) controller for DAB converter with multiple DOF modulation is proposed. First, to derive the state space equations of the nonlinear dynamics of DAB, a general voltage prediction model based on RNN is developed, which is lightweight and capable of adapting to different operation modes and various circuit parameters. Second, a state-attention-based hybrid RNN controller with three hidden layers is introduced to adaptively generate optimal dynamic modulation online. By integrating conventional control laws into RNN, the proposed hybrid RNN cells in hidden layer 1 exhibit enhanced interpretability and low computational burden. Besides, the novel state-attention mechanism in layer 2 is designed to adaptively adjust the attention weights among different control laws derived from hybrid RNNs, yielding superior control performance characterized by rapid convergence, reduced voltage ripple, and enhanced robustness. Finally, based on the preceding layer’s control law, the optimal dynamic modulation is derived in hidden layer 3 by inverting the state-space equations. Simulations and comparative experiments are presented to validate the effectiveness and superiority of the proposed approach.