<p>Hybrid electric vehicles (HEVs) have attracted much attention due to their high fuel economy and low emissions, and the performance of their energy management strategies (EMSs) is crucial to their energy efficiency. However, existing energy management strategies based on deep reinforcement learning (DRL) still face challenges such as slow convergence, inefficient training, and limited adaptability to hybrid action spaces that involve both discrete and continuous variables. Therefore, based on the Double Actor-Critic with TD Error-Driven Regularization (TDDR) algorithm, this paper investigates the EMS of a two-mode power-split HEV equipped with a complex transmission system consisting of three planetary gear sets. Through kinematic and dynamic analysis of the working modes of a three-planetary-gear two-mode hybrid vehicle, the discrete driving modes and continuous engine power jointly form a hybrid action space. The Gumbel-Softmax reparameterization technique is introduced to enable differentiable sampling of the discrete actions output by the Actor network in the TDDR framework. Simulation results indicate that the TDDR-based EMS outperforms traditional DRL methods in training efficiency, convergence speed, and fuel economy. This study provides an effective energy management solution for two-mode hybrid vehicles with complex planetary gear systems and contributes to the advancement of hybrid vehicle control optimization and fuel-saving technologies.</p>