<p>This paper proposes a neural network-based model predictive control (NN-MPC) scheme, trained with Bayesian regularization, to address the rapid growth of online computational load that arises in conventional MPC for modular multilevel converter (MMC) regenerative braking energy-feedback devices as the number of submodules increases. A discrete model of the device is first formulated to define the MPC inputs and outputs. System state data under classical MPC are then generated in MATLAB/Simulink; the AC-side current is preprocessed using variational mode decomposition (VMD) to suppress high-frequency harmonic interference while preserving amplitude–frequency characteristics. On this basis, a back-propagation neural network is constructed with electrical state variables as inputs and the per-arm submodule insertion commands as outputs. Bayesian regularization is used to optimize the network weights, enhancing generalization and prediction accuracy. The trained NN-MPC effectively approximates the MPC optimal control law in real time, thereby avoiding online rolling optimization. Simulation and hardware-in-the-loop (HIL) results show that, relative to classical MPC, the proposed NN-MPC achieves comparable performance in DC current tracking, submodule capacitor voltage balancing, and circulating current suppression, while reducing the online computational burden to a constant, horizon-independent level and improving dynamic response. Under grid voltage sag and sudden load change conditions, the controller maintains low AC current total harmonic distortion and stable DC-link regulation, demonstrating robustness.</p>