DDPG-Based RMS Current Optimization for Dual Active Bridge Converters with Improved Generalization Capability

Conventional intelligent modulation schemes for the dual active bridge (DAB) converter typically depend on detailed circuit parameters. Consequently, modulation strategies obtained through intelligent optimization algorithms must be adjusted or even retrained whenever converter parameters change. To enhance generalization capability and avoid the need for retraining on new hardware platforms, this paper adopts a normalization method for the parameter-independent modeling and optimization of the DAB converter. The normalized model for the DAB converter under triple phase shift (TPS) control is developed using harmonic analysis. This model enables the computation of the normalized values of both output power and root mean square (RMS) current without requiring detailed circuit parameters, especially the switching frequency, series inductance, and the turns ratio of the transformer. Using the proposed normalized model, a parameter-independent optimization problem is summarized. The deep deterministic policy gradient (DDPG) algorithm is employed to solve this parameter-independent optimization problem. The TPS modulation based on the trained neural network can be directly applied to DAB modulation and provides a unified optimization result applicable to converters with different parameters. Furthermore, network distillation and quantization are employed to reduce computational complexity, enabling efficient deployment of the optimized neural network on a field-programmable gate array (FPGA) platform. Experimental results verify that the proposed model and optimization strategy can achieve effective control performance under different circuit parameter conditions.