DC–DC boost converters present persistent engineering challenges due to their nonlinear behavior, wide operating ranges, and non-minimum-phase characteristics under continuous conduction. Managing large transient signals, component variations, and constant-power-load effects has traditionally relied on classical control methods with inherent limitations.
A comprehensive technical review examined how deep reinforcement learning can address these control problems through three primary approaches: learning-assisted classical controllers, direct duty-cycle management, and hybrid architectures combining both methods. Rather than proposing AI as a simple replacement for conventional control, the research emphasizes integration of converter physics with machine learning design.
Key findings indicate that successful deep reinforcement learning deployment requires careful attention to several interdependent factors. Action design must reflect physical converter constraints. Reward functions must capture both performance metrics and safety boundaries. Observation timing must align with converter switching dynamics. Safety constraints must remain explicit throughout training and operation.
The review synthesized evidence from Scopus, Web of Science, and IEEE Xplore databases, organizing findings around converter-specific challenges rather than algorithm rankings alone. This approach revealed that hardware-oriented validation—testing on actual power electronic devices rather than simulations alone—remains essential for credible practical implementation.
The research suggests that progress toward real-world deployment requires treating algorithm selection, physical knowledge, action and reward design, observation timing, safety constraints, and validation fidelity as an integrated system rather than independent choices. Hybrid control architectures appear most promising, leveraging machine learning's adaptability while maintaining classical control's predictability and safety guarantees.
For power electronics engineers, this work provides structured guidance for evaluating deep reinforcement learning in converter applications, emphasizing that successful implementation demands equal attention to artificial intelligence methodology and fundamental power electronics physics.
