Deep Neural Network-Based Surrogate Model for Optimal Component Sizing of Power Converters Using Deep Reinforcement Learning

The optimal design of power converters often requires a huge number of simulations and numeric analyses to determine the optimal parameters. This process is time-consuming and results in a high computational cost. Therefore, this paper proposes a deep reinforcement learning (DRL)-based optimization algorithm to optimize the design parameters for power converters using a deep neural network (DNN)-based surrogate model. The surrogate model of power converters can quickly estimate the power efficiency from input parameters without requiring any simulation. The proposed optimization model includes two major steps. In the first step, the surrogate model is trained offline using a large dataset. In the second step, a soft actor-critic-based optimization model interacts with the surrogate model from step 1 to determine the optimal values of design parameters in power converters. Unlike deep Q learning-based methods, the proposed method is able to handle large state and action spaces. In addition, using entropy-regularized reinforcement learning, our proposed method can accelerate and stabilize the learning process and also prevent trapping in local optima. Finally, to show the effectiveness of the proposed method, the performance of different optimization algorithms is compared, considering over ten power converter topologies.

Automated summary

This paper proposes a deep reinforcement learning-based optimization algorithm for power converter design parameters. By using surrogate models and optimization models, the power efficiency can be estimated quickly, and optimal design parameters can be determined. The proposed algorithm can handle large state and action spaces, and can accelerate and stabilize the learning process.