Neural Network Trained Cold-plate Models for Modular Converter Exploration

Designing power electronics for medium voltage (MV) and applications, such as the Modular Multilevel Converter (MMC) MVac-MVdc converter for Multi-Terminal (MT) MVdc systems, requires a multidisciplinary approach to optimize a highly intricate system. One of the main challenges in this field is exploring various MMC implementations and evaluating them through appropriate metrics to determine the most effective approach. To address this challenge, this paper introduces a Virtual Prototyping Process (VPP) for Power Electronic Building Block (PEBB) based systems that leverages neural network-trained coldplate surrogate models. These coldplate surrogate model is useful in managing the complexity of a single thermal interface point connected to multiple coldplate-cooled elements within an MMC Sub-Module Drawer (SMD) to a system-level thermal management system. This approach enables the correlation of the intrinsic power capability of point of source to point of load Power Trains as a function of coolant approach and thermal management system capability to facilitate solutuion space exploration. This exploration is conducted in an evolutionary environment (within the VPP) that balances competing objectives like power density, efficiency, specific power and specific cost. The shipboard Integrated Power and Energy System (IPES) application is used as a use-case example.