Learning the Dynamics of Future Marine Microgrids Using Temporal Convolutional Neural Network

TL;DR

The paper addresses dynamic modeling challenges in MVDC shipboard microgrids arising from highly dynamic loads and confidentiality constraints by proposing a data-driven identification approach using Temporal Convolutional Networks (TCNs). It introduces a TCN architecture with dilated causal convolutions, residual connections, weight normalization, and a fully connected output head trained with an $L_{MSE}$ loss to learn state dynamics from measurements. Case studies show high predictive accuracy and strong generalization to unseen pulsed power patterns, using 100 s of training data at a sampling rate of 0.5 ms. The work demonstrates potential for data-driven stability assessment and control design (e.g., MPC) in naval MVDC SMGs, with future work on incorporating physical priors and closed-loop performance.

Abstract

Medium-voltage direct-current (MVDC) ship-board microgrids (SMGs) are the state-of-the-art architecture for onboard power distribution in navy. These systems are considered to be highly dynamic due to high penetration of power electronic converters and volatile load patterns such as pulsed-power load (PPL) and propulsion motors demand variation. Obtaining the dynamic model of an MVDC SMG is a challenging task due to the confidentiality of system components models and uncertainty in the dynamic models through time. In this paper, a dynamic identification framework based on a temporal convolutional neural network (TCN) is developed to learn the system dynamics from measurement data. Different kinds of testing scenarios are implemented, and the testing results show that this approach achieves an exceptional performance and high generalization ability, thus holding substantial promise for development of advanced data-driven control strategies and stability prediction of the system.

Figures (9)

• Figure 1: Topology of the MVDC SMG
• Figure 2: Reduced-order model of the MVDC SMG.
• Figure 3: Structure of the proposed TCN architecture for identifying dynamics of SMGs.
• Figure 4: Dilated causal convolutions operation process.
• Figure 5: PPL variation for training and testing sets