Accurate and Efficient Linear Surrogate Models for Conversion Losses in Hybrid AC/DC Microgrids based on Hard Constrained Neural Networks

Fast and accurate conversion loss models are becoming crucial for the reliable and efficient operation of hybrid AC/DC microgrids (MGs). However, the traditionally applied conversion loss surrogate modeling methods are either too simple to capture the nonlinearity of the conversion losses, or too complex to fit in MG energy management problems. In this study, a neural network-based linear surrogate modeling method is developed to provide fast and accurate approximations of conversion losses. We first generate the training and test data using PLECS simulation models. Then, the neural network is trained using the augmented Lagrangian method to enforce additional hard constraints to the conversion loss-related variables. Once trained, the proposed neural network model is reformulated into a mixed-integer linear programming (MILP) model, which can be subsequently used in MG energy management problems. We compare the proposed model against other commonly used linear surrogate models on the test data to examine the model performance. The experiment results indicate that the proposed model yields significantly smaller relative error than other commonly used linear surrogate models and can be solved efficiently using state-of-the-art MILP solvers.