A Compact Neural Network-Based Conversion Loss Model With Hard Constraints for Energy Management

The evolving microgrid (MG) technology integrates various converters for different energy sources and loads, leading to increased complexity in optimization problems due to additional decision variables introduced by power conversion processes in modern energy management systems (EMS). Existing conversion loss models are too complex for optimization. This article presents a neural network-based linear surrogate model to accurately and efficiently approximate power conversion losses in EMS. In EMS, the black-box nature of neural networks (NNs) raises concerns about violating optimization constraints. To address this, the proposed NN model is trained using the augmented Lagrangian method (ALM), enforcing additional hard constraints on the network input/output variables. Moreover, the trained NN is reformulated as mixed-integer linear programming (MILP) constraints, allowing the model to be integrated into energy management (EM) problems solvable by MILP solvers. To optimize the model architecture, a novel neural architecture search (NAS) algorithm is introduced, striking a balance between accuracy and complexity. The case study results demonstrate that the proposed model effectively approximate conversion loss with small absolute errors while satisfying the additional hard constraints. Compared to the manually tuned models, the NAS-generated model achieves superior performance with a compact architecture. Moreover, the reformulated MILP model can be solved efficiently using state-of-the-art MILP solvers, facilitating the integration of surrogate models into EM problems. Lastly, NN example optimal power flow (OPF) problem is presented in the case study, showcasing the practical usability of the proposed model in EM problems.