Physics-Informed Fully Convolutional Network-Based Power Flow Analysis for Multi-Terminal MVDC Distribution Systems

Numerical methods in power flow (PF) studies for medium-voltage direct current (MVDC) distribution systems require repetitive computations, particularly in scenarios with time-variable facilities that often alter the system operation points. Conventional neural networks (NNs), though efficient in rapid PF calculations, face accuracy challenges with untrained data distributions and varied topology structures. This highlights the need for more robust approaches to improve reliability in diverse scenarios. This paper proposes a physics-informed fully convolutional network (PI-FCN) to address this issue. The architecture of the PI-FCN is enhanced with the inclusion of two additional layers: i) a channel combination layer and ii) a physics operation layer. The former channel combination layer strengthens the model feature extraction capability by converting all input channels constituted by initial PF data matrix into dc voltage, current and line conductance matrix channels. The latter physics operation layer reformulates the combined input channels by physical connections in MVDC systems. The new layers enhance the prediction accuracy and allow generalization of the model. Five multi-terminal MVDC (MT-MVDC) distribution networks with different dc voltage levels and network layouts are used to verify the superiority of proposed PI-FCN compared to other NNs in fixed and varied topology structures.