Neural Network Power Flow Approach to Detect Overload and Voltage Anomalies in Low-Voltage Unbalanced Networks, Agnostic of Network Topology

The application of Power Flow (PF) algorithms at Low Voltage (LV) becomes essential, to ensure safe and cost-effective operation. Deterministic approaches do not appear suitable and scalable for LV networks, with a higher risk of non-convergence. The proposed Neural Network Power Flow model (NN-PF) provides accurate power loading, voltage magnitudes and angles in LV unbalanced network, based on nodal consumption and generation power, while being agnostic of the LV network model. Broader dataset is generated for training and testing purposes, including solar generation and undesired voltage events. Despite challenges posed by limited dataset size and the absence of the network topology and features, the NN-PF demonstrates robust performance and high accuracy to identify voltage anomalies and overloads in LV networks. The highest Mean Absolute Error (MAE) is 2e-4 p.u. (0.48 V), 4.6 kW active and 1.51 kVAr reactive power flow at extreme steady-state conditions (V < 0.95 p.u.).