Model-Free Voltage Calculations for PV-Rich LV Networks: Smart Meter Data and Deep Neural Networks

The widespread adoption of residential photovoltaic (PV) systems is causing voltage rise issues on low voltage (LV) networks. Operationally, voltage calculations can help determining specific settings (e.g., PV curtailment). From a planning perspective, voltage calculations can be used to assess the PV hosting capacity. However, voltage calculations normally require power flow analyses and, consequently, detailed three-phase LV network models which, in practice, are not readily available. This paper proposes a model-free voltage calculation approach that uses a Deep Neural Network (DNN) trained to capture the nonlinear relationships among historical single-phase smart meter data (P, Q, V) and the corresponding LV feeder. A methodology is proposed to determine hyperparameters and parameters suitable for the investigated LV feeder. Results using an Australian realistic LV feeder, demonstrate that the approach calculates voltages accurately for PV injections that are significantly different from the historical data, i.e., is suitable for any type of what-if scenario.