Assessing the robustness of machine learning-based voltage calculations for LV networks

The integration of distributed energy resources (DERs) into low-voltage (LV) distribution networks requires distribution companies to assess customer voltages for new scenarios involving larger generation (from solar PVs) or increased demand (from electric vehicles [EVs]). Voltage calculations typically depend on power flow analyses, which require detailed three-phase electrical models that are often unavailable for LV networks. As an alternative, machine learning (ML) models can leverage historical smart meter data (active power [P], reactive power [Q] and voltage magnitudes [V]) to capture the underlying physics of the LV network and calculate customer voltages for new scenarios without relying on electrical models. However, their robustness in scenarios beyond their training scope is often overlooked, leading to errors in determining how much DER capacity an LV network can handle. This paper evaluates the robustness of ML-based voltage calculations using Neural Networks (NNs) and Linear Regression (LR), in scenarios that remain within (in-domain) and beyond (out-of-domain) the historical data ranges, such as having more solar PV or EVs. A voltage sensitivity analysis assesses each model’s ability to capture the network’s response to changes in P and Q. The study uses synthetic data from a realistic Australian LV network comprising 31 single-phase customers and 25 % PV penetration. Results indicate that LR models calculate voltages more accurately than NNs, especially in out-of-domain scenarios, although all models exhibit limitations in capturing the network’s sensitivity to P and Q. These findings highlight the need for improving ML models to ensure reliable voltage calculations for applications involving DER integration.