Invited talk | Nando Ochoa: Learning Where No Electrical Model Exists: Voltage Calculations from Real Smart Meter Data

Most machine learning for power grids assumes an electrical model exists, so training data can be generated by running power flows. While that might hold at transmission and medium voltage levels, it fails in low voltage (LV) networks, the residential infrastructure where rooftop solar, batteries and electric vehicles connect, because distribution companies rarely have accurate electrical models. This talk takes a different starting point. Using real smart meter data from Victoria, Australia, it shows that machine learning can capture the physics of three-phase LV networks and calculate voltages for any what-if scenario, with no electrical model and no power flow simulations. This enables fast, accurate assessment of what more of these technologies would do to voltages, without ever building an electrical model.