Using Dueling Double Q-learning for Voltage Regulation in PV-Rich Distribution Networks

Abstract

The widespread adoption of photovoltaic (PV) systems results in reverse power flows that are already causing overvoltage problems in many distribution networks. To mitigate these voltage problems, distribution companies are using different voltage regulation approaches that exploit the flexibility of existing devices, such as on-load tap changers (OLTC) at primary substations, and the PV inverters themselves. However, the real-time process that determines the most adequate settings becomes more complex with the increasing number of devices that need to be controlled. This work presents a voltage regulation approach that uses the Machine Learning technique called Double Dueling Q-learning (DDQN) as an extremely fast alternative to coordinate the OLTC-fitted transformer at primary substations and the power factor of PV inverters. The proposed voltage regulation approach is assessed using a real Brazilian MV/LV feeder with 108 residential customers and 15 industrial/commercial customers, in which 60% of the residential customers have a PV system. Results validate the ability of using the DDQN control for voltage regulation in real time applications.