Data-driven Volt/Var control based on constrained temporal convolutional networks with a corrective mechanism

Abstract

The increasing renewable energy sources such as photovoltaics (PVs) systems in distribution networks cause frequent voltage violations. Volt-Var control (VVC) has been successfully integrated into distribution networks for reducing power loss and optimizing voltage profiles. By coordinating conventional devices like capacitor banks (CBs) as well as smart inverters, this paper proposes an innovative data-driven VVC strategy using constrained temporal convolutional networks (C-TCN) with a corrective mechanism to reduce power losses and mitigate voltage violations. In the offline training process, the reactive power optimization problem involving CBs and PV systems is formulated as a mixed-integer second-order cone programming (MISOCP). The solutions and the historical operational data form the dataset for training purposes. The proposed C-TCN with a corrective layer gives an additional loss penalty to networks for results which not meet the operational constraints. In online implementations, the well-trained networks give optimization solutions based on real-time system measurements to deal with fluctuations in distribution networks. The proposed method is validated on a modified IEEE 33-bus testing distribution feeder, and numerical results verify effectiveness in mitigating voltage deviations and improving the optimization results compared with other benchmarks.