Quantized deep learning model based Volt-Var control for hosting capacity maximization: a practical case study

Abstract

This study introduces a data-driven Volt-Var control strategy aimed at maximizing hosting capacity (HC) and mitigating voltage violations in distribution networks. Central to the proposed methodology is a quantized one-dimensional convolutional neural network (QCNN), developed to emulate optimal Volt-Var control decisions. Post-training quantization is applied, resulting in tenfold reduction in model size, making QCNN well-suited for deployment on resource-constrained edge devices. The training data is prepared using Modified Reptile Search Algorithm, which determines optimal Volt-Var control set points for HC maximization. The QCNN model trained offline processes measurements from the test network to predict the optimal reactive power reference points, which are supplied to the Decoupled Finite Control Set Model Predictive Controller (D-FCS-MPC). The D-FCS- MPC subsequently determines the optimal inverter switching states to directly regulate reactive powers to the predicted reference values. The methodology is validated using model of a real-world medium voltage (MV) network situated in Vaasa, Finland, known as the Sundom Smart Grid (SSG). Both simulation and OPAL-RT based real time results confirms the effectiveness of the proposed methodology in maximizing hosting capacity and ensuring grid stability.