Time Series Forecasting of Reactive Power Support from Smart Converters in SDN Using Machine Learning

Abstract

The climate changes in the last few years created a major need to integrate more renewable energy sources and other low-carbon technologies into the power system network. This causes to face more changes in the power system network, which are particularly visible in distribution power networks. A higher penetration of the distributed energy resources, installed at low- voltage and or medium-voltage levels, creates new challenges for distribution system operators. One of the important issues is to effectively manage the reactive power in a smart distribution network, as the mismatch of the reactive power in a power system network can cause voltage violations in the network. By timely predicting the reactive power, a distribution system operator can make better decisions to avoid any voltage violations. Several conventional techniques like optimal power flow (OPF) control and droop control are used, which are highly dependent on grid models. But machine learning (ML) is an effective approach due to its capability of handling multiple variable data sets and its performance is being independent of grid constraints. Therefore, in this thesis, we propose a machine learning-based approach for time series prediction of reactive power in a smart distribution network. After going through a literature review to find research gaps, a detailed methodology is discussed, highlighting tools used and how they impact on our objectives. Moving forward, a power flow analysis is performed to see the impact of reactive power on SDN. After acquiring all the data required for training algorithms, ML is implemented, and the results are also compared with the optimal power flow method. The results show that the predicted reactive power by the ML approach is very close to the OPF results. However, more improvements can be made by increasing the dataset and changing in the layers of ML algorithms.