A comparison of machine learning regression models for critical bus voltage and load mapping with regards to max reactive power in pv buses.

Abstract

The aim of this master thesis is to compare voltage and system loading mapping capabilities of a variety of regression algorithms, such as Adaptive Network based Fuzzy Inference System (ANFIS), Artificial Neural Networks (ANN), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), and Decision Tree (DT). A voltage sensitivity matrix is generated from the power flow Jacobian matrix for a loading scenario near the unstable point. Principal Component Analysis (PCA) is used to separate the system, close to the critical point, in order to group the buses into coherent voltage controlling areas. For different reactive power injection scenarios, we have different bus voltages that can be mapped by the aforementioned regression algorithms. The algorithms are trained with limited amounts of data, in order to establish a fair comparison between them. The present work shows that ANFIS and KNN have a better performance in critical voltage and load prediction when compared to the rest. The academic IEEE 14 and 118 bus systems are employed with all its limits considered, so the results may be reproduced.