Machine learning and control applications for intelligent grids

• Main Author: Fong, Ye Cheng
• Other Authors: Hung Dinh Nguyen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering
• Online Access: https://hdl.handle.net/10356/176492
• Institution: Nanyang Technological University

In this project, we investigate the capabilities of Gaussian Process Regression (GPR) in predicting power flow parameters in a power grid. Data points are produced with different types of probability distribution, uniform distribution, and Gaussian distribution, for the GPR model to perform on. Additionally, different types of kernels, the Squared Exponential, Matern, and Rational Quadratic kernels, were used to sieve out the most suited kernel type for our GPR model for this experiment. Combining kernels are also trialed to see if their performance edges out on the standard kernels. This project attempts to explore the versatility of GPR because of its unique feature of predicting uncertainty alongside predicted values. We aim to see if our GPR model is able to perform well under different conditions, for example, comparing the performance of predicting a uniform distributed dataset with the Matern kernel and the performance of a GPR model using the Rational Quadratic kernel. Tests were done using Matlab and Python, with readily available packages like MATPOWER and GPML for us to generate data and excute GPR properties respectively. The 33-bus system from Baran & Wu is used to produce power flow results data, then GPR training and testing is done in Matlab and Python. Results display that the type of probability distribution plays a part in GPR performance as the GPR predictions on a Gaussian distributed dataset produced better accuracy results than the predictions on a uniform distributed dataset. The type of kernel used did not play a big part in the accuracy results.