Electrical and Computer Engineering ETDs

Publication Date

Summer 8-1-2022

Abstract

The elevated costs that incur power grid stakeholders due to forecasting errors in power load demand have created the need for forecasting methods that provide accurate predictions and allow for assessing the reliability of their predictions. This thesis proposes a probabilistic forecasting method for multi-step ahead forecasting.

In particular, it presents a probabilistic method to perform a 24-hours-ahead power load forecasting that arises as the combination of Gaussian Process regressors with NMF (nonnegative matrix factorization) and integrates the advantages of both methods. Instead of training 24 independent processes for each hour of the predicted day, this work proposes to factorize the 24 hours power profile as the additive composition of a reduced number of K latent components estimated with K independent Gaussian Processes. This reduces considerably the training time invested in the regressors and introduces some interpretability to the model. The proposed method is compared with models built exclusively with Gaussian Processes in a state-of-the-art data set of aggregated load. In all cases, the methods showed comparable results to existing methods, in similar conditions in the literature, with MAPE values between 2% and 5% and improved density estimation.

Keywords

machine learning, multi-task Gaussian processes, nonnegative matrix factorization, short-term power load forecasting, probabilistic power load forecasting

Document Type

Dissertation

Language

English

Level of Degree

Doctoral

Department Name

Electrical and Computer Engineering

First Committee Member (Chair)

Dr. Manel Martinez-Ramon

Second Committee Member

Dr. Marios Pattichis

Third Committee Member

Dr. Ramiro Jordan

Fourth Committee Member

Dr. Sandra Biedron

Fifth Committee Member

Dr. Fernando Moreu

Download

Share

COinS