Repository logo
 

Analytic Methods for Electricity Price Forecasting: Application to New Zealand Electricity Market

Files

Thesis(3.25 MB)

Date

2024

Authors

Kapoor, Gaurav

Supervisor

Wichitaksorn, Nuttanan
Zhang, Wenjun

Item type

Thesis

Degree name

Doctor of Philosophy

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

Electricity price forecasting has become a crucial focus for energy market participants in the last few decades. Its importance stems from the lack of efficient electricity storage options, and the uncertainty in its generation and ability to meet real-time demand. This thesis presents three independent forecasting studies for New Zealand electricity prices using diverse methods and models from electricity price forecasting literature. The models, methodologies, and variables considered in these studies are tailored for the New Zealand electricity market, and as such, the results are of direct interest to participants in the New Zealand electricity sector. The first study forecasts daily electricity prices using Markov regime-switching (MRS) models and compares them to an extreme-value theory (EVT) framework. Due to the application of the generalised Pareto distribution for extreme prices, the EVT framework is able to perform better in in-sample density fitting than other models, despite its relative lack of complexity. However, the three-regime MRS model with time-varying transition probabilities presents the best out-of-sample price density fits. The second study employs a mixed-frequency framework to forecast half-hourly electricity prices using hourly weather variables. A mixed-frequency vector autoregressive (MF-VAR) model and the reverse unrestricted mixed-data sampling (RU-MIDAS) model are compared. LASSO regularization stands out as a key factor, consistently enhancing the forecasting performance when it is applied. Inspired by LASSO’s success, the third study further explores the impact of feature selection on forecasting performance. This study compares statistical (GARCH and stochastic volatility) and machine learning models (LASSO-estimated auto-regressive, deep neural network (DNN), long short-term memory (LSTM), gated recurrent unit (GRU), extreme gradient boosting (XGBoost)) for daily electricity price forecasting. A meticulous comparison methodology involves a large number of external features and a variety of feature selection methods, including the LASSO, mutual information, and recursive feature selection. GARCH-t, SV-t, GARCH, and SV with LASSO-selected features consistently outperform benchmark models LEAR and DNN, showcasing performance increases of over 40% compared to GARCH and SV models with all features.

Description

Keywords

Source

DOI

Publisher's version

Rights statement

Permanent link

http://hdl.handle.net/10292/17525

Collections

Doctoral Theses