Author

Jie Han

ORCID Identifier(s)

0000-0001-8003-8604

Graduation Semester and Year

2023

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Industrial Engineering

Department

Industrial and Manufacturing Systems Engineering

First Advisor

Shouyi Wang

Abstract

Electricity price forecasting (EPF) is a crucial task for market participants seeking informed decisions in day-ahead electricity markets. The increasing penetration of stochastic renewable energy and the deregulation of electricity markets pose challenges to electricity price forecasting. Given the dependence of electricity prices on stochastic factors such as weather conditions, market dynamics, and customer behaviors, deterministic forecasting methods offer limited insight into the potential future states of energy prices in highly stochastic markets. In this study, a transformer-based electricity price forecasting (TDEPF) model was developed, utilizing a two-step training process and demonstrating superior performance compared to typical RNN models. Subsequently, three probabilistic forecasting models were introduced: quantile regression (QR-Transformer), transformer-based composite quantile regression (TCQR), and Gaussian Process combined with transformer (GP-Transformer). These models can produce multi-step ahead forecasts using multivariate time series as inputs. The application of these models was demonstrated in the context of day-ahead electricity price forecasting, utilizing five years of data from the Electric Reliability Council of Texas (ERCOT). Among the probabilistic models, TCQR emerged as the most effective, as evidenced by metrics such as CRPS, pinball loss, and Winkler score. To evaluate the accuracy of peak time prediction, a novel metric called Winkler-peak score was introduced. Additionally, to regulate the convergence of quantiles near the 0% or 100% level, MSE-regularized pinball loss was proposed. Simultaneously, to prioritize peak time prediction, Winkler-peak score regularized pinball loss was proposed. Both regularization methods on pinball loss were demonstrated to effectively achieve their respective goals when applied in TCQR. Given that deep learning methods operate as black boxes and lack a guarantee of coverage rate, a deeper investigation into uncertainty quantification was undertaken on probabilistic forecasting results. Conformal prediction has been established as capable of constructing prediction intervals with statistically guaranteed coverage rates. However, existing research has not explored applications in either multivariate time series forecasting or multi-step ahead forecasting. Hence, in this study, adaptive quantile random forest (AQRF) and adaptive conformal residual fitting (ACRF) were introduced. Both methods can attain the target coverage rate, and AQRF, in particular, can achieve a narrower bandwidth compared to benchmark models such as conformal prediction and quantile random forest methods. In conclusion, this research can furnish reliable day-ahead electricity price forecasts, aiding decision-makers in making informed decisions within the electricity market.

Keywords

Probabilistic forecasting, Multivariate time series, Conformal prediction, Deep learning

Disciplines

Engineering | Operations Research, Systems Engineering and Industrial Engineering

Comments

Degree granted by The University of Texas at Arlington

32000-2.zip (11771 kB)

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.