University of Thessaly (UTH)
Design strategies towards the enhancement of short-term forecasting in the energy sector
Abstract
dc:descriptionShort-term forecasting processes constitute an integral part of data analysis in the energy sector since their integration in demand response programs, energy management systems, smart grid and energy market applications is valuable towards the examination of complex and volatile time series variables such as load and electricity price. Short-term load forecasting models offer valuable insight towards consumption patterns through the inspection of influential factors and introduce intelligent ways of monitoring electricity demand as well as the occurrence of irregular events in order to improve the decision-making processes of electric utilities and reinforce grid stability respectively. Short-term electricity price forecasting models address the challenge of price volatility and contribute towards the development of robust strategies towards efficient resource management and optimal energy transactions for all types of energy market participants and consumers. It is evident that methods focusing on load and electricity price time series follow a similar structure including preprocessing, forecasting and output modules for the estimation of the target variables after data collection. Therefore, this dissertation acknowledges the shared and overlapping structure of those forecasting processes and addresses prominent challenges and research gaps associated with each component through the development of optimal design strategies that improve the overall model performance. The study of the preprocessing module led to the assessment of robust feature selection and highlighted the role of rule generation for efficient examination of the studied environments. Since prominent challenges in data preprocessing are connected to dataset dimensionality and feature interpretability, a method towards the generation of a compact and interpretable set of rules through hybrid feature selection was proposed. Furthermore, the study of the main forecasting framework denoted challenges with regards to standalone, combinatorial and meta-modeling design philosophies. In standalone modeling, the uncertainty surrounding estimator selection due to the insufficient exploration of edge cases often hinders research progress and leads to confusion with regards to training behavior. Consequently, a comparative study examining the baseline performance of neural network methodologies for high resolution predictions addresses one of the edge cases where the brevity of the training process and time constraints could provoke this uncertainty. Regarding combinatorial modeling, the uncertainty of estimation member selection in ensemble methods coupled with the challenges of concept and data drift could lead to arbitrary design decisions and suboptimal model combinations. As a result, a novel design strategy focusing on the deterministic selection of estimator members based on the structural characteristics of peak and non-peak indices was proposed in order to generate performant ensemble learning models. Moreover, the examination of meta-modeling approaches highlighted the performance benefits of additional forecasting layers and led to the introduction of a forecasting approach that estimated load consumption through the inspection of similarity and causality for the derivation of alternative time series representations. This approach improved the error metrics compared to the base LSTM ensemble model, denoting the impact of community factors when the quality of the input dataset is far from ideal. Lastly, following this a posteriori design method, the study of the output module identified the need for performance refinement through additional structures that estimate and minimize error values towards increased model stability and improved accuracy. In this scope, an error compensation module was developed towards the performance improvement of a deep learning structure for the task of short-term electricity price forecasting. This approach introduced an autoregressive model for the estimation of residual training error, resulting in more consistent predictions and overall lower error metrics when tested in different training scenarios. Additionally, this method discussed the potential addition of hyperparameters that configure the error compensation module for future applications and benchmarks. The extension of the strategies presented in this dissertation could enable the development of more flexible and adaptive forecasting pipelines that could enhance the capabilities of future energy applications.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kontogiannis, Dimitrios
- Κοντογιάννης, Δημήτριος
Subjects
dc:subject × 32- Πρόβλεψη φορτίου
- Πρόβλεψη τιμής ηλεκτρικής ενέργειας
- Ανταπόκριση ζήτησης
- Μηχανική μάθηση
- Νευρωνικά δίκτυα
- Τεχνητή νοημοσύνη
- Παλινδρόμηση
- Εποπτευόμενη μάθηση
- Ασαφής λογική
- Μηχανική χαρακτηριστικών
- Επιλογή μοντέλου
- Μετα-μοντελοποίηση
- Σχεδίαση βραχυπρόθεσμης πρόβλεψης
- Load forecasting
- Electricity price forecasting
- Demand response
- Machine learning
- Neural networks
- Artificial intelligence
- Regression
- Supervised learning
- Fuzzy logic
- Feature engineering
- Model selection
- Meta-modeling
- Short-term forecasting design
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Ηλεκτρική και Ηλεκτρονική μηχανική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Electrical and Electronic Engineering
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/54045