University of West Attica
Ανάπτυξη επιβλεπόμενων και μη επιβλεπόμενων μεθόδων μηχανικής μάθησης για την μοντελοποίηση χρονικά εξελισσόμενων συστημάτων με εφαρμογές σε έξυπνα ηλεκτρικά δίκτυα
Abstract
dc:descriptionThe present doctoral thesis concerns the study of time evolving systems with the aim of developing methodologies for their modeling and, their consequent use in the problem of forecasting the consumption of electrical load from medium and high voltage substations. Their operation is based on the algorithmic adaptation of the relationship that describes the input and output data in the form of time sequences or time series. The main purpose of this doctoral thesis is to propose and investigate the application of new supervised and unsupervised machine learning methods for prediction of features extracted from time series. Consequently, an important part of the theoretical section of this work is occupied by time series analysis as a tool for modeling of time evolving systems and the study of time series forecasting. A thorough literature review is then conducted on the methods that have been proposed for electric load forecasting, through which, specific research gaps that arise during the investigation of the above problem are highlighted. Initially, it is observed that despite the fact that a large number of machine learning methodologies have been proposed for the prediction of electrical load, however, none of them can outperform all the rest. This is mainly due to the existence of non-linearity and the different stochastic properties that characterize the time series of electric load consumption. An additional factor that worsens these characteristics stems from the new electrical grid paradigm imposed by the transition from conventional to so-called smart grids. As a result, the smart grid environment introduces two-way power flow between users and the utility grid, resulting in the load data containing both production and consumption measurements, and the so-called mixed load time series exhibiting high amount of uncertainty. It is therefore deemed necessary to search for novel and robust machine learning methods, able to effectively deal with the aforementioned phenomena and demonstrate global superiority over other competitors. Another shortcoming that has been identified is the scant consideration of mixed load forecasting, as opposed to the extensive study of forecasting electricity generation or consumption individually. Based on these observations, the development of three new forecasting methods is proposed. The first of these is based on the theory of sparse representation and dictionary learning, with its main proposition being the integration of the atoms and the corresponding sparse coefficients in a hierarchical tree structure. The superiority of the proposed method lies in its low complexity and its inherent ability to efficiently represent the nonlinear characteristics of the mixed load while remaining robust to overfitting effects. The next methodology that is introduced concerns an ensemble model which is trained by various machine learning methods. Specifically, techniques from: artificial neural networks, linear regression, support vector regression, random forests and sparse representation are combined. The final predicted values are calculated with the use of an online decision mechanism based on the weighting of the sub-models according to their past performance. The set of proposed forecasting methodologies is completed with a GNN model, which is enhanced with the properties of Visibility graphs. According to this approach, which is called a Visibility graph neural network (VGNN), the mixed load time series is transformed into an undirected graph using the criterion of natural visibility between the nodes of a graph, and then an adjacency matrix is employed for training a GNN. The next step was the evaluation of the proposed methods. The scope for their experimental testing is drawn from the field of smart grids. Specifically, the mixed electric load forecasting is taking place, using real data from a high and medium voltage substation. After the detailed introduction to smart grids and their essential contribution to load forecasting and, furtherly, to their efficient operation, follows the description of the experimental protocol of the aforementioned simulations. According to the obtained results, the proposed machine learning methodologies manage to provide reliable, robust and accurate forecasts for cases of different forecast time horizons ranging from 15 minutes to 24 hours ahead but also different execution scenarios involving data spanning periods of idiosyncratic consumption patterns.
Degree
thesis:*- Grantor dc:publisher
- University of West Attica
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Giamarelos, Nikolaos
- Γιαμαρέλος, Νικόλαος
Subjects
dc:subject × 26- Επιβλεπόμενη μάθηση
- Mη επιβλεπόμενη μάθηση
- Aραιή αναπαράσταση
- Πρόβλεψη μικτού φορτίου
- Νευρωνικά δίκτυα γράφων
- Γράφοι ορατότητας
- Εκμάθηση ανσάμπλ
- Supervised learning
- Unsupervised learning
- Sparse representation
- Mixed load forecasting
- Graph neural networks
- Visibility graphs
- Ensemble learning
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Ηλεκτρική και Ηλεκτρονική μηχανική
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Electrical and Electronic Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/54881