Classification of electricity consumer profiles
using Artificial Intelligence techniques

Σπανού Καπαντώνη, Αγγελική; Spanou Kapantoni, Angeliki

dc.contributor.author	Σπανού Καπαντώνη, Αγγελική	el
dc.contributor.author	Spanou Kapantoni, Angeliki	en
dc.date.accessioned	2026-09-02T06:47:21Z
dc.date.available	2026-09-02T06:47:21Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/65668
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.33362
dc.rights	Default License
dc.subject	Electricity consumer classification	en
dc.subject	Machine learning pipeline	en
dc.subject	Transformers	en
dc.subject	Smart grids	en
dc.subject	Time series data	en
dc.subject	Κατανάλωση ενέργειας	el
dc.subject	Μηχανική μάθηση	el
dc.subject	Χρονοσειρές	el
dc.subject	Προφίλ καταναλωτή ενέργειας	el
dc.subject	Ταξινόμηση καταναλωτών	el
dc.title	Classification of electricity consumer profiles using Artificial Intelligence techniques	en
dc.title	Ταξινόμηση Προφίλ Καταναλωτών Ηλεκτρικής Ενέργειας με Τεχνικές Τεχνητής Νοημοσύνης	el
dc.contributor.department	Artificial Intelligence and Learning Systems Laboratory	el
heal.type	bachelorThesis
heal.classification	Μηχανική Μάθηση	el
heal.classification	Τεχνητή Νοημοσύνη	el
heal.language	en
heal.access	campus
heal.recordProvider	ntua	el
heal.publicationDate	2026-07-08
heal.abstract	The mass rollout of smart electricity meters by the Hellenic Electricity Distribution Network Operator (HEDNO/DEDDIE), produces per-consumer hourly consumption time series that enable a range of analytics tasks. This thesis addresses one such task: the au- tomatic classification of electricity consumers into behavioural categories based on their consumption signature, rather than their self-declared registry label. The work pursues two complementary goals: a verification use case, in which a mismatch between a con- sumer’s declared category and the category inferred from actual usage is flagged as a potential case of misdeclaration or non-technical loss; and a general-purpose categorisa- tion use case supporting tariff design, demand forecasting, and network planning. Using two years of hourly active-load data for tens of thousands of HEDNO con- sumers, the thesis classifies six behavioural groups: the four registry-native categories Agricultural, Industrial, Commercial, and Public Services, together with Tourism and Food Service. An unsupervised cluster analysis (UMAP, DBSCAN, k-means) is first used to characterise class separability in raw feature space, followed by an extensive feature- engineering stage, statistical, temporal, seasonal, and activity-based descriptors, reduced from 373 to 174 features via SHAP-guided selection. Classification is then approached with both classical machine-learning models (LightGBM, Random Forest, Logistic Regres- sion) and deep time-series architectures (PatchTST, with weekly/daily/ monthly patch branches and self-supervised pretraining, TimesNet, MiniRocket, InceptionTime), and fi- nally combined through a soft-vote ensemble and a confidence-gated cascade that routes only the consumers LightGBM itself flags as ambiguous to binary PatchTST experts. The tabular-fused cascade of LightGBM and PatchTST is the best-performing config- uration overall, reaching a test Macro-F1 of 0.8648 (86.6% accuracy) on the four-class dataset, a gain of 4.67 Macro-F1 points over standalone LightGBM and larger than the corresponding soft-vote ensemble’s improvement. Notably, plain LightGBM already out- performs every deep-learning model trained from scratch while requiring only CPU train- ing, underscoring that careful data handling and feature fusion matter more than archi- tectural sophistication on this dataset. These results support the proposed pipeline as both an accurate general-purpose classifier and a practical, interpretable tool for flagging declared-versus- behavioural consumption mismatches for distribution-operator review.	en
heal.advisorName	Voulodimos, Athanasios	en
heal.committeeMemberName	Vescoukis, Vassilios	en
heal.committeeMemberName	Voulodimos, Athanasios	en
heal.committeeMemberName	Stamou, Giorgos	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών. Τομέας Τεχνολογίας Πληροφορικής και Υπολογιστών. Εργαστήριο Συστημάτων Τεχνητής Νοημοσύνης και Μάθησης	el
heal.academicPublisherID	ntua
heal.numberOfPages	130
heal.fullTextAvailability	false