Vrednovanje algoritama za detekciju stršećih vrijednosti i anomalija u vremenskim nizovima podataka solarnih elektrana

master's thesis

• Title (english): Evaluation of Algorithms for Outlier and Anomaly Detection in Solar Plant Time Series Data
• Author: Domagoj Mustač
• Mentor: Tomislav Petković (mentor)
• Committee member: Tomislav Petković (predsjednik povjerenstva)
• Committee member: Zvonko Kostanjčar (član povjerenstva)
• Committee member: Stjepan Begušić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-11, Croatia

Abstract

Evaluacija algoritama generalno je važan aspekt dobivanja estimacije performansi, te omogućuje dodatne mogućnosti poboljšanja algoritma u obliku optimizacije parametara i slično. Često stručnjaci u domeni solarnih elektrana ručno provjeravaju takve algoritme, što nije idealno zbog nemogućnosti skaliranja na rastući broj individualnih elektrana. Zbog toga je razvijen radni okvir za automatsku validaciju algoritama. Standardne metrike, poput točnosti se nisu pokazale kao dobrim odabirom zbog svoje nesposobnosti da razlikuju i vrednuju specifične elemente koji su bitni za evaluaciju algoritama za detekciju anomalija i stršećih vrijednosti u domeni solarnih elektrana. Ti elementi uključuju stabilnost, kasno otvaranje i zatvaranje zastavica za kvarove, i ostale specifične informacije koje su bitne za evaluaciju. Zbog toga su uz pomoć stručnjaka u domeni kreirane nove metrike pomoću novo napravljenih značajki, gdje svaka metrika evaluira svaki pojedinačni aspekt algoritma, te su sve te metrike kasnije kombinirane sa različitim težinama kako bi dale bolju reprezentaciju konačne kombinirane mjere u usporedni sa standardnim metrikama poput točnosti.

Abstract (english)

Algorithm evaluation is generally an important aspect to get performance estimates, as well as allowing for further algorithm improvements in the form of parameter tuning. It is common that domain experts in the solar power plant industry manually validate machine learning algorithms which is not an ideal solution, as it is not scalable to a growing number of individual power plants. Therefore, an automatic evaluation framework was developed to tackle this issue. The conventional metrics, such as accuracy, were deemed inadequate due to their inability to account for specific aspects that are essential for evaluating a solar power plant anomaly detection algorithm. These aspects include stability, opening and closing faults speed, and other specific information that must be thoroughly tested. New metrics using newly engineered features were therefore created with the help of domain experts, where each metric tests a specific algorithm aspect, and all of these metrics are later combined with different weights to give the most representative final combined score compared to the standard metrics such as accuracy.

• Keywords: anomalija; outlier; strojno učenje; umjetna inteligencija; solarne elektrane; solarna energija; predikcija proizvodnje solarne elektrane; detekcija outliera; evaluacija detekcije outliera
• Keywords (english): anomaly; outlier; machine learning; artificial intelligence; solar power plants; solar energy; solar power plant power prediction; outlier detection; outlier detection evaluation
• Language: english
• URN:NBN: urn:nbn:hr:168:787755
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-10-11 09:54:33