| Abstract | U današnje vrijeme, elektroenergetski sustavi (EES) postaju sve složeniji zbog uključivanja obnovljivih izvora energije i težih ispunjavanja zahtjeva stabilnosti, pouzdanosti i održive proizvodnje električne energije. S obzirom na navedene izazove sve se češće koriste suvremene metode strojnog učenja i stohastičkih modela u svrhu predikcije proizvodnje pojedine jedinice unutar sustava. Cilj ovog rada je detaljno istražiti osovne algoritme strojnog učenja i stohastičke modele, opisati njihove glavne funkcionalnosti i način rada i na izgrađenom modelima vidjeti prednosti i nedostatke pojedinog modela u primjeni problema predikcije. Napravljeni modeli su testirani na cijelom skupu podataka (10 godina) sunčane elektrane Trina 10,5 kW na lokaciji Australia Springs i na manjem skupu podataka od tjedan dana. Na temelju analize dobivenih rezultata, utvrđene su prednosti i nedostatci svakog od modela, robusnost na veće promjene u podacima, osjetljivost na ulazne podatke i sposobnost praćenja trenda. Nakon same evaluacije modela zaključeno je da je model neuronske mreže puno bolji za rješavanje problema predikcije izlazne snage SE dan unaprijed jer bolje prati nagle promjene vrijednosti, ima dugoročno pamćenje, ali dobro su uočena kretanja vrijednosti i na manjem skupu od tjedan dana. Za razliku od NM, ARIMA model je pokazao puno lošije rezultate na velikom skupu podataka, a također rezultati ARIMA modela su dosta lošiji kada se smanji skup podataka na tjedan dana što dokazuje da teže prati nagle promjene u vrijednosti i uočava pravilnosti u kretanju istih. Na temelju usporedbe predviđenih i stvarnih vrijednosti izlazne snage, dobivene pogreške, kao i kvalitete modela i njegove prikladnosti za odabrani skup podataka, model NM je dokazano bolji odabir u predikciji izlazne snage sunčane elektrane. |
| Abstract (english) | In today's world, power systems are becoming increasingly complex due to the integration of renewable energy sources and the growing challenges in meeting the demands for stability, reliability, and sustainable electricity production. Given these challenges, modern machine learning methods and stochastic models are increasingly used to predict the output of individual units within the system. The aim of this paper is to thoroughly explore the basic machine learning algorithms and stochastic models, describe their main functionalities and operating principles, and evaluate the advantages and disadvantages of each model when applied to the problem of prediction. The developed models were tested on the entire dataset (10 years) of the Trina 10.5 kW solar power plant located in Australia Springs and on a smaller dataset spanning one week. Based on the analysis of the obtained results, the strengths and weaknesses of each model were determined, including their robustness to significant data changes, sensitivity to input data, and their ability to follow trends. After evaluating the models, it was concluded that the neural network model is far superior for solving the day-ahead solar power output prediction problem, as it better captures sudden value changes, has long-term memory, and effectively tracks value trends even in the smaller one-week dataset. In contrast to the neural network, the ARIMA model showed significantly poorer results on the larger dataset, and its performance was even worse when the dataset was reduced to one week, proving that it struggles to track sudden changes in values and to recognize patterns in value trends. Based on the comparison of predicted and actual output power values, the errors obtained, as well as the quality of the model and its suitability for the selected dataset, it has been proven that the neural network model is the better choice for predicting the output power of the solar power plant. |
