Data stream fusion for predicting electricity price

Konda, Jaka (author), Bosnić, Zoran (mentor), Legenstein, Robert (comentor)

In this work, we tackle the problem of building a predictive model for the electricity market. In it, market players compete for the best prices to increase their profits and now with transition to renewable energy sources, the market has become more volatile and dependant on different environmental factors. In our experiments we use different statistical and machine learning methods in a combination with data sources from European energy platform and European meteorological institute to obtain weather information. We fuse three different data sources together into four different datasets. We follow the machine learning pipeline where we select features, model hyper-parameters and test the models on the test set. Contrary to our expectations, additional weather information had a negative impact on the error and the variance of the models. Some improvements in the prediction accuracy were noted only when we included additional datasets with most important selected features being the lagged values of the target variable and also past and forecasted energy load. Our best method is obtained using late level fusion by combining all trained regressors together into an ensemble. It achieves an sMAPE error of 13.084, compared to our second best, neural network, with an error of 14.932 and baseline with the value of 22.963.

• Language: English
• Keywords: machine learning, time-series forecasting, incremental learning, data fusion, data streaming, electricity price prediction
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2019
• PID: 20.500.12556/RUL-113304
• COBISS.SI-ID: 1538499779
• Publication date in RUL: 19.12.2019

Secondary language

• Language: Slovenian
• Title: Zlivanje podatkovnih tokov pri napovedovanju cene električne energije

Abstract

V delu se lotimo problema napovedovanja cen električne energije, kjer na trgu udeleženci tekmujejo za čim večji profit. V zadnjem obdobju s prehodom na obnovljive vire energije je trg postal bolj nepredvidljiv in odvisen od okoljskih dejavnikov. V naših eksperimentih uporabimo različne statistične modele in modele strojnega učenja v kombinaciji s podatkovno zbirko Evropske mreže operaterjev električnega omrežja in Evropskega meteorološkega inštituta. Vse zbirke združimo v štiri različne podatkovne množice in sledimo cevovodu strojnega učenja. Najprej izberemo atribute, nato hiper-parametre modelov in nazadnje ovrednotimo modele na testni množici. V nasprotju s pričakovanji, dodatne vremenske informacije niso pripomogle k izboljšanju uspeha modelov ter so imele celo negativen vpliv na napovedno uspešnost. Rezultati smo uspeli izboljšati, ko smo uvedli dodatne atribute iz Evropske mreže operaterjev, kjer so se za najpomembnejše izkazale zakasnjene vrednosti ciljne spremenljivke in pretekle ter napovedane porabe električne energije. Najboljši rezultat smo dosegli z metodo ansamblov z uporabo poznega zlivanja. Ansambel doseže napako sMAPE 13,084, drugi najboljši model, nevronska mreža, 14,392, referenčni model pa napako 22,963.

• Keywords: strojno učenje, napovedovanje časovnih vrst, inkrementalno učenje, zlivanje podatkov, podatkovni takovi, napovedovanje cene električne energije