Napovedovanje koncentracij onesnaževalcev zraka in prepoznavanje izvornih regij

Ropret, Matevž (author), Štrumbelj, Erik (mentor)

Onesnaževalci zraka lahko predstavljajo velik problem za zdravje ljudi. Onesnaževalci se lahko prenašajo z gibanjem zračnih mas iz izvorne v druge regije. Zanima nas, ali lahko s pomočjo gibanja zračnih mas napovemo, kakšna bo koncentracija onesnaženosti za nek dan in ali lahko ugotovimo, od kod prihaja onesnaženost. Obstaja že veliko literature na to temo, razvitih pa je bilo tudi nekaj metod za reševanje tega problema. Mi smo želeli uporabiti strojno učenje, da bi naredili nove, boljše metode. Naredili smo dve novi metodi. Prva temelji na principu 2D mreže, druga pa neposredno uporabi kar koordinate o trajektorijah gibanja zračnih mas. Ti dve metodi smo primerjali z obstoječima metodama CF in RCF. Končni rezultati so pokazali, da so nekatere naše metode vsaj dvakrat boljše pri napovedovanju onesnaženosti. Vseeno končni rezultati niso tako dobri, kot smo pričakovali. Na vizualizacije, od kod prihaja onesnaženost, se ne moramo preveč zanesti, saj nam je uspelo vizualizirati samo metodo 2D mreže, ki pa ne daje boljših rezultatov od obstoječih. Kombinirali smo tudi rezultate večih postaj v upanju, da bi to izboljšalo vizualizacijo, ampak tudi ta pristop ni dal boljših rezultatov.

• Language: English
• Keywords: strojno učenje, vir onesnaženosti, CF, RCF, napovedovanje onesnaženosti, naključni gozdovi, blasso, bayesovska regresija
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2017
• PID: 20.500.12556/RUL-88907
• Publication date in RUL: 30.01.2017

Secondary language

• Language: Slovenian
• Title: Forecasting air pollutant concentrations and identifying source regions

Abstract

Air pollutants are hazardous to human health. Pollutants can be transported by air masses from one region to another. We were interested if we could use air mass movement to predict daily pollution concentrations and to visualize where this pollution came from. This area is rich in related work and there already exist methods that solve this problem. Our goal was to use machine learning to create new and better performing methods. We created two new methods. The first is based on a 2D grid, while the second is based on raw coordinate data. We compared these two methods with existing CF and RCF methods. Results show that some of our methods perform more than twice as good as existing methods. However, the results are still below our initial expectations. We cannot rely on source attribution visualization, because we were able to get it working only with the 2D grid method, which is not much better than existing CF and RCF methods. We also tried combing results of multiple stations in hopes that we could make better source attribution visualization, but this also performed worse than expected.

• Keywords: machine learning, pollution source, CF, RCF, source attribution, pollutant forecasting, pollutant prediction, random forest, blasso, bayesian regression