Post-processing of air pollutant predictions using machine learning

Majkić, Dragan (author), Faganeli Pucer, Jana (mentor)

Air pollution is a critical environmental and public health problem, making accurate air pollutant forecasting important for decision-makers and the general public. Numerical air quality models such as CAMx provide operational forecasts, but their predictions can contain errors and therefore require corrections (post-processing). This thesis investigates machine learning and deep learning methods for post-processing CAMx air pollutant forecasts in Slovenia. The developed approach combines CAMx air quality forecasts, ALADIN meteorological forecasts, and observations from Slovenian monitoring stations. We compared baseline, machine learning, and deep learning methods using an expanding window walk-forward evaluation. The analysis focused on hourly PM10 and O3 concentrations and forecasting exceedances of limit values. All post-processing methods improved the original CAMx forecasts. XGBoost and the Temporal Fusion Transformer (TFT) achieved the best results, with TFT reducing RMSE by approximately 36% for hourly PM10 and 51% for hourly O3. Post-processing also improved exceedance detection. Overall, the results show that the proposed methodology can improve CAMx forecasts and provide more reliable next-day hourly predictions.

• Language: English
• Keywords: environmental data science, machine learning, time series, air pollution, post-processing
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-186066
• COBISS.SI-ID: 289220355
• Publication date in RUL: 26.08.2026

Secondary language

• Language: Slovenian
• Title: Poprocesiranje napovedi onesnaževal zraka z uporabo strojnega učenja

Abstract

Onesnaženost zraka je nujen okoljski in javnozdravstveni problem, zato so natančne napovedi koncentracij onesnaževal pomembne za odločevalce in širšo javnost. Numerični modeli kakovosti zraka, kot je CAMx, omogočajo operativno napovedovanje, vendar lahko njihove napovedi vsebujejo napake, zato je potrebno napovedi popraviti (poprocesirati). V magistrskem delu smo preučili uporabo metod strojnega in globokega učenja za poprocesiranje napovedi modela CAMx v Sloveniji. Razviti pristop uporablja napovedi modela CAMx, meteorološke napovedi modela ALADIN in meritve s slovenskih merilnih postaj. Primerjali smo preproste metode poprocesiranja in preverili, ali uporaba bolj kompleksnih modelov strojnega in globokega učenja prinese večje izboljšave. Modele smo ovrednotili po metodi razširjajočega okna. Napovedovali smo urne koncentracije PM10 in O3 ter preseganja mejnih vrednosti. Z uporabo metod poprocesiranja smo izboljšali napovedi CAMx. Najboljše rezultate sta dosegla XGBoost in Temporal Fusion Transformer (TFT). TFT je zmanjšal RMSE za približno 36% pri PM10 in 51% pri O3. Poprocesiranje je izboljšalo tudi zaznavanje preseganj mejnih vrednosti ter zagotovilo zanesljivejše urne napovedi za naslednji dan.

• Keywords: okoljska podatkovna znanost, strojno učenje, časovne vrste, onesnaževanje zraka, poprocesiranje