Napovedovanje delcev PM10 v Slovenji s pomočjo nevronskih mrež

MARINČEK, MARK (author), Faganeli Pucer, Jana (mentor)

V diplomski nalogi smo se posvetili problematiki onesnaženosti zraka z delci PM10 in njihovemu vplivu na zdravje. Osredotočili smo se na primerjavo različnih modelov globokih nevronskih mrež (LSTM, CNN, MLP) in ansambelskega modela naključnih gozdov (RF) pri napovedovanju dnevne koncentracije delcev PM10. Uporabili smo meteorološke podatke in napovedi modela ECMWF za napovedovanje v Ljubljani, Celju, Zagorju in Kopru za današnji in jutrišnji dan. Za iskanje optimalnih hiperparametrov modelov smo uporabili metodo iskanja po mreži. Učinkovitost napovednih modelov smo ocenjevali s tremi metrikami: MAE, MAPE in RMSE. Ugotovili smo, da se je LSTM najbolje odrezal pri napovedih za današnji in še zlasti pri napovedih za jutrišnji dan. Rezultati so pokazali, da je med primerjanimi modeli LSTM najboljša izbira za napovedovanje delcev PM10, medtem ko so se drugi modeli odrezali različno glede na kraj in dan napovedi. MLP se je pričakovano odrezal najslabše, medtem ko je CNN presenetil z dobro uspešnostjo, zlasti pri napovedih za današnji dan. Diplomska naloga ponuja vpogled v učinkovitost uporabe globokih nevronskih mrež za napovedovanje onesnaženosti zraka ter prispeva k razumevanju, kako lahko različni modeli in metrike vplivajo na točnost napovedi.

• Language: Slovenian
• Keywords: globoke nevronske mreže, LSTM, CNN, MLP, RF, PM10
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-155321
• COBISS.SI-ID: 190883843
• Publication date in RUL: 26.03.2024

Secondary language

• Language: English
• Title: Forecasting PM10 particles in Slovenia using neural networks

Abstract

In this thesis, we focused on the prediction of air pollution in particularly with PM10. We concentrated on comparing various deep neural network models (LSTM, CNN, MLP) and the ensemble model random forests (RF) for predicting the daily concentration of PM10 particles. Data extracted from meteorological measurements and ECMWF model forecasts were used for PM10 predictions in Ljubljana, Celje, Zagorje, and Koper for the current and next day. To find the optimal hyperparameters of the models, we employed a grid search method. The effectiveness of the predictive models was evaluated using three metrics: MAE, MAPE, and RMSE. We found that LSTM performed best for predictions for today and was even more precise for predictions for tomorrow. The results showed that among the compared models, LSTM is the best choice for predicting PM10 concentrations, while other models varied in performance depending on the location and the day of prediction. As expected, MLP performed the worst, while CNN surprised with good efficiency, especially in today's predictions. This thesis provides insight into the effectiveness of deep neural networks and for predicting air pollution and contributes to understanding how different models and metrics can affect the accuracy of predictions.

• Keywords: deep neural network, LSTM, CNN, MLP, RF, PM10