Subject: аir pollution forecast; LSTM & Meta models; univariate vs. multivariate comparison
Year: 2019
Type: Proceedings
Title: Air Pollution Prediction Using LSTM Neural Networks
Author: Evkoski, Bojan
Author: Stojanovski, Zafir
Author: Trajkovski, Aleksandar
Author: Gjorgjevikj, Dejan
Abstract: Air pollution in North Macedonia is 20 times over the EU limit. Recently Skopje is mentioned as the most polluted city in Europe. As a result, this is believed to contribute to 2000 annual premature deaths in Skopje, Tetovo and Bitola only. Being able to forecast air pollution levels to take timely precaution could drastically reduce these numbers. Using state of the art recurrent neural networks known as LSTMs, we were able to predict these levels by combining historical pollution data and weather forecasts through meta models, achieving mean RMSE for all sensors around 20, with the best results having RMSE as low as 8.78, with PM10 measurements ranging from 0 to above 1000 and are usually accompanied by a lot of noise. In this paper we present several approaches we have tried for solving the problem and a basic comparison between them and we also propose a way to expand these models into a realtime system for multitarget predictions.
Publisher:
Relation: The 16th International Conference for Informatics and Information Technology (CIIT 2019)
Identifier: oai:repository.ukim.mk:20.500.12188/23862
Identifier: http://hdl.handle.net/20.500.12188/23862
