Sequence-to-Sequence Models for Pollutant Load Forecasting
Air pollution is still a major problem in cities all over the world, therefore a short term forecast of pollutant loads across critical regions is of interest for decision makers to take appropriate measures. In this thesis, a sequence-to-sequence recurrent neural network has been implementented and trained to perform a 48h pollutant load forecast for measuring stations across North-Rine-Westfalia on the basis of measured pollutant load and weather data. The network is based on Gated Recurrent Units (GRU) and extended with an attention mechanism. The network outperforms a baseline neural network and the attention mechanism provides interpretability of the results. The thesis is developed in cooperation with the Fraunhofer Heinrich-Hertz-Institute as part of project SAUBER.