Postprocessing visibility ensemble forecasts using Neural Networks
| dc.contributor.advisor | Baran, Sándor | |
| dc.contributor.author | Tisza, Gréta | |
| dc.contributor.department | DE--Informatikai Kar | |
| dc.date.accessioned | 2023-04-25T13:16:35Z | |
| dc.date.available | 2023-04-25T13:16:35Z | |
| dc.date.created | 2023-04-20 | |
| dc.description.abstract | Even the most advanced state-of-the-art numerical weather forecasting models suffer from uncertainty coming from various sources. Ensemble forecasts can approximate uncertainty and give more reliable results; however, even this output of different runs of the model needs post-processing to approximate the best fitting probabilistic distribution. Several statistical post-processing methods were proposed in the last two decades including both parametric and non-parametric approaches. Recenly, machine learning (ML) based techniques gain more and more popularity in this field. Neural networks were used in the regression phase of some popular statistics-based methods (like NGR or quantile-based solutions) for calculating coefficients and parameters of the statistical model, some of them showed significant improvement. However, only very few attempts were made to apply ML models to a higher abstraction level and step towards end-to-end ML-based post-processing. In this paper we try to exploit the pattern-recognition ability of Convolutional NN-s (CNN) and give a deep learning based solution which directly infers the predicting distribution from the ensemble data and previous observations. The model was trained and tested on 51-member visibility ensemble forecasts of the European Centre of Medium-Range Weather Forecasts, which weather quantity is crucial in many parts of aviation or navigation and predicting it more accurately can result in direct economic benefits. Visibility is usually reported in discrete categories (ranges), therefore the prediction was handled as a classification problem. Our model showed significant improvement compared to the raw ensemble and even to the classical multilayer perceptron neural network. | |
| dc.description.corrector | N.I. | |
| dc.description.course | Programtervező Informatikus | |
| dc.description.degree | MSc/MA | |
| dc.format.extent | 45 | |
| dc.identifier.uri | https://hdl.handle.net/2437/350817 | |
| dc.language.iso | en | |
| dc.rights.access | Hozzáférhető a 2022 decemberi felsőoktatási törvénymódosítás értelmében. | |
| dc.subject | ensemble postprocessing | |
| dc.subject | neural networks | |
| dc.subject.dspace | DEENK Témalista::Informatika::Alkalmazott matematika | |
| dc.title | Postprocessing visibility ensemble forecasts using Neural Networks |