Show simple item record

dc.rights.licenseCC-BY-NC-ND
dc.contributorSjoerd Dirksen, Kirien Whan, Maurice Schmeits
dc.contributor.advisorDirksen, S
dc.contributor.authorTeixeira Soares Tolomei, Daniel
dc.date.accessioned2022-04-20T00:00:36Z
dc.date.available2022-04-20T00:00:36Z
dc.date.issued2022
dc.identifier.urihttps://studenttheses.uu.nl/handle/20.500.12932/41500
dc.description.abstractWeather forecasts issued by Numerical Weather Prediction (NWP) systems often display systematic bias and do not quantify the inherent uncertainty of the forecast. It is the task of statistical postprocessing to use these NWP predictions to issue probabilistic forecasts that address these issues. In this work we focus on multivariate postprocessing, which also requires statistical modelling of the spatial, temporal or inter-variable dependencies. More especifically, we use NWP forecasts from the Harmonie-Arome model to issue multivariate probabilistic forecasts for hourly wind speed predictions from initialization at 0 UTC up to 48h ahead. We propose a new statistical model for multivariate forecasting, the ARMOS(p) model, that exploits the autoregressive property of forecast errors to estimate an explicit parametric distribution, and compare it to a benchmark obtained from a combination of Ensemble Output Statistics (EMOS) with the Schaake Shuffle. We further extend these models by estimating the distribution parameters using neural networks, which incorporate spatial and temporal information from the NWP forecasts by using LSTM and Convolutional layers. In our experiments we verify model performance by computing proper multivariate scores and by performing marginal verification on the test set. The results show that the LSTM/EMOSnet and the ARMOS(2)net improve on the benchmarks, and are the best models overall.
dc.description.sponsorshipUtrecht University
dc.language.isoEN
dc.subjectWe perform multivariate statistical postprocessing of the wind speed forecasts of the Harmonie-Harome numerical weather prediction model for multiple lead times (0 - 48h ahead) using new statistical and machine learning methods. We introduce the new ARMOS statistical model, that issues explicit multivariate forecast distributions, and compare it to benchmarks by estimating the parameters of the distributions linearly and using neural networks.
dc.titleMultivariate Postprocessing of Temporal Dependencies with Autoregressive and LSTM Neural Networks
dc.type.contentMaster Thesis
dc.rights.accessrightsOpen Access
dc.subject.keywordsPostprocessing; Wind speeds; Neural Networks; LSTM; Autoregressive; TIme Series; Multivariate forecast; Probabilistic forecasting
dc.subject.courseuuMathematical Sciences
dc.thesis.id3447


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record