Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural Networks

Simon Veldkamp; Kirien Whan; Sjoerd Dirksen; Maurice Schmeits

doi:10.1175/MWR-D-20-0219.1

Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural Networks

Simon Veldkamp, Kirien Whan, Sjoerd Dirksen, Maurice Schmeits

Royal Netherlands Meteorological Institute

Research output: Contribution to journal › Article › Academic › peer-review

Abstract

Current statistical postprocessing methods for probabilistic weather forecasting are not capable of using full spatial patterns from the numerical weather prediction (NWP) model. In this paper, we incorporate spatial wind speed information by using convolutional neural networks (CNNs) and obtain probabilistic wind speed forecasts in the Netherlands for 48 h ahead, based on KNMI's deterministic HARMONIE-AROME NWP model. The probabilistic forecasts from the CNNs are shown to have higher Brier skill scores for medium to higher wind speeds, as well as a better continuous ranked probability score (CRPS) and logarithmic score, than the forecasts from fully connected neural networks and quantile regression forests. As a secondary result, we have compared the CNNs using three different density estimation methods [quantized softmax (QS), kernel mixture networks, and fitting a truncated normal distribution], and found the probabilistic forecasts based on the QS method to be best.

Original language	English
Pages (from-to)	1141-1152
Number of pages	12
Journal	Monthly Weather Review
Volume	149
Issue number	4
DOIs	https://doi.org/10.1175/MWR-D-20-0219.1
Publication status	Published - 1 Apr 2021

Bibliographical note

Funding Information:
We thank the following people from KNMI: Toon Moene for executing the reforecasting runs of the HARMONIE-AROME model, and Andrea Pagani and Dirk Wolters for assisting with the practical implementation of the deep learning methods used in this paper. Besides, we are grateful to the 3 anonymous reviewers for their comments, which have helped to improve a previous version of the manuscript. S.D. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under SPP 1798 (COSIP - Compressed Sensing in Information Processing).

Funding Information:
Acknowledgments. We thank the following people from KNMI: Toon Moene for executing the reforecasting runs of the HARMONIE-AROME model, and Andrea Pagani and Dirk Wolters for assisting with the practical implementation of the deep learning methods used in this paper. Besides, we are grateful to the 3 anonymous reviewers for their comments, which have helped to improve a previous version of the manuscript. S.D. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under SPP 1798 (COSIP - Compressed Sensing in Information Processing).

Publisher Copyright:
© 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Keywords

Deep learning
Forecast verification/skill
Machine learning
Model output statistics
Neural networks
Probability forecasts/models/distribution

Access to Document

10.1175/MWR-D-20-0219.1

[15200493 - Monthly Weather Review] Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural NetworksFinal published version, 866 KB

Cite this

@article{600dca6dab194d298d9a01a0b9970461,

title = "Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural Networks",

abstract = "Current statistical postprocessing methods for probabilistic weather forecasting are not capable of using full spatial patterns from the numerical weather prediction (NWP) model. In this paper, we incorporate spatial wind speed information by using convolutional neural networks (CNNs) and obtain probabilistic wind speed forecasts in the Netherlands for 48 h ahead, based on KNMI's deterministic HARMONIE-AROME NWP model. The probabilistic forecasts from the CNNs are shown to have higher Brier skill scores for medium to higher wind speeds, as well as a better continuous ranked probability score (CRPS) and logarithmic score, than the forecasts from fully connected neural networks and quantile regression forests. As a secondary result, we have compared the CNNs using three different density estimation methods [quantized softmax (QS), kernel mixture networks, and fitting a truncated normal distribution], and found the probabilistic forecasts based on the QS method to be best.",

keywords = "Deep learning, Forecast verification/skill, Machine learning, Model output statistics, Neural networks, Probability forecasts/models/distribution",

author = "Simon Veldkamp and Kirien Whan and Sjoerd Dirksen and Maurice Schmeits",

note = "Funding Information: We thank the following people from KNMI: Toon Moene for executing the reforecasting runs of the HARMONIE-AROME model, and Andrea Pagani and Dirk Wolters for assisting with the practical implementation of the deep learning methods used in this paper. Besides, we are grateful to the 3 anonymous reviewers for their comments, which have helped to improve a previous version of the manuscript. S.D. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under SPP 1798 (COSIP - Compressed Sensing in Information Processing). Funding Information: Acknowledgments. We thank the following people from KNMI: Toon Moene for executing the reforecasting runs of the HARMONIE-AROME model, and Andrea Pagani and Dirk Wolters for assisting with the practical implementation of the deep learning methods used in this paper. Besides, we are grateful to the 3 anonymous reviewers for their comments, which have helped to improve a previous version of the manuscript. S.D. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under SPP 1798 (COSIP - Compressed Sensing in Information Processing). Publisher Copyright: {\textcopyright} 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).",

year = "2021",

month = apr,

day = "1",

doi = "10.1175/MWR-D-20-0219.1",

language = "English",

volume = "149",

pages = "1141--1152",

journal = "Monthly Weather Review",

issn = "0027-0644",

publisher = "American Meteorological Society",

number = "4",

}

TY - JOUR

T1 - Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural Networks

AU - Veldkamp, Simon

AU - Whan, Kirien

AU - Dirksen, Sjoerd

AU - Schmeits, Maurice

N1 - Funding Information: We thank the following people from KNMI: Toon Moene for executing the reforecasting runs of the HARMONIE-AROME model, and Andrea Pagani and Dirk Wolters for assisting with the practical implementation of the deep learning methods used in this paper. Besides, we are grateful to the 3 anonymous reviewers for their comments, which have helped to improve a previous version of the manuscript. S.D. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under SPP 1798 (COSIP - Compressed Sensing in Information Processing). Funding Information: Acknowledgments. We thank the following people from KNMI: Toon Moene for executing the reforecasting runs of the HARMONIE-AROME model, and Andrea Pagani and Dirk Wolters for assisting with the practical implementation of the deep learning methods used in this paper. Besides, we are grateful to the 3 anonymous reviewers for their comments, which have helped to improve a previous version of the manuscript. S.D. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under SPP 1798 (COSIP - Compressed Sensing in Information Processing). Publisher Copyright: © 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

PY - 2021/4/1

Y1 - 2021/4/1

N2 - Current statistical postprocessing methods for probabilistic weather forecasting are not capable of using full spatial patterns from the numerical weather prediction (NWP) model. In this paper, we incorporate spatial wind speed information by using convolutional neural networks (CNNs) and obtain probabilistic wind speed forecasts in the Netherlands for 48 h ahead, based on KNMI's deterministic HARMONIE-AROME NWP model. The probabilistic forecasts from the CNNs are shown to have higher Brier skill scores for medium to higher wind speeds, as well as a better continuous ranked probability score (CRPS) and logarithmic score, than the forecasts from fully connected neural networks and quantile regression forests. As a secondary result, we have compared the CNNs using three different density estimation methods [quantized softmax (QS), kernel mixture networks, and fitting a truncated normal distribution], and found the probabilistic forecasts based on the QS method to be best.

AB - Current statistical postprocessing methods for probabilistic weather forecasting are not capable of using full spatial patterns from the numerical weather prediction (NWP) model. In this paper, we incorporate spatial wind speed information by using convolutional neural networks (CNNs) and obtain probabilistic wind speed forecasts in the Netherlands for 48 h ahead, based on KNMI's deterministic HARMONIE-AROME NWP model. The probabilistic forecasts from the CNNs are shown to have higher Brier skill scores for medium to higher wind speeds, as well as a better continuous ranked probability score (CRPS) and logarithmic score, than the forecasts from fully connected neural networks and quantile regression forests. As a secondary result, we have compared the CNNs using three different density estimation methods [quantized softmax (QS), kernel mixture networks, and fitting a truncated normal distribution], and found the probabilistic forecasts based on the QS method to be best.

KW - Deep learning

KW - Forecast verification/skill

KW - Machine learning

KW - Model output statistics

KW - Neural networks

KW - Probability forecasts/models/distribution

UR - http://www.scopus.com/inward/record.url?scp=85102667607&partnerID=8YFLogxK

U2 - 10.1175/MWR-D-20-0219.1

DO - 10.1175/MWR-D-20-0219.1

M3 - Article

SN - 0027-0644

VL - 149

SP - 1141

EP - 1152

JO - Monthly Weather Review

JF - Monthly Weather Review

IS - 4

ER -

Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural Networks

Abstract

Bibliographical note

Keywords

Access to Document

Other files and links

Fingerprint

Cite this