TY - GEN
T1 - A Novel Bayesian Deep Learning Approach to the Downscaling of Wind Speed with Uncertainty Quantification
AU - Gerges, Firas
AU - Boufadel, Michel C.
AU - Bou-Zeid, Elie
AU - Nassif, Hani
AU - Wang, Jason T.L.
N1 - Publisher Copyright:
© 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - Wind plays a crucial part during adverse events, such as storms and wildfires, and is a widely leveraged source of renewable energy. Predicting long-term daily local wind speed is critical for effective monitoring and mitigation of climate change, as well as to locate suitable locations for wind farms. Long-term simulations of wind dynamics (until year 2100) are given by various general circulation models (GCMs). However, GCM simulations are at a grid with coarse spatial resolution (>100 km), which renders spatial downscaling to a smaller scale an important prerequisite for climate-impacts studies. In this work, we propose a novel deep learning approach, named Bayesian AIG-Transformer, that consists of an attention-based input grouping (AIG), transformer, and uncertainty quantification. We use the proposed approach for the spatial downscaling of daily average wind speed (AWND), formulated as a multivariate time series forecasting problem, over four locations within New Jersey and Pennsylvania. To calibrate and evaluate our deep learning approach, we use large-scale observations extracted from NOAA’s NCEP/NCAR reanalysis dataset (2.5° × 2.5° resolution), which provides a proxy for GCM data when evaluating the model. Results show that our approach is suitable for the downscaling task, outperforming related machine learning methods.
AB - Wind plays a crucial part during adverse events, such as storms and wildfires, and is a widely leveraged source of renewable energy. Predicting long-term daily local wind speed is critical for effective monitoring and mitigation of climate change, as well as to locate suitable locations for wind farms. Long-term simulations of wind dynamics (until year 2100) are given by various general circulation models (GCMs). However, GCM simulations are at a grid with coarse spatial resolution (>100 km), which renders spatial downscaling to a smaller scale an important prerequisite for climate-impacts studies. In this work, we propose a novel deep learning approach, named Bayesian AIG-Transformer, that consists of an attention-based input grouping (AIG), transformer, and uncertainty quantification. We use the proposed approach for the spatial downscaling of daily average wind speed (AWND), formulated as a multivariate time series forecasting problem, over four locations within New Jersey and Pennsylvania. To calibrate and evaluate our deep learning approach, we use large-scale observations extracted from NOAA’s NCEP/NCAR reanalysis dataset (2.5° × 2.5° resolution), which provides a proxy for GCM data when evaluating the model. Results show that our approach is suitable for the downscaling task, outperforming related machine learning methods.
KW - Climate change
KW - Deep learning
KW - Time series forecasting
KW - Transformer
KW - Wind speed
UR - http://www.scopus.com/inward/record.url?scp=85130214523&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85130214523&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-05981-0_5
DO - 10.1007/978-3-031-05981-0_5
M3 - Conference contribution
AN - SCOPUS:85130214523
SN - 9783031059803
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 55
EP - 66
BT - Advances in Knowledge Discovery and Data Mining - 26th Pacific-Asia Conference, PAKDD 2022, Proceedings
A2 - Gama, João
A2 - Li, Tianrui
A2 - Yu, Yang
A2 - Chen, Enhong
A2 - Zheng, Yu
A2 - Teng, Fei
PB - Springer Science and Business Media Deutschland GmbH
T2 - 26th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2022
Y2 - 16 May 2022 through 19 May 2022
ER -