Author ORCID Identifier
0000-0003-0640-9782
Document Type
Dissertation
Date of Award
12-31-2022
Degree Name
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Department
Computer Science
First Advisor
Jason T. L. Wang
Second Advisor
Michel Boufadel
Third Advisor
Fadi P. Deek
Fourth Advisor
Ioannis Koutis
Fifth Advisor
Jing Li
Sixth Advisor
Elie Bou-Zeid
Abstract
The impact of climate change on the environment has become increasingly visible today. One important topic in climate change monitoring is to forecast future climate events, which relies on long-term prediction of climate variables at a local regional scale. This topic is crucial for disaster risk management and for resolving climate change at the local level. General Circulation Models (GCMs) allow for the simulation of several climate variables, decades into the future. GCM simulations, however, are at a global large scale and are too coarse to monitor climate change at the local small scale. This dissertation explores new machine learning methods for performing the spatial downscaling of global large-scale climate variables. In the dissertation, three downscaling tasks are tackled: downscaling of temperatures, wind speed, and precipitation. The dissertation also explores the use of downscaled GCM simulations for the long-term prediction of solar irradiance, an important variable for solar energy generation.
First, a novel deep learning approach, named AIG-Transformer, is presented for the spatial downscaling of minimum and maximum temperatures. The core of this approach is to cluster similar input features and perform implicit feature reduction using a novel attention-based input grouping (AIG) deep neural network, before feeding the data to a stacked transformer network. AIG-Transformer is further used with future simulation data from GCMs to produce temperature projections over the 20-year period between 2030 and 2049.
Second, uncertainty quantification is incorporated into the AIG-Transformer framework to construct a Bayesian deep learning model, named Bayesian AIG-Transformer. The essence of this approach is to adopt the Monte Carlo dropout sampling technique to quantify the inherent uncertainty in the input data, depicted by the aleatoric uncertainty, as well as the uncertainty of the model, depicted by the epistemic uncertainty. The Bayesian AIG-Transformer model is used for the task of spatial downscaling of the daily average wind speed. The model is further extended to adopt multi-head convolutional neural networks (CNNs) and to perform future projections.
Third, a novel convolution-based deep learning approach is proposed for forecasting daily precipitation via downscaling. This approach, named Precipitation CNN or PCNN, employs multi-head CNNs followed by Bandanau attention blocks and an uncertainty quantification component with Bayesian inference. PCNN is further used to produce local-scale precipitation projections for multiple periods into the future (up to year 2100).
Finally, a Bayesian multi-head deep learning model, named DeepSl (denoting Deep Solar Irradiance), is proposed for the long-term forecasting of solar irradiance using downscaled GCM simulations. DeepSl consists of BLSTM autoencoders, prefixed to a transformer, with a Bayesian inference component. DeepSl is used to project daily solar irradiance up to year 2099.
Experimental results based on different datasets demonstrate the superiority of the proposed methods over existing machine learning algorithms and the feasibility of the proposed methods for climate change monitoring.
Recommended Citation
Gerges, Firas, "Monitoring climate change with machine learning and uncertainty quantification" (2022). Dissertations. 1795.
https://digitalcommons.njit.edu/dissertations/1795