LSTMs for Inferring Planetary Boundary Layer Height (PBLH)
Loading...
Links to Files
Permanent Link
Author/Creator
Author/Creator ORCID
Date
2020-01-01
Type of Work
Department
Computer Science and Electrical Engineering
Program
Computer Science
Citation of Original Publication
Rights
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan through a local library, pending author/copyright holder's permission.
This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see http://aok.lib.umbc.edu/specoll/repro.php or contact Special Collections at speccoll(at)umbc.edu
This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see http://aok.lib.umbc.edu/specoll/repro.php or contact Special Collections at speccoll(at)umbc.edu
Abstract
In this paper, we describe how machine learning could be used to augment existing methods for calculating and estimating the Planetary Boundary Layer Height (PBLH). We show the performance of using a Long Short-Term Memory (LSTM) Network to learn PBLH changes over time for different geographical locations across the United States, in conjunction with the Weather Research and Forecasting coupled with Chemistry (WRF-Chem) model. If a machine learning method could accurately emulate the model-based PBLH calculations, and there are significant efficiencies in computing time, then it is feasible to consider replacing the PBLH model with the deep learning model of the WRF-Chem model. We compared three different stacked LSTM models to understand their usefulness for forecasting PBLH based on learning from the physical model of past PBLH outputs. We conducted four main experiments to train the behavior of these three varied models and measured their performance using RMSE. We also compared their performance to regression models using the same data. The first experiment used a univariate multi-location LSTM network trained on 20 geographical locations for a two-month period of hourly PBLH calculated from WRF-Chem model. We evaluated how well this type of LSTM could learn PBLH by using a limited amount of data across a set of locations located geographically near each other. The model achieved an RMSE of 0.11. The second experiment used a univariate single location LSTM network for multiple locations, each network was trained using one year of hourly PBLH calculations from the WRF-Chem model for that specific location. We evaluated three different locations, achieving RMSE scores of 0.04, 0.01 and 0.05, respectively. The third experiment used the same single location LSTM network; however, the locations were deliberately selected from across the U.S. to understand how well the LSTM model generalizes across the gridded geographical locations, achieving an RMSE of 0.05. The fourth experiment used a multivariate LSTM network which included as features the location's latitude and longitude values, giving an RMSE of 0.2. The overall results suggest that the LSTM model proves proficient to potentially predict PBLH for a site located as far as 43 miles away from the trained location. We further present future plans for resolving some new questions arising from this work.