Loading...

Enhancing the Accuracy of Temperature and Pressure Variables in Global Weather Models Using Machine Learning and Wavelet Transformation in the Urmia Basin

Jazini, Ali | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57362 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Safaie Nematollahi, Ammar
  7. Abstract:
  8. The objective of this research is to enhance the accuracy of temperature and pressure variables derived from global meteorological models such as ERA5 using machine learning and deep learning (ML-DL) techniques. Since variables like temperature and pressure are critical metrics for various processes and decision-making in fields such as agriculture and limnology, and they serve as input variables in most hydrological, hydrodynamic, and meteorological models, accessing high-precision time series with appropriate temporal resolution is of great importance. Currently, global meteorological models like ERA5 provide relatively long time series of these variables on a grid with spatial resolutions of 0.25 and 0.1 degrees. In Iran, the primary boundary conditions for these models are derived from the data of 73 synoptic stations with international distribution. However, there are over 500 active synoptic stations in Iran. Despite the often lower accuracy of these data, they are utilized in most grid-based hydrological models, and efforts are made to manage their errors through model calibration. Nevertheless, when input data have low accuracy or observational data are scarce, model calibration cannot significantly enhance modeling quality. To increase spatial resolution, models such as the WRF (Weather Research and Forecasting) model are employed, which require substantial computational resources and time for execution. Given the importance of this issue, this study aims to improve the accuracy of historical data from the ERA5 product, which already has high temporal and spatial resolution, using ML-DL methods and wavelet transformations. Therefore, for each of the temperature and pressure variables, modeling was performed using the time series from 1990 to 2015 as the training set and the time series from 2015 to 2022 as the testing set. For the temperature variable, the best case yielded a mean absolute error (MAE) of 1.65°C and a Nash–Sutcliffe efficiency coefficient of 0.95. For the pressure variable, the best case achieved a mean absolute error of 0.35 hPa and a Nash–Sutcliffe efficiency coefficient of 0.99. Additionally, another finding of this study is that although the use of wavelet transformation reduces modeling errors at each frequency level, it does not necessarily decrease the total modeling error, which is the sum of the modeling errors at each frequency and the errors introduced by the transformation itself. This is because the transformation process inherently introduces some degree of error into the problem
  9. Keywords:
  10. Machine Learning ; Deep Learning ; Wavelet Transform ; Weather Research and Forecasting (WRF)Modeling System ; ERA5 Database ; Meteorological Variables ; Time Series Accuracy Enhancement ; Lake Urmia Watershed

 Digital Object List

 Bookmark

No TOC