Loading...

Evaluating the Effectiveness of Physics-Informed Machine Learning Models in Enhancing Dust Particle Concentration Forecasting

Giahchin, Kimia | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58374 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Danesh Yazdi, Mohammad
  7. Abstract:
  8. This research aims to evaluate the capability of a physics-based machine learning model in forecasting PM₁₀ concentration (wich serves as dust particle concentration). Previous studies have primarily employed two methodological approaches for dust concentration prediction namely one based on the physical dynamics governing dust events and the other relying on data-driven techniques through machine learning. In this study, a Physics-Informed Neural Network (PINN) model was employed, integrating the strengths of both approaches. Training the PINN model required gridded meteorological inputs and corresponding gridded PM₁₀ data. To ensure reliability, correlation analyses were conducted between multiple reanalysis meteorological datasets and in-situ measurements for the period 2013–2017, and the most consistent dataset was selected. Natural dust events from 2001 to 2018 were subsequently identified using meteorological station codes and pollutant monitoring records. The neural network model was trained using meteorological variables derived from the reanalysis meteorological database, while ground-based PM₁₀ measurements for the current time and three hours ahead served as the target variables. The model achieved mean absolute errors of 59.94 μg/m³ for current-time prediction and 58.15 μg/m³ for the three-hour-ahead forecast. In the next phase, for each identified dust event, gridded meteorological inputs were generated by interpolating the selected remote sensing product, and PM₁₀ values for the current and three-hour-ahead time steps were predicted using the previously trained neural network model. The PINN model was subsequently trained using the gridded datasets for each dust event under four distinct scenarios including regional non-urban, local non-urban, regional urban, and local urban events. Across all selected events, the PINN model achieved an average MAE of 12.58 μg/m³ for current-time prediction and 14.38 μg/m³ for the three-hour-ahead forecast, thereby outperforming the neural network results. In the validation phase, comparison with ground-based PM₁₀ data yielded an average MAE of 38.42 μg/m³ for the current time step, further confirming the improved performance of the PINN over the neural network. In the final step, a scenario analysis was conducted to evaluate the performance of the PINN model against the four mentioned scenarios. Although the PINN effectively reconstructed individual dust events, generalization across events required parameter adjustments due to fundamental differences in atmospheric conditions, land-surface properties, and dust source dynamics. Nonetheless, the high degree of agreement with observations and the capacity to reproduce event-specific processes underscore the potential of the PINN framework for integration into dust early warning systems. In particular, its reduced computational demand and faster processing relative to conventional numerical models highlight its value as both an efficient alternative and a complementary enhancement to existing forecasting methods
  9. Keywords:
  10. Dust Storm ; Machine Learning ; Interpretability ; Particle Transport Physics ; Dust Concentration

 Digital Object List

 Bookmark

No TOC