A NOVEL TRANSFORMER-BASED MODEL FOR SPATIOTEMPORAL PM2.5 PREDICTION

Author

Jordan C. Limperis, Georgia Southern UniversityFollow

Term of Award

Fall 2022

Degree Name

Master of Science, Computer Science (M.S.C.S.)

Document Type and Release Option

Thesis (open access)

Copyright Statement / License for Reuse

This work is licensed under a Creative Commons Attribution 4.0 License.

Department

Department of Computer Sciences

Committee Chair

Weitian Tong

Committee Member 1

Lixin Li

Committee Member 2

Felix Hamza-Lup

Abstract

Forecasting time series data is a big challenge due to the temporal and multivariate dependencies in the data combined with complex geospatial relations. In this project, we present a new approach named as TSPPM25 (Transformer-based Spatiotemporal Prediction of PM2.5) for forecasting PM2.5, a key air quality indicator. It is based on the state-of-the-art Transformer neural network and various data embedding techniques. By performing attention calculations among features over time steps as well across geospatial locations, TSPPM25 mimics cognitive attention and selectively enhances essential parts of the input data while diminishing other parts. TSPPM25 can effectively capture temporal and spatial relations in tandem with multiple influencing meteorological features as well as satellite-derived Aerosol Optical Depth. Experiments demonstrate its effectiveness by comparing it with a cutting-edge ensemble deep learning model from Zhang et al. [Information Sciences 544:427–445 (2021)]. Our TSPPM25 model outperforms Zhang et al.’s model under the same experimental setting on a well-researched benchmark dataset. As Zhang et al.’s model is restricted to univariate PM2.5 prediction, our TSPPM25 model bypasses this restriction and further improves the prediction accuracy when considering more influencing meteorological features. Moreover, our TSPPM25 model can maintain high prediction accuracy over longer periods of time as compared to the Long-Short Term Memory (LSTM) and Bidirectional LSTM models (BiLSTM). Our TSPPM25 model’s spatial learning capabilities are evaluated on a novel robust California PM2.5 dataset including Aerosol Optical Depth. Our model effectively predicts PM2.5 across a multitude of stations and is better at predicting far outlier data points than the LSTM, BiLSTM, and Linear Regression models. Our TSPPM25 model is also more stable than LSTM and BiLSTM models by maintaining higher accuracy on more complex periodic PM2.5 as determined by Fourier pattern analysis.

Recommended Citation

Limperis, Jordan C., "A NOVEL TRANSFORMER-BASED MODEL FOR SPATIOTEMPORAL PM2.5 PREDICTION" (2022). Electronic Theses and Dissertations. 2503.
https://digitalcommons.georgiasouthern.edu/etd/2503

Research Data and Supplementary Material

Yes