Deep learning and statistical approaches for area-based PM2.5 forecasting in Hat Yai, Thailand

PM2.5 pollution poses a significant environmental and health concern across Southeast Asia, including Thailand. This study aims to forecast area-based PM2.5 concentrations in Hat Yai city, Songkhla province, using daily data collected from the Hat Yai monitoring station from 2013 to 2023. To achieve...

Full description

Saved in:
Bibliographic Details
Published in:Journal of big data Vol. 12; no. 1; pp. 36 - 24
Main Authors: Damkliang, Kasikrit, Chumnaul, Jularat
Format: Journal Article
Language:English
Published: Cham Springer International Publishing 01.12.2025
Springer Nature B.V
SpringerOpen
Subjects:
Area-based PM2.5 forecasting
Averages
Big Data
Big Data and Artificial Intelligence in Emerging Scientific Fields
Classification
Communications Engineering
Computational Science and Engineering
Computer Science
ConvLSTM1D-BiLSTM
Data Mining and Knowledge Discovery
Database Management
Deep learning
Effectiveness
Forecasting
Information Storage and Retrieval
Machine learning
Mathematical Applications in Computer Science
Mean
Methodology
Networks
Pollution
Robustness
Sliding window configuration
Southern Thailand
Thailand
Deep learning
Sliding window configuration
Statistical analysis
Area-based PM2.5 forecasting
Machine learning
Southern Thailand
Time-series analysis
ConvLSTM1D-BiLSTM
ISSN:2196-1115, 2196-1115
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:PM2.5 pollution poses a significant environmental and health concern across Southeast Asia, including Thailand. This study aims to forecast area-based PM2.5 concentrations in Hat Yai city, Songkhla province, using daily data collected from the Hat Yai monitoring station from 2013 to 2023. To achieve this, we developed and evaluated forecasting models utilizing both Deep Learning (DL) and Machine Learning (ML) techniques, with performance assessed using statistical tools. Among the tested models, the ConvLSTM1D-BiLSTM model, optimized with a 7-input window, 1-output window, and 6 strides, demonstrated the highest effectiveness, achieving a mean MAE of 2.05 and a mean R 2 score of 0.68. In PM2.5 level classification, this model attained macro-average accuracy, sensitivity, and F1 scores of 0.86, 0.80, and 0.80, respectively. External validation using data from four nearby stations further confirmed the model’s effectiveness, yielding a mean MAE of 1.38 and a mean R 2 score of 0.90. These results underscore the robustness of our approach, supporting its practical application, particularly for local stations in Southern Thailand.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2196-1115
2196-1115
DOI:10.1186/s40537-025-01079-9