Scholarlink

PM2.5 is a significant contributor to air pollution worldwide, necessitating accurate prediction methods for effective prevention and early warning. While initial studies focused on temporal forecasting, recent efforts have incorporated spatiotemporal modeling, though performance challenges remain. In this paper, we aim to enhance such techniques for predicting future PM2.5 concentrations by utilizing the latest video prediction model, “IAM4VP,” as the backbone. The model incorporates features such as PM2.5 concentration, number of hotspots on the Earth’s surface, and wind speed moving eastward and northward. To enhance temporal learning, cyclical encoding was integrated, addressing the importance of hourly PM2.5 variation. Experiments conducted on PM2.5 data from Thailand (2022–2023) demonstrated significant improvements, including a 13.63% reduction in MAE and a 15.27% increase in F1 score compared to the ConvLSTM baseline. The model showed particular strength in predicting Moderate-to-Unhealthy PM2.5 levels and achieved higher accuracy across different prediction horizons, especially during the initial hours.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing PM2.5 Forecasting Using Video-Based Spatiotemporal Models and Cyclical Encoding

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jitti Pranonsatit,
  • Kritchart Wongwailikhit,
  • Pisut Painmanakul,
  • Peerapon Vateekul

摘要

PM2.5 is a significant contributor to air pollution worldwide, necessitating accurate prediction methods for effective prevention and early warning. While initial studies focused on temporal forecasting, recent efforts have incorporated spatiotemporal modeling, though performance challenges remain. In this paper, we aim to enhance such techniques for predicting future PM2.5 concentrations by utilizing the latest video prediction model, “IAM4VP,” as the backbone. The model incorporates features such as PM2.5 concentration, number of hotspots on the Earth’s surface, and wind speed moving eastward and northward. To enhance temporal learning, cyclical encoding was integrated, addressing the importance of hourly PM2.5 variation. Experiments conducted on PM2.5 data from Thailand (2022–2023) demonstrated significant improvements, including a 13.63% reduction in MAE and a 15.27% increase in F1 score compared to the ConvLSTM baseline. The model showed particular strength in predicting Moderate-to-Unhealthy PM2.5 levels and achieved higher accuracy across different prediction horizons, especially during the initial hours.