南京大学学报（自然科学版）2026，Vol.62Issue(3)：442-450,9.DOI:10.13232/j.cnki.jnju.2026.03.010
基于降噪编码与时序分块机制的PM2.5浓度预测模型DAE-PatchTST
A PM2.5 concentration prediction model DAE-PatchTST based on denoising autoencoder and temporal patch segmentation
摘要
Abstract
To address the issues of low efficiency and insufficient accuracy in multivariate long-sequence forecasting of PM2.5 concentration time series data,a DAE-PatchTST-based PM2.5 concentration prediction model is proposed.The data were collected from seven meteorological observation stations and atmospheric composition stations in different regions of Chongqing.Six sets of hourly historical meteorological and environmental data were selected using Pearson correlation coefficient analysis,and linear interpolation and reversible instance normalization methods were applied for data preprocessing.The constructed model utilizes a denoising autoencoder mechanism to reduce noise in the original input,thereby ensuring model stability.Meanwhile,channel independence and time series patch mechanisms were introduced into the classical Transformer architecture to extensively capture long-term dependencies in the input sequences.Comparative experiments with existing research methods such as RNN,GRU and LSTM demonstrate that the proposed model achieves the best performance in MAE,RMSE and R² metrics for short-and medium-term PM2.5 concentration time series forecasting tasks.Additionally,the model exhibits strong reliability in predicting PM2.5 concentrations in surrounding regions.关键词
时间序列/PM2.5浓度预测/降噪自编码/DAE-PatchTSTKey words
time series/PM2.5 concentration prediction/denosing auto encoder/DAE-PatchTST分类
资源环境引用本文复制引用
李文钊,董晓炜,王新,赵芳..基于降噪编码与时序分块机制的PM2.5浓度预测模型DAE-PatchTST[J].南京大学学报（自然科学版）,2026,62(3):442-450,9.基金项目
重庆市三峡库区危岩地灾防治气象保障项目(TC249D043),中国气象局智能观测重点实验室开放基金重点项目(ZNGC2025ZD05) （TC249D043）
