PNC-based trend feature extraction method for time series data

Bo He

doi:10.1016/j.neucom.2024.129174

PNC-based trend feature extraction method for time series data

2025-02-01

0

PRE

AI

DOI：10.1016/j.neucom.2024.129174

原文链接

原文求助

摘要

En 中文

With the rapid development of manufacturing industry, air pollution has become a global hot issue, and the accurate prediction of PM2.5 has been an important topic, which can provide valuable information for governmental decision-making in environmental management affairs, among which the trend feature information is quite important. In this paper, in order to solve the problem of insufficient trend feature extraction for time series data, a trend feature extraction method based on PNC (Positive-Negative Correlation) for time series data is proposed. The method focuses on the trend feature extraction of time series data, and skillfully adopts the positive-negative constraints and positive-inverse ratio constraints of PNC to extract trend features of time series data. Firstly, the missing values of the features are filled in the time series data by using the backward and forward sequence pattern, and secondly, the regression algorithm is applied to optimally fill in the missing values of the target columns; then the EMD algorithm is used to extract more features of the data with different frequencies as the new features; then the original unsupervised learning data is transformed into the supervised learning data; and finally, the PNC method is used to extract the trend features of the time series data. In this paper, four deep learning models are used in the experiments to train the prediction of PM2.5 experimental data in four cities. The experimental results show that the RMSE, MSE, MAE, and R2-Score metrics of the model set (Model-Ours) using the PNC-based trend feature extraction method for time-series data are better than those of the model set without trend feature extraction (Model-Base) in the four datasets; among which the MSE metrics are the most obvious, and Model-Ours are better than those of the model set with no trend feature extraction in each of the four datasets, with prediction steps of 48,96,192,384,768 are lower than Model-Base by 67.882, 29.152, 19.559, 9.57 on average.This proves that the trend feature extraction method of time series data based on PNC can improve the prediction accuracy of the model and reduce the model error, and it also proves the effectiveness and superiority of the method.

Keyword:

PM2.5 concentration prediction

Time series trend features extraction

Correlation analysis

Neural network

PNC-based trend feature extraction method for time series data

摘要

期刊

机构

引用论文