Jin Fan , Zehao Wang , Feiwei Qin , Huifeng Wu , Danfeng Sun , Jia Wu
{"title":"面向长序列时间序列预测的双相关分布特征提取网络","authors":"Jin Fan , Zehao Wang , Feiwei Qin , Huifeng Wu , Danfeng Sun , Jia Wu","doi":"10.1016/j.neucom.2025.130806","DOIUrl":null,"url":null,"abstract":"<div><div>Long-sequence time-series forecasting (LSTF) has broad real-world applications, and numerous methods have driven significant advancements in this field. However, challenges remain, such as dealing with distribution shifts and capturing the features from both global and local perspectives, such as trends and seasonal changes in time series. To address these issues, we propose an Auto/Cross-correlation Distribution Feature Extraction Network (ACDN), a linear complexity model that integrates two key modules and a mechanism: the Distribution Processing Module normalizes the input sequence and dynamically predicts the distribution of the forecasted sequence to capture the features of distribution shifts. The Auto/Cross-Correlation Module captures evolving trend components by computing both the autocorrelation of individual time series segments and the cross-correlation between different segments. The Subtle Feature Preservation Mechanism compensates for feature loss caused by dimensionality reduction in the encoder–decoder structure, ensuring critical fine-grained patterns are retained. Extensive experiments on nine datasets from diverse domains demonstrate the effectiveness of ACDN in multivariate LSTF tasks.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130806"},"PeriodicalIF":6.5000,"publicationDate":"2025-06-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"A distribution feature extracting network with dual correlation for long sequence time-series forecasting\",\"authors\":\"Jin Fan , Zehao Wang , Feiwei Qin , Huifeng Wu , Danfeng Sun , Jia Wu\",\"doi\":\"10.1016/j.neucom.2025.130806\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Long-sequence time-series forecasting (LSTF) has broad real-world applications, and numerous methods have driven significant advancements in this field. However, challenges remain, such as dealing with distribution shifts and capturing the features from both global and local perspectives, such as trends and seasonal changes in time series. To address these issues, we propose an Auto/Cross-correlation Distribution Feature Extraction Network (ACDN), a linear complexity model that integrates two key modules and a mechanism: the Distribution Processing Module normalizes the input sequence and dynamically predicts the distribution of the forecasted sequence to capture the features of distribution shifts. The Auto/Cross-Correlation Module captures evolving trend components by computing both the autocorrelation of individual time series segments and the cross-correlation between different segments. The Subtle Feature Preservation Mechanism compensates for feature loss caused by dimensionality reduction in the encoder–decoder structure, ensuring critical fine-grained patterns are retained. Extensive experiments on nine datasets from diverse domains demonstrate the effectiveness of ACDN in multivariate LSTF tasks.</div></div>\",\"PeriodicalId\":19268,\"journal\":{\"name\":\"Neurocomputing\",\"volume\":\"649 \",\"pages\":\"Article 130806\"},\"PeriodicalIF\":6.5000,\"publicationDate\":\"2025-06-30\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Neurocomputing\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S092523122501478X\",\"RegionNum\":2,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Neurocomputing","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S092523122501478X","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
A distribution feature extracting network with dual correlation for long sequence time-series forecasting
Long-sequence time-series forecasting (LSTF) has broad real-world applications, and numerous methods have driven significant advancements in this field. However, challenges remain, such as dealing with distribution shifts and capturing the features from both global and local perspectives, such as trends and seasonal changes in time series. To address these issues, we propose an Auto/Cross-correlation Distribution Feature Extraction Network (ACDN), a linear complexity model that integrates two key modules and a mechanism: the Distribution Processing Module normalizes the input sequence and dynamically predicts the distribution of the forecasted sequence to capture the features of distribution shifts. The Auto/Cross-Correlation Module captures evolving trend components by computing both the autocorrelation of individual time series segments and the cross-correlation between different segments. The Subtle Feature Preservation Mechanism compensates for feature loss caused by dimensionality reduction in the encoder–decoder structure, ensuring critical fine-grained patterns are retained. Extensive experiments on nine datasets from diverse domains demonstrate the effectiveness of ACDN in multivariate LSTF tasks.
期刊介绍:
Neurocomputing publishes articles describing recent fundamental contributions in the field of neurocomputing. Neurocomputing theory, practice and applications are the essential topics being covered.