Scholarlink

<p>Convolutional neural network (CNN)-long short-term memory (LSTM) models have been extensively utilized in time-series forecasting (TSF) owing to their robust capabilities in feature extraction and capturing temporal dependencies. However, they face challenges in long-term time series forecasting (LTSF), because the recurrent iterations inherent to LSTM over extended lookback and forecast horizons result in the accumulation of errors. Based on the PatchTST method, this study introduces ECLNet, a novel approach for LTSF. ECLNet is an enhanced CNN-LSTM framework that improves forecasting accuracy and efficiency. ECLNet incorporates a hidden state parallel forecasting (HSPF) strategy to reduce the number of recurrent iterations, thereby decreasing cumulative errors, improving predictive accuracy, and enhancing training speed. Distinct from the prevalent channel-independent strategies in existing models, we propose a dual ConvFFN interrelation module (DCIM) to model the complex relationships among multiple variables and channels, thereby enhancing the capture of intricate dependencies, while leveraging LSTM units to capture temporal dependencies within the data. This approach improved the forecasting performance and robustness of the model. We further compared our ECLNet model with all baselines through experiments on six public datasets, which indicated that our proposed model demonstrated a competitive performance comparable to state-of-the-art models on the evaluated benchmarks.</p>

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

ECLNet: enhancing the CNN-LSTM networks for multivariate long-term time series forecasting

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Jiachen Xie,
  • Jiwei Qin,
  • Xizhong Qin,
  • Qiang Li,
  • Daishun Cui,
  • Dezhi Sun

摘要

Convolutional neural network (CNN)-long short-term memory (LSTM) models have been extensively utilized in time-series forecasting (TSF) owing to their robust capabilities in feature extraction and capturing temporal dependencies. However, they face challenges in long-term time series forecasting (LTSF), because the recurrent iterations inherent to LSTM over extended lookback and forecast horizons result in the accumulation of errors. Based on the PatchTST method, this study introduces ECLNet, a novel approach for LTSF. ECLNet is an enhanced CNN-LSTM framework that improves forecasting accuracy and efficiency. ECLNet incorporates a hidden state parallel forecasting (HSPF) strategy to reduce the number of recurrent iterations, thereby decreasing cumulative errors, improving predictive accuracy, and enhancing training speed. Distinct from the prevalent channel-independent strategies in existing models, we propose a dual ConvFFN interrelation module (DCIM) to model the complex relationships among multiple variables and channels, thereby enhancing the capture of intricate dependencies, while leveraging LSTM units to capture temporal dependencies within the data. This approach improved the forecasting performance and robustness of the model. We further compared our ECLNet model with all baselines through experiments on six public datasets, which indicated that our proposed model demonstrated a competitive performance comparable to state-of-the-art models on the evaluated benchmarks.