Scholarlink

<p>Multivariate time-series (MTS) forecasting is still a difficult problem because of high dimensionality, complex temporal dependencies, and noisy inter-feature relationships. Most models find it difficult to be both accurate, interpretable, able to handle nonlinearities, and computationally efficient at the same time. In a bid to overcome these challenges, this work presents QTL-Net, a new hybrid forecasting model that combines quartet principal component analysis (QPCA), temporal convolutional networks (TCNs), and bidirectional long short-term memory (Bi-LSTM) networks with logistic–sigmoid curve-based layer normalization. QPCA successfully eliminates dimensionality while retaining critical variance and temporal structure, thus reducing redundancy in high-dimensional data. TCNs utilize dilated and residual convolutions to obtain localized temporal features with low computational cost. Bi-LSTM layers capture intricate bidirectional dependencies, bolstering long-term forecasting capacity. The application-specific logistic–sigmoid normalization also stabilizes training by mitigating vanishing and exploding gradient problems. Large-scale experiments on benchmark MTS datasets show that QTL-Net outperforms traditional methods like moving average, LSTM, and GRU in every aspect of forecasting accuracy, training stability, and computational scalability. By combining statistical feature extraction, deep temporal encoding, and recurrent sequence modeling, QTL-Net presents an effective and interpretable solution for pushing state-of-the-art multivariate time-series forecasting to a higher level. Additionally, proposed envisioned QTL-Net architecture is tailored for real-time forecasting on HPC-enabled platforms, leveraging parallelizable dimensionality reduction and hybrid temporal learning for scalable performance on big multivariate datasets. Its computational architecture naturally is amenable to GPU acceleration and distributed training and thus is well-suited for deployment in real-time high-throughput analytical platforms.</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

QTL-Net: an advanced quartet PCA-driven temporal learning model with logistic–sigmoid normalization for accurate multivariate forecasting

  • 期刊论文

  • 发表日期：

    2025年11月14日

查看全文
  • Yuvaraja Boddu,
  • A Manimaran

摘要

Multivariate time-series (MTS) forecasting is still a difficult problem because of high dimensionality, complex temporal dependencies, and noisy inter-feature relationships. Most models find it difficult to be both accurate, interpretable, able to handle nonlinearities, and computationally efficient at the same time. In a bid to overcome these challenges, this work presents QTL-Net, a new hybrid forecasting model that combines quartet principal component analysis (QPCA), temporal convolutional networks (TCNs), and bidirectional long short-term memory (Bi-LSTM) networks with logistic–sigmoid curve-based layer normalization. QPCA successfully eliminates dimensionality while retaining critical variance and temporal structure, thus reducing redundancy in high-dimensional data. TCNs utilize dilated and residual convolutions to obtain localized temporal features with low computational cost. Bi-LSTM layers capture intricate bidirectional dependencies, bolstering long-term forecasting capacity. The application-specific logistic–sigmoid normalization also stabilizes training by mitigating vanishing and exploding gradient problems. Large-scale experiments on benchmark MTS datasets show that QTL-Net outperforms traditional methods like moving average, LSTM, and GRU in every aspect of forecasting accuracy, training stability, and computational scalability. By combining statistical feature extraction, deep temporal encoding, and recurrent sequence modeling, QTL-Net presents an effective and interpretable solution for pushing state-of-the-art multivariate time-series forecasting to a higher level. Additionally, proposed envisioned QTL-Net architecture is tailored for real-time forecasting on HPC-enabled platforms, leveraging parallelizable dimensionality reduction and hybrid temporal learning for scalable performance on big multivariate datasets. Its computational architecture naturally is amenable to GPU acceleration and distributed training and thus is well-suited for deployment in real-time high-throughput analytical platforms.