Scholarlink

<p>Forecasting Multivariate Time Series remains a challenging task due to high dimensionality, nonlinear temporal dependencies, and noise-induced feature correlations. Traditional forecasting and dimensionality reduction methods, including PCA and ICA, often lose temporal coherence and discard essential variance, leading to suboptimal predictive accuracy. For addressing the attendant limitations, this study presents a novel Quartet Truncated Singular Value Decomposition–Temporal Convolutional Network Bidirectional Long Short-Term Memory QtSVD–TCN–Bi-LSTM hybrid framework integrated with Logistic–Sigmoid Curve-Based Hybrid Layer Normalization. The proposed QtSVD method, in this context, serves as an effective means to reduce dimensionality while retaining essential temporal and statistical variance. The TCN learns local temporal features due to causal and dilated convolutions and enables efficient modeling of sequences, while the Bi-LSTM model’s long-range bidirectional dependencies to refine temporal learning. Logistic–Sigmoid HLN dynamically regularizes the gradient flow and prevents vanishing and exploding gradients, hence enhancing convergence stability. Extensive experiments on benchmark MTS datasets of various problems, such as crop growth, infrared thermography, ERP, and air pollution, substantiate that the proposed framework outperforms conventional and hybrid deep learning architectures involving LSTM, GRU, and PCA-LSTM. The model exhibits lower RMSE, MAE, and MAPE, while providing higher R² and generalization capability. This paper is thus establishing QtSVD–TCN–Bi-LSTM combined with HLN as a scalable, robust, and interpretable paradigm for high-dimensional time series forecasting, effectively bridging statistical reduction, convolutional feature extraction, and recurrent temporal learning.</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

Leveraging Quartet Dimensionality Reduction, Deep Learning, and Logistic-Sigmoid Curve Transformation Layer Normalization for Accurate Multivariate Time Series Forecasting

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Yuvaraja Boddu,
  • A. Manimaran

摘要

Forecasting Multivariate Time Series remains a challenging task due to high dimensionality, nonlinear temporal dependencies, and noise-induced feature correlations. Traditional forecasting and dimensionality reduction methods, including PCA and ICA, often lose temporal coherence and discard essential variance, leading to suboptimal predictive accuracy. For addressing the attendant limitations, this study presents a novel Quartet Truncated Singular Value Decomposition–Temporal Convolutional Network Bidirectional Long Short-Term Memory QtSVD–TCN–Bi-LSTM hybrid framework integrated with Logistic–Sigmoid Curve-Based Hybrid Layer Normalization. The proposed QtSVD method, in this context, serves as an effective means to reduce dimensionality while retaining essential temporal and statistical variance. The TCN learns local temporal features due to causal and dilated convolutions and enables efficient modeling of sequences, while the Bi-LSTM model’s long-range bidirectional dependencies to refine temporal learning. Logistic–Sigmoid HLN dynamically regularizes the gradient flow and prevents vanishing and exploding gradients, hence enhancing convergence stability. Extensive experiments on benchmark MTS datasets of various problems, such as crop growth, infrared thermography, ERP, and air pollution, substantiate that the proposed framework outperforms conventional and hybrid deep learning architectures involving LSTM, GRU, and PCA-LSTM. The model exhibits lower RMSE, MAE, and MAPE, while providing higher R² and generalization capability. This paper is thus establishing QtSVD–TCN–Bi-LSTM combined with HLN as a scalable, robust, and interpretable paradigm for high-dimensional time series forecasting, effectively bridging statistical reduction, convolutional feature extraction, and recurrent temporal learning.