Scholarlink

<p>Multivariate Time Series (MTS) analysis presents challenges due to complex feature correlations, non-linear data structures, and intricate temporal dynamics. Effective preprocessing, including data scaling, is essential to enhance dimensionality reduction techniques like Principal Component Analysis (PCA) and Truncated Singular Value Decomposition (tSVD). Traditional scaling methods, such as min–max and standard scaling, often limit the explained variance captured by principal components, reducing model efficacy. Addressing this, a novel hybrid scaling approach integrates standard scaling with min–max scaling to enhance normalization while maximizing variance retention and minimize the loss of information. This method improves the explained variance ratio by around 10%, allowing the initial components in PCA and tSVD transformations to capture a greater significance of the dataset's total variance, thereby enhancing dimensionality reduction effectiveness and retaining critical data information. Following this enhanced preprocessing, a Long Short-Term Memory (LSTM) network with Hybrid Layer Normalization (HLN) is introduced to capture MTS data non-linearity. The HLN strategically combines Logistic and Sigmoid-Curve normalization to balance local feature scaling and global trend adjustments, resulting in notable improvements in model generalization and predictive accuracy. Comprehensive evaluations across multiple metrics demonstrate this framework's superior performance compared to traditional scaling and contemporary techniques.</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

Optimizing Multivariate Time Series Forecasting with LSTM: A Hybrid Scaling and Layer Normalization Framework Utilizing Logistic and Sigmoid-Curve transformations for Enhanced Predictive Accuracy

  • 期刊论文

  • 发表日期：

    2025年9月25日

查看全文
  • Yuvaraja Boddu,
  • Manimaran A

摘要

Multivariate Time Series (MTS) analysis presents challenges due to complex feature correlations, non-linear data structures, and intricate temporal dynamics. Effective preprocessing, including data scaling, is essential to enhance dimensionality reduction techniques like Principal Component Analysis (PCA) and Truncated Singular Value Decomposition (tSVD). Traditional scaling methods, such as min–max and standard scaling, often limit the explained variance captured by principal components, reducing model efficacy. Addressing this, a novel hybrid scaling approach integrates standard scaling with min–max scaling to enhance normalization while maximizing variance retention and minimize the loss of information. This method improves the explained variance ratio by around 10%, allowing the initial components in PCA and tSVD transformations to capture a greater significance of the dataset's total variance, thereby enhancing dimensionality reduction effectiveness and retaining critical data information. Following this enhanced preprocessing, a Long Short-Term Memory (LSTM) network with Hybrid Layer Normalization (HLN) is introduced to capture MTS data non-linearity. The HLN strategically combines Logistic and Sigmoid-Curve normalization to balance local feature scaling and global trend adjustments, resulting in notable improvements in model generalization and predictive accuracy. Comprehensive evaluations across multiple metrics demonstrate this framework's superior performance compared to traditional scaling and contemporary techniques.