Scholarlink

<p>Precise multivariate time series (MTS) forecasting, particularly in atmospheric applications like air quality monitoring, is still a challenging task because of high dimensionality, temporal correlations, and non-stationary interactions between features. The classical methods such as Auto Regressive Integrated Moving Average) ARIMA and isolated Long Short-Term Memory (LSTM) are likely to fail to capture nonlinear relationships and are highly sensitive to the scale of features, and Principal Component Analysis (PCA) based dimensionality reduction is likely to result in information loss. To mitigate these constraints, we introduce PIHS-Bi-LSTM-GRU, a deep learning hybrid model that combines PCA-ICA-based reduction of dimensions, multi-level hybrid scaling of features, and an improved Bidirectional LSTM- Gated Recurrent Unit (GRU) architecture with dual-layer normalization and dropout. Our approach begins by taking a weighted ensemble of Min-Max, Z-Score, and Robust scalers for stabilizing heterogeneous distributions of features. PCA is used to alleviate redundancy, followed by Independent Component Analysis (ICA) to yield statistically independent latent signals. The deep learning model subsequently learns temporal patterns from the transformed sequences. A new component-wise inverse transformation mechanism provides exact reconstruction in the original feature space. Comprehensive evaluation on actual air quality data shows that the proposed model considerably outperforms baseline methods in Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and <InlineEquation ID="IEq1"><EquationSource Format="TEX">\(R^{2}\)</EquationSource></InlineEquation> on all features. The findings verify the effectiveness of the framework in picking up on intricate temporal-spatial relations and enhancing predictive reliability under multivariate prediction situations.</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

A hybrid PCA-ICA and multi-level feature scaling framework with bidirectional LSTM-GRU architecture improves multivariate time series forecasting accuracy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月18日

查看全文
  • Yuvaraja Boddu,
  • A. Manimaran,
  • Jayanth Talabathula,
  • M. Sucharitha

摘要

Precise multivariate time series (MTS) forecasting, particularly in atmospheric applications like air quality monitoring, is still a challenging task because of high dimensionality, temporal correlations, and non-stationary interactions between features. The classical methods such as Auto Regressive Integrated Moving Average) ARIMA and isolated Long Short-Term Memory (LSTM) are likely to fail to capture nonlinear relationships and are highly sensitive to the scale of features, and Principal Component Analysis (PCA) based dimensionality reduction is likely to result in information loss. To mitigate these constraints, we introduce PIHS-Bi-LSTM-GRU, a deep learning hybrid model that combines PCA-ICA-based reduction of dimensions, multi-level hybrid scaling of features, and an improved Bidirectional LSTM- Gated Recurrent Unit (GRU) architecture with dual-layer normalization and dropout. Our approach begins by taking a weighted ensemble of Min-Max, Z-Score, and Robust scalers for stabilizing heterogeneous distributions of features. PCA is used to alleviate redundancy, followed by Independent Component Analysis (ICA) to yield statistically independent latent signals. The deep learning model subsequently learns temporal patterns from the transformed sequences. A new component-wise inverse transformation mechanism provides exact reconstruction in the original feature space. Comprehensive evaluation on actual air quality data shows that the proposed model considerably outperforms baseline methods in Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and \(R^{2}\) on all features. The findings verify the effectiveness of the framework in picking up on intricate temporal-spatial relations and enhancing predictive reliability under multivariate prediction situations.