Deep belief improved bidirectional LSTM for multivariate time series forecasting.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multivariate time series (MTS) data is prevalent but challenging to forecast due to complex temporal dependencies.
• Traditional MTS forecasting methods face limitations in speed and complexity.
• Existing Long Short-Term Memory (LSTM) networks partially capture spatial relationships, with shallow LSTMs struggling with high-dimensional data.

Purpose of the Study:

• To develop an improved Bidirectional LSTM (BiLSTM) network, termed DBI-BiLSTM, for enhanced MTS forecasting.
• To leverage Deep Belief Networks (DBN) and a chained structure for deeper feature extraction and bidirectional learning.
• To improve the accuracy and efficiency of MTS forecasting compared to existing methods.

Main Methods:

• A novel DBI-BiLSTM architecture integrating a DBN layer with stacked BiLSTM layers.
• Input data processed by DBN for comprehensive feature extraction.
• Clustered features, identified via global sensitivity analysis, fed into BiLSTM layers.
• Outputs from shallower layers combined with clustered features for deeper layer inputs.

Main Results:

• The DBI-BiLSTM model demonstrated superior one-step-ahead prediction performance on four real-world MTS datasets.
• Significant improvements were observed compared to traditional Artificial Neural Networks (ANNs), deep LSTMs, and other advanced LSTM variants.
• Percentage improvements over conventional LSTM ranged from 30.72% to 85.41% across the datasets.

Conclusions:

• The proposed DBI-BiLSTM effectively captures intricate spatial relationships and enhances feature representation in MTS data.
• DBI-BiLSTM offers a powerful and accurate solution for multivariate time series forecasting challenges.
• The model's deep structure and bidirectional learning capabilities contribute to its advanced performance.