Multivariate Time Series Forecasting Using Multiscale Recurrent Networks With Scale Attention and Cross-Scale...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Multivariate time series (MTS) forecasting is complex due to nonlinear interdependencies.
• Existing deep learning models often lose scale information by being single-scale oriented.
• Recurrent neural networks (RNNs) with attention mechanisms model temporal patterns but struggle with multiscale data.

Purpose of the Study:

• To develop novel deep learning frameworks for MTS forecasting that incorporate multiscale analysis.
• To address the limitation of scale information loss in existing single-scale forecasting models.
• To propose two new multiscale recurrent network (MRN) models: MRN-SA and MRN-CSG.

Main Methods:

• Integration of multiscale analysis into deep learning frameworks to create scale-aware recurrent networks.
• MRN-SA model utilizes scale attention, input attention, and temporal attention for dynamic information selection.
• MRN-CSG model employs a cross-scale guidance mechanism for efficient, lightweight forecasting.

Main Results:

• Both MRN-SA and MRN-CSG models achieved state-of-the-art performance on five diverse MTS datasets.
• MRN-CSG demonstrated effectiveness as a lightweight and easily trainable model without significant accuracy compromise.
• The proposed models successfully handle complex temporal patterns and inter-series dependencies.

Conclusions:

• The developed multiscale recurrent networks (MRNs) significantly advance MTS forecasting capabilities.
• Integrating multiscale analysis is crucial for improving the accuracy and robustness of time series forecasting models.
• The MRN models offer promising solutions for real-world applications requiring accurate MTS predictions.