A Novel Encoder-Decoder Model for Multivariate Time Series Forecasting.

Area of Science:

• Data Mining
• Machine Learning
• Deep Learning

Background:

• Multivariate time series (MTS) data presents challenges due to high dimensionality, dynamic nature, and noise.
• Accurate MTS forecasting is critical in the era of big data for predicting variation trends.

Purpose of the Study:

• To propose a novel deep learning architecture for improved MTS forecasting.
• To enhance the accuracy and feasibility of MTS prediction models.

Main Methods:

• An encoder-decoder framework utilizing gated recurrent units (GRU) for feature extraction.
• Integration of an attention mechanism (AM) to weigh historical data importance during decoding.
• Implementation of skip connections from residual networks for feature reuse and reduced historical influence.
• Inclusion of convolutional and fully connected modules to boost performance and discriminative ability.

Main Results:

• The proposed deep learning architecture demonstrated effectiveness in MTS forecasting.
• Extensive experiments on stock and shared bicycle data validated the model's feasibility.
• The method successfully extracted successive features and utilized historical data importance.

Conclusions:

• The novel deep learning architecture offers a promising approach for accurate multivariate time series forecasting.
• The integration of GRU, attention mechanisms, and residual connections significantly enhances prediction capabilities.
• The model's effectiveness is confirmed across diverse real-world datasets.