A novel extreme adaptive GRU for multivariate time series forecasting

Area of Science:

• Machine Learning
• Deep Learning
• Time Series Analysis

Background:

• Multivariate time series forecasting is crucial for real-world applications.
• Imbalanced data with extreme events pose a significant challenge for existing forecasting models.
• Current methods often neglect extreme events, leading to suboptimal performance.

Purpose of the Study:

• To introduce a novel deep learning model, the Extreme Event Adaptive Gated Recurrent Unit (eGRU), for accurate multivariate time series forecasting.
• To address the challenge of imbalanced time series data containing extreme events.
• To improve the learning of both normal and extreme event patterns.

Main Methods:

• Developed the Extreme Event Adaptive Gated Recurrent Unit (eGRU) model.
• Introduced a time series data segmentation technique to process sequences at different resolutions and reduce input length.
• Conducted experiments on four real-world benchmark datasets.

Main Results:

• The eGRU model demonstrated superior performance compared to vanilla RNNs, LSTMs, GRUs, and other state-of-the-art RNN variants.
• Ablation studies confirmed the consistently superior forecasting accuracy of eGRU.
• The model effectively incorporated diverse labeling results.

Conclusions:

• The eGRU model offers a significant advancement in multivariate time series forecasting, particularly for imbalanced datasets.
• The proposed segmentation technique enhances the model's ability to capture complex temporal patterns.
• eGRU provides a robust solution for forecasting tasks impacted by extreme events.