AA-forecast: anomaly-aware forecast for extreme events.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Real-world time series data frequently contain extreme events and anomalies, complicating accurate probabilistic forecasting.
• Effective risk management, particularly for events like pandemics or hurricanes, necessitates reliable predictions even amidst data irregularities.
• Current methods often struggle with automatic anomaly detection and learning in large datasets, leading to increased manual intervention.

Purpose of the Study:

• To develop an anomaly-aware forecasting framework capable of automatically detecting and learning from extreme events and anomalies.
• To enhance the accuracy and reduce the uncertainty of time series predictions during periods of high volatility.
• To provide a more efficient and automated approach to forecasting for large-scale, anomaly-prone datasets.

Main Methods:

• Proposed an anomaly-aware forecast framework utilizing an attention mechanism to incorporate detected anomalies.
• Implemented an automated anomaly extraction and learning process within the model.
• Employed a dynamic uncertainty optimization algorithm for online uncertainty reduction.

Main Results:

• The framework demonstrated superior accuracy compared to existing prediction models across datasets with diverse anomalies.
• The model successfully reduced forecast uncertainty in an online manner.
• Automated anomaly detection and incorporation led to improved prediction performance during extreme events.

Conclusions:

• The proposed anomaly-aware framework offers a robust solution for improving time series forecasting accuracy and reducing uncertainty.
• Automated anomaly handling is crucial for effective risk management in the presence of extreme events.
• The framework provides a scalable and efficient alternative to manual data processing for anomaly-rich datasets.