Optimized ensemble deep learning framework for scalable forecasting of dynamics containing extreme events.

Area of Science:

• Physics
• Computer Science
• Data Science

Background:

• Deep learning and ensemble methods are powerful tools for analyzing physical phenomena.
• These techniques are traditionally used independently, limiting their synergistic potential.
• Forecasting unpredictable chaotic dynamics, especially extreme events, presents significant scientific challenges.

Purpose of the Study:

• To develop an optimized ensemble deep learning (OEDL) framework integrating deep learning and ensemble methods.
• To achieve synergistic improvements in accuracy, stability, scalability, and reproducibility for dynamic system predictions.
• To advance the forecasting of nonlinear systems, with a specific focus on predicting extreme events.

Main Methods:

• Developed an optimized ensemble deep learning (OEDL) framework.
• Employed a best convex combination of feed-forward neural networks, reservoir computing, and long short-term memory (LSTM).
• Validated the framework on numerically simulated data and real-world datasets, including chaotic systems, epidemiological data, and climate data.

Main Results:

• The OEDL framework demonstrated superior out-of-sample performance compared to individual deep learners and standard ensemble methods.
• Achieved significant improvements in model accuracy, stability, scalability, and reproducibility.
• Successfully forecasted extreme events from a Liénard-type system, COVID-19 cases in Brazil, dengue cases in San Juan, and sea surface temperature in the Niño 3.4 region.

Conclusions:

• The proposed OEDL framework offers a powerful approach for synergistic improvements in forecasting complex dynamics.
• This integrated methodology is highly effective for predicting extreme events in both simulated and real-world scenarios.
• The OEDL model represents a significant advancement in applying machine learning for dynamic system prediction.