Conditional POD for predicting extreme events in turbulent flow time signals

Area of Science:

• Fluid Dynamics
• Turbulence Research
• Data-Driven Science

Background:

• Extreme events in turbulent flows are rare but impactful, posing prediction challenges.
• Traditional data-intensive methods struggle with the intermittent and infrequent nature of these events.
• Accurate prediction is crucial for system reliability and performance in various engineering applications.

Purpose of the Study:

• To develop a novel data-driven approach for early-stage prediction of extreme events in time signals.
• To enable on-the-fly detection of rare events without extensive training data.
• To improve the reliability of turbulent flow predictions by flagging potentially inaccurate forecasts.

Main Methods:

• Identifies the most energetic time-only Proper Orthogonal Decomposition (POD) mode from segments preceding extreme events.
• Uses a support vector machine (SVM) to classify incoming signals based on similarity to the identified POD mode.
• Applies the conditional POD method to predict extreme dissipation events in wall-bounded shear flows.

Main Results:

• The method demonstrates robust performance in predicting extreme dissipation events across various Reynolds numbers and wall distances.
• Achieves prediction lead times that match or exceed those of the Hankel-DMD method, even with limited training data.
• Successfully flags incoming extreme events, allowing for the potential discarding or shortening of unreliable forecasts from other methods.

Conclusions:

• The conditional POD method offers a viable data-driven solution for early prediction of rare extreme events in turbulent flows.
• This approach is particularly effective for events with limited data availability.
• Enables enhanced reliability of turbulent flow prediction systems by providing timely alerts for extreme events.