Closed-loop Feedback based Adaptive Model Predictive Control of a Bluff-Body Wake

Achieving robust closed-loop feedback control of turbulent bluff-body wakes poses challenges due to the non-linear response of the wake dynamics to actuation. A strategy based on Model Predictive Control (MPC) with a Long Short-Term Memory (LSTM) Neural Network model to achieve closed-loop control of a bluff-body wake is outlined and its performance assessed experimentally.

The wake control strategy is presented for a square-section cylinder with two moving surface actuators at the leading corners. The LSTM is trained using actuation rate and feedback surface-pressure history to forecast future pressure states. An adaptive learning algorithm adjusts the model to new conditions beyond the original training data, increasing robustness over a Reynolds number range.

Recovery of mean pressure setpoint, and both drag and wake fluctuation intensity optimization tests were used to assess controller performance. The controller showed significant improvements over benchmark open-loop tests results. The model trained at a single Re quickly adapted to conditions in the range 10000 < Re < 20000. Particle Image Velocimetry was used to determine the flow mechanisms affected by the different actuations. It is found that low and high frequency actuations affect controller performance differently.