Reservoir computing for system identification and model predictive control.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Dynamical Systems

Background:

• Model predictive control (MPC) requires accurate and efficient dynamical models for real-time operation.
• Complex systems often lack precise models or have computationally intensive ones, necessitating data-driven surrogate models.
• Echo state networks (ESNs) are recurrent neural networks suitable for approximating complex system dynamics.

Purpose of the Study:

• To evaluate Echo State Networks (ESNs) as data-driven surrogate models for system dynamics in Model Predictive Control (MPC).
• To compare the performance of ESN-based MPC against other architectures, specifically Long Short-Term Memory (LSTM) networks.
• To assess the efficiency and sample requirements of ESNs in the context of MPC.

Main Methods:

• Implementing MPC with ESNs as surrogate models for system dynamics.
• Benchmarking performance on control tasks including the Lorenz system and fluid flow.
• Comparing ESN-based MPC against LSTM-based MPC in terms of control objective achievement, cost, and variability.
• Analyzing training time and sample efficiency of ESNs versus LSTMs.

Main Results:

• ESN-based MPC consistently achieved control objectives on challenging benchmarks where LSTM-based MPC often failed.
• ESN-based MPC reduced average control cost by up to 10% and decreased variability by up to 85% compared to LSTM-based MPC.
• ESNs demonstrated superior sample efficiency and were over an order of magnitude faster to train than LSTMs.

Conclusions:

• ESNs are accurate and efficient data-driven surrogate models for MPC in complex systems with unknown dynamics.
• ESNs outperform LSTMs in terms of control performance, cost reduction, and training efficiency.
• ESNs enable scalable data-driven MPC, particularly when training data is limited.