(458e) Machine Learning-Based Model Predictive Control Using Real-Time Transfer Learning

Transfer learning, a well-known framework in the field of machine learning, involves transferring knowledge from a data-rich source process to a data-scarce target process [1]. Typically, the construction of high-fidelity machine learning models relies primarily on utilizing extensive training datasets that are often not available in practice in engineering systems. However, transfer learning offers a method to significantly enhance the prediction accuracy by only modifying a pre-existing model trained on an extensive data set from a source process through subsequent integration of limited data and/or relevant physics from a target process [2]. This adaptation aims to not only reduce the data requirements but also the training time for model training, significantly expediting the training process compared to developing a machine learning model for the target process entirely from the beginning [1].

In this study, our objective is to improve a machine learning model via transfer learning in real-time. Specifically, starting from a model trained for a source process, defined here as time-varying operation within a restricted operating region of the state-space, using an extensive dataset generated offline operation, the goal is to update the model in real-time to capture the dynamics of the target process, where the process evolves in an unexplored operating region. Transfer learning occurs only when the model accuracy is observed to deteriorate and new data becomes available. Subsequently, we integrate this machine learning model into a real-time model predictive control (MPC) framework, while initiating the process from the unexplored new operating region. As the MPC computes the optimal control action and the process evolves in real-time, we conduct real-time transfer learning by collecting real-time data alongside an offline, pre-existing dataset, defined within the new operating region, which is used to update the model weights after freezing part of the network. The real-time transfer learning framework is applied to a non-isothermal CSTR operating at an unstable steady state under a Lyapunov-based MPC to demonstrate its effectiveness at maintaining the state near the setpoint by updating the neural network model as necessary.

References:

[1] Xiao, M., Hu, C., & Wu, Z. (2023). Modeling and predictive control of nonlinear processes using transfer learning method. AIChE Journal, 69, e18076.

[2] Weber, M., Auch, M., Doblander, C., Mandl, P., & Jacobsen, H. A. (2021). Transfer learning with time series data: a systematic mapping study. IEEE Access, 9, 165409-165432.