As data is becoming more readily accessible from various systems, data-driven modeling and control approaches have gained increased popularity among both researchers and practitioners. Compared to its first-principle counterparts, data-driven approaches require minimal domain knowledge and calibration efforts, which is especially appealing to nonlinear and complex systems. In this thesis, two efficient data-driven frameworks, indirect and direct, for the control of nonlinear and complex systems are presented.The indirect method first involves an efficient online system identification approach with a composite local model structure. We introduce the concept of evolving Spatial Temporal Filters (eSTF) that dynamically transforms an incoming input-output data stream into a nonlinear combination of local models. Each local model is assigned with an ellipsoid-shaped cluster that is used to define its validity zone, and a distance metric that combines the Mahalanobis distance to the clusters and the scaled local model prediction error is exploited to compute the local model composition weights. The cluster and model parameters are efficiently updated online using input-output data stream, enabling adaptive system identification with efficient computations. With the identified eSTF model structure, we then develop an efficient quasi-linear parameter varying (qLPV) based stochastic model predictive controller (SMPC) for a class of nonlinear systems subject to chance constraints and additive disturbance. The qLPV form is established with the scheduling variable and a set of linear time-invariant models obtained from the eSTF system identification approach. To handle chance constraints, probabilistic reachable sets -- the probabilistic analogy of robust reachable sets -- are exploited to tighten the constraints to robustly guarantee constraint satisfaction despite model uncertainties and additive disturbances.A shifted scheduling variable strategy is designed such that the resultant MPC optimization can be efficiently solved by solving a series of quadratic programming problems. The indirect data-driven modeling and control pipeline is successfully applied to automotive powertrain systems with great system identification and control performance demonstrated. On the other hand, we further develop a direct data-driven control paradigm that leverages behavioural system theory, and directly generates control commands from input/output data without the need of a parametric model. We exploit singular value decomposition (SVD)-based order reduction to significantly reduce the online computation complexity without degrading the control performance. This control paradigm is successfully applied to battery fast charging, which has complex dynamics and is difficult to model. Furthermore, the direct data-driven approach heavily relies on the intensive collection and sharing of data, which raises serious privacy concerns, especially for systems with multiple agents. Therefore, we develop a privacy-preserving data-enabled predictive control scheme where we exploit affine-masking to protect the privacy of shared input/output data. It is then applied to the control of connected and automated vehicles (CAVs) in a mixed traffic environment with promising results demonstrated through comprehensive simulations.
Read
