A Safe Hierarchical Control Scheme Incorporated with Constrained Deep Neural Networks

A methodology is proposed for the control design of hierarchical systems. The high layer makes a decision without knowing the overall systems dynamics precisely. This procedure is implemented by a deep neural network constructed offline. Besides, to handle the safe learning issue, an additional technique is developed to guarantee constraints satisfaction of the trained neural network. At the underlying layer, local model predictive control (MPC) controllers are designed to regulate the subsystems to follow the optimal trajectories derived from high layer. The sample data from lower layer is utilized as the training set of neural network in high layer. The efficacy of the presented algorithm is illustrated in a simulation example.