Learning control for flex-fuel CI engine and fuel cell

Sammanfattning: This thesis investigated the modeling and control problems in the context of the flex-fuel compression-ignition (CI) engine and fuel cell, which shows great potential in the transition from fossil fuel to renewable energy sources.The modeling parts included the flex-fuel engine combustion process and intake system, and the system scale fuel cell model.The flex-fuel engine gas system models describing the intake pressure, temperature, oxygen concentration dynamics were established and validated with experimental data. The ignition delay was one key indicator of the combustion process and fuel properties and was modeled with a physical model and data-based models. A fuel cell physical model was built to illuminate the electrochemical behavior, and Gaussian process (GP) models were used to predict the voltage and hydrogen pressure with the collected data.Model predictive control (MPC) approaches based on physical models were applied to the flex-fuel CI engine and the fuel cell.An adaptive MPC method was proposed to control the combustion process of the flex-fuel CI engine. The control targets were combustion phasing and ignition delay. The adaptivity was done by estimating the physical ignition delay model parameters with real-time data online by Kalman filter. The proposed adaptive MPC approach showed the successful application in the fuel transition scenario with diesel, gasoline/n-heptane mixture, and ethanol/n-heptane mixture. An MPC with control constraints was developed to keep the fuel cell voltage at a reference value under current disturbance while satisfying the hydrogen pressure safety requirements. The state-space model was built by the simplification and linearization of the detailed system model. The proposed MPC controller fulfilled the control task and was compared with a PI controller.Learning-based MPC (LBMPC) integrated the learning models to the state-space model to improve the controller performance.One learning-based MPC method that decoupled the robustness and performance by maintaining two system models was proposed and applied to the control of combustion phasing when running with diesel. The comparison of LBMPC and MPC showed the improvement of performance by LBMPC.A GP MPC was developed to solve the fuel cell voltage control task with current disturbance and hydrogen pressure limit. Two GP predicting voltage and hydrogen pressure were integrated into the state-space model. The GP MPC showed comparable performance with MPC based on a detailed system physical model while requires less system information during operation.