Constantly growing demands on the pollutant emissions and fuel consumption of motor vehicles are, among other things, driving the development of locally emission-free powertrains. The resulting increase in the degree of freedom and complexity of the powertrain, especially in hybrid vehicles, requires new control concepts, also known as control strategies. Researchers agree that efficiency can be further increased by incorporating predictive information into control strategies. This leads to the

more »

... model predictive control methods, which however are often not real-time capable or only achieve this by linearizing the simplest physical models. The goal of this dissertation is the development of data-based, model predictive methods, which do not contain any linearizations, but nevertheless allow a real-time capable application. After a detailed description of the state of the art, a two-part methodology is described. In the first step, the hybrid powertrain's nonlinear, dynamic process behavior is approximated using system identification methods. Since the common methods of system identification produce insufficient model accuracy due to the complex, dynamically coupled system, an augmented time-delay neural network (TDNN) architecture is proposed. This architecture shows sufficient model accuracy and advantages in computational efficiency, especially with respect to the intended application. The second step of the overall methodology consists of the nonlinear model predictive control. In order to deliver optimal results for the control, the cost function is first established and a suitable optimization algorithm is selected. Furthermore, stabilizing measures are investigated and implemented to ensure the robust operation of the control strategy. To further collateralize the model predictive approach, a redundant, locally optimal control strategy is incorporated. In the next step, the developed methodology is applied in a simulation environment, respecting certification cycles and everyday driving. The new approach is c [...]