Control of future low Temperature Combustion Technologies with nonlinear Model based Predictive Control based on Neural Networks

The combustion in future engines will work with a very high amount of recirculated exhaust gas in part load conditions to enable a low peak combustion temperature. This combustion suffers from instabilities of the process and a highly nonlinear behaviour. The paper presents the use of neural nets for observing the engine. A nonlinear model without feedback of measurements is linearised online and combined with an extended Kalman filter. This observer is compared to a neural net with observer structure by application to two different valve timing strategies. The more promising observer is combined with a model based predictive controller with a quadratic cost function. Its analytic solution is compared with quadratic programming for respecting constraints in the prediction for improving the control error.