Systematic Framework for Deep Learning-based Predictive Injection Control with Bayesian Hyperparameter Optimization for a Hydrogen/Diesel Dual-fuel Engine

Climate change and global warming concerns promote interest in alternative fuels, especially zero-carbon fuels like hydrogen. Modifying existing combustion engines for dual-fuel operation can decrease emissions of vehicles that are already on the road. The procedure of a deep learning-based model predictive control as a machine learning implementation, practical for complex nonlinear systems with input and state constraints, has been developed and tested on a hydrogen/diesel dual-fuel (HDDF) engine application. A nonlinear model predictive controller (NMPC) utilizing a deep neural network (DNN) process model is proposed to control the injected hydrogen and diesel. This DNN model has eight inputs and four outputs and has a short computational time compared to the physics-based model. The architecture and hyperparameters of the DNN model of the HDDF process are optimized through a two-stage Bayesian optimization to achieve high accuracy while minimizing the complexity of the model described. The final DNN architecture has two hidden layers with 31 and 23 neurons. A modified engine capable of HDDF operation is compared to standard diesel operation to evaluate the engine performance and emissions. During experimental engine testing, the controller required an average computational time of 2 ms per cycle on a low-cost processor, satisfying the real-time requirements, and was faster than recurrent networks. The control performance of the DNN-NMPC for the HDDF engine showed a mean absolute error of 0.19 bar in load tracking while maximizing average hydrogen energy share (68%) and reducing emissions. Specifically, the particulate matter emissions decrease by 87% compared to diesel operation.