Research on multi-objective optimization control of diesel engine combustion process based on model predictive control-guided reinforcement learning method

As environmental protection regulations become increasingly stringent, traditional control strategies for diesel engines often struggle to simultaneously satisfy real-time performance, fuel economy, and low emission requirements during transient processes. Consequently, this study presents a multi-objective real-time optimization control method that integrates model predictive control (MPC) with reinforcement learning (RL) to regulate rail pressure and injection timing for the synergistic optimization of nitrogen oxide (NOx) emissions and indicated mean effective pressure (IMEP). The results indicate that while the MPC algorithm demonstrates strong performance in short-term optimization during simulation validation, its reliance on limited short-term predictions leads to substantial fluctuations in control parameters. Conversely, the RL algorithm exhibits superior synergistic optimization capabilities, however, it suffers from poor training stability. After incorporating superior samples from the MPC to inform RL training, both stability and training efficiency were significantly enhanced, resulting in a NOx emission reduction of 31.4 %, which markedly surpasses the 19.25 % improvement achieved by the MPC alone. Furthermore, the RL algorithm demonstrated higher real-time computational efficiency. In experimental validation, compared to the original control map settings, the MPC-guided RL algorithm reduced NOx emissions by 16 % and increased IMEP by 0.32 %, thereby achieving multi-objective real-time optimization of the diesel engine combustion process.