This article presents an infinite horizon model predictive control (infinite MPC) framework with terminal cost learning for energy management of hybrid electric vehicles. The proposed framework integrates the model-based predictive capability of MPC with the learning potential of a neural network-based terminal cost function. The Monte Carlo method is employed to estimate long-term consumption, aiming to minimize the equivalent fuel consumption of the vehicle. To enhance the generalization of the terminal cost network, a vehicle speed Markov chain is utilized to create a stochastic environment. Leveraging historical data to train the terminal cost network extends the optimization horizon of MPC to infinity, enabling more energy-efficient energy management decisions. Additionally, a knowledge-guided cross-entropy method is adopted for rolling optimization. Simulation results demonstrate the effectiveness of the proposed approach in reducing equivalent fuel consumption. Compared to adaptive equivalent consumption minimization strategy (AECMS), the proposed method increases the proportion of operating points in the high-efficiency zone by 12% and reduces the equivalent fuel consumption by 11.3%. Using three different vehicle weights to test the robustness of the algorithm, the simulation results indicate that infinite MPC still shows better fuel economy than AECMS when facing parameter variations. Furthermore, engine-generator-set-in-the-loop experimental results reveal a 10% reduction in equivalent fuel consumption between AECMS and infinite MPC outcomes, validating the practical applicability of the approach.
Hybrid heavy-duty off-road vehicles frequently experience rapid acceleration and deceleration, as well as frequent uphill and downhill motion. Consequently, the engine must withstand aggressive transients which may drastically worsen the fuel economy and even cause powertrain abnormal operation. When the engine cannot respond to the transient demand power quickly enough, the battery must compensate for the large amount of power shortage immediately, which may cause excessive battery current that adversely affects the battery safety and life span. In this paper, a nonlinear autoregressive with exogenous input neural network is used to recognize the driver’s intention and translate it into subsequent vehicle speed. Combining energy management with vehicle speed control, a co-optimization-based driver-oriented energy management strategy for manned hybrid vehicles is proposed and applied to smooth the engine power to ensure efficient operation of the engine under severe transients and, at the same time, to regulate battery current to avoid overload. Simulation and the hardware-in-the-loop test demonstrate that, compared with the filter-based energy management strategy, the proposed strategy could yield a 38.7% decrease in engine transient variation and an 8.2% decrease in fuel consumption while avoiding battery overload. Compared with a sequential-optimization-based energy management strategy, which is recognized as a better strategy than a filter-based energy management strategy, the proposed strategy can achieve a 16.2% decrease in engine transient variation and a 3.2% decrease in fuel consumption.
27 February 2025 in Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering
Cooling system is a crucial subsystem essential for the engines, and its condition monitoring plays an important role in the engine safety and reliability. This paper proposes an innovative deep digital twin (DDT) model that combines Gradient Boosting Decision Tree (GBDT) based ensemble learning and Stacked Sparse Autoencoder (SSAE) based deep learning to enhance the sensitivity and accuracy of engine cooling system condition monitoring (CM). The Gradient Boosting Decision Tree (GBDT) algorithm is employed to generate coolant temperature baselines of healthy state under varying operational conditions. Then, taking the coolant temperature as the characteristic parameter, the health feature extraction model is constructed using Stacked Sparse Autoencoder (SSAE) network to extract the feature representations from coolant temperature under different health states. Specifically, to enhance the efficiency of feature extraction, this paper introduced modifications to the structure of the SSAE. To assess the health state of the cooling system quantitively, the probability density function (PDF) of the feature representations was calculated, with Kullback-Leibler (KL) divergence serving as the health indicators (HI). The severity of cooling system’s degradation was indicated by comparing the deviation in KL divergence between healthy and degradation states. Simulation and experimental data validation demonstrate the capability of the proposed method in cooling system condition monitoring.
This paper describes the application of a NSGA-II-based integrated optimization of battery sizing and control parameter for a hybrid amphibious vehicle. Firstly, the amphibious vehicle model is built for verification. Secondly, main power system parameters are selected through theoretical analysis and filter-based energy management strategy is introduced. Then NSGA-I is applied to optimize the number of battery cells and cutoff frequency, which is the key parameter of energy management strategy. The simulation results show that the NSGA-I -based integrated optimization exhibits better performance in terms of battery SOH and fuel economy than traditional matching method.