Driving cycle reproduction supported deep learning co-optimization of freight EV powertrains in stochastic operation environments

Vehicles developed and validated under the fixed standard driving cycle lack the robustness of performance when they are utilized in real stochastic environments. This article presents a deep-learning aided method that simultaneously optimizes the expectation and susceptibility of the energy efficiency of FEV powertrains in stochastic driving environments, addressing the intertwined restrictions among driving cycles, control policies, and powertrain design. To aid the optimization, deep learning models are developed to represent the stochastic environments by generating new driving cycles, clarify design constraints, and efficiently evaluate candidate designs. Compared with collected driving cycles, the result demonstrates that the generated cycles have a consistent influence on the powertrain design decision, as the energy consumption regression factor of candidate designs in the two types of driving cycles reaches over 0.9. For deep learning models aiding the evaluation of design schemes, the overall accuracy of the classifier reaches 95.2 % reflecting an accurate representation of complex design constraints, and the R-square of the predictor is 0.97, leading to an effective and efficient evaluation of the optimal performance of candidate designs. Ultimately, the expectation and susceptibility of the energy efficiency of the FEV are comprehensively optimized, as the infeasible rate of completing diverse driving cycles declines from 31 % to 3 % within an energy economy gap of about 5 %, compared with the benchmark design.