Graph deep-learning method for integrated topological and parametric optimization of hybrid electric vehicle powertrain

This paper presents a novel graph deep-learning aided optimization methodology for the integrated optimization of hybrid electric vehicle (HEV) powertrain topologies, component parameters, and energy management strategy. Jointly representing topological structures and component parameters, clarifying the design domain, and efficiently evaluating the optimal performance of candidate designs are correlated challenges in current topology and component size optimization methods. To address this, an innovative graph data-driven representation model is developed to extract the comprehensive features of system topology structure and component attributes, and it outputs a feature of the powertrain system in a latent low-dimensional real space. Supported by this, graph deep-learning-based classifier and predictor models are constructed for feasibility discrimination and optimal energy consumption prediction, respectively, which initiates a graph data-driven integrated optimization method of HEV powertrains with a two-layer framework. The verification results indicate that the optimized design achieves fuel economy improvements ranging from 4.43 % to 27.47 % compared to benchmarks adopting various system topologies and component sizing schemes, while the fuel consumption of the superior design produced by the proposed method is only 0.1 % different from the global optimal solution, demonstrating the significant effectiveness of this approach to enhance fuel economy of HEV powertrains.