Intelligent energy management in hybrid electric cycling work vehicles: A multi-task learning approach

Learning-based approaches have become mainstream solutions for achieving intelligent energy management in hybrid electric vehicles (HEVs). For hybrid electric cycling work vehicles (HECWVs), predictable driving routes facilitate model learning. Nevertheless, effectively tailoring and deploying learning-based energy management strategies specifically for HECWVs encounters significant hurdles. These primarily involve substantial data requirements to capture diverse operational tasks, achieving robust generalization across heterogeneous scenarios, and ensuring the interpretability of the resultant control policies. To address these issues in HECWVs, this study pioneers the introduction of a multi-task learning mechanism and presents a customized energy management approach. Specifically, a work condition clustering module using Self-Organizing Maps (SOM) enables efficient dimensionality reduction and intuitive visual clustering. Additionally, the prioritized sweeping method is introduced within the multi-task learning framework to enhance learning efficiency and establish an efficient cross-task sharing mechanism. Finally, this study introduces an asymmetric sharing mechanism to accommodate diverse working conditions among HECWVs, thereby enhancing strategic decision-making responsiveness. This enhancement is achieved by guiding targeted interactions among tasks through an asymmetric parameter-sharing structure aligned with the proportion of working conditions. Real data analysis demonstrates improved model interpretability and adaptability in sparse data scenarios. Additionally, it outperforms traditional model-based learning approaches in both optimality and strategic response speed in environments with heterogeneous driving characteristics. Notably, fuel economy is improved by 5.3 %, and the learning response speed is enhanced by 388 %.