基于GCN-TCN的大功率HMCVT拖拉机混合动力学建模方法

GCN-TCN-based Hybrid Longitudinal Dynamic Modeling of High-power HMCVT Tractor

摘要:

针对大功率液压机械无级变速（HMCVT）拖拉机在复杂田间作业工况下纵向动力学模型精度不足、传统机理模型适应性受限等问题，在AMESim-Simulink联合仿真模型与实车试验数据基础上，本文提出了一种融合图卷积网络（GCN）与时间卷积网络（TCN）的混合纵向动力学建模方法。该方法以拖拉机动力系统实际结构为约束，将发动机、传动系统及整车纵向状态抽象为动力学状态图，通过GCN表征动力学状态间的结构耦合关系，并利用TCN描述纵向动力学状态的时间演化特性，实现时空特征的联合建模。在多种典型田间作业工况下对所建模型进行了试验验证，并采用工况加权灰色关联评价方法对瞬态与稳态预测误差进行统一量化评价。结果表明，所提出的GCN-TCN混合模型在不同工况下均表现出较高预测精度与稳定性，多工况综合预测精度达到97.693%，在负载突变和非稳态工况下优势尤为明显。研究结果可为大功率拖拉机传动系统控制策略优化与整机性能分析提供可靠的纵向动力学模型支撑。

Abstract:

Aiming to overcome the insufficient accuracy of longitudinal dynamic models and the limited adaptability of conventional mechanism-based approaches for high-horsepower hydraulic mechanical continuously variable transmission (HMCVT) tractors operating under complex field conditions, a hybrid longitudinal dynamic modeling method that integrated graph convolutional networks (GCN) with temporal convolutional networks (TCN) was proposed. The proposed approach was developed based on an AMESim-Simulink co-simulation framework combined with real-vehicle experimental data. Constrained by the actual structural topology of the tractor powertrain, the engine, transmission system, and vehicle longitudinal states were abstracted into a dynamic state graph. GCN was employed to capture the structural coupling relationships among dynamic states, while TCN was utilized to model the temporal evolution characteristics of longitudinal dynamics, enabling joint spatiotemporal feature learning. The proposed model was experimentally validated under multiple typical field operation scenarios. A condition-weighted grey relational analysis method was further introduced to uniformly quantify prediction errors in both transient and steady-state operating conditions. Experimental results indicated that the GCN-TCN hybrid model consistently achieved high prediction accuracy and robustness across different operating conditions, with an overall multi-condition prediction accuracy of 97.693%. Notably, the proposed method demonstrated superior performance under load transients and non-steady-state conditions. The results can provide a reliable foundation for longitudinal dynamics. This modeling support was essential for optimizing powertrain control strategies and conducting overall performance analysis for high-horsepower tractors.