Adaptive DDPG energy management strategy for hybrid electric vehicles based on online robust driving condition identification

Energy management under variable driving conditions is a critical aspect for improving the fuel efficiency of hybrid electric vehicles (HEVs). Addressing the issue that insufficient real-time accuracy and robustness in recognizing complex and variable driving conditions hinders the overall energy-saving performance of HEVs, this paper proposes an online robust driving condition identification method integrating a Temporal Convolutional Network (TCN) and a Kernel Extreme Learning Machine (KELM). Base on this basis, constructs an adaptive Deep Deterministic Policy Gradient (TCN-KELM-DDPG) energy management framework for HEVs. To balance the real-time requirement and robustness of driving-condition identification, this study adopts a cascaded TCN–KELM architecture: the TCN extracts temporal features from historical and current vehicle-speed sequences and produces short-horizon representations, which are then fed into the KELM to complete condition classification via Gaussian-kernel mapping into a high-dimensional space. Based on the results of robust condition identification, an adaptive DDPG is applied to address the continuous and complex action space under different driving conditions, enabling dynamic optimization of torque distribution in the HEV powertrain and improving the global energy efficiency of HEVs in complex driving scenarios. Hardware-in-the-loop experimental results based on typical standard driving cycles and real-world road tests demonstrate the effectiveness of the proposed TCN-KELM-DDPG strategy in improving condition identification performance and reducing the overall energy consumption cost of HEVs. • A short-term driving condition prediction method based on TCN is proposed. • A robust condition identification model based on KELM is developed. • An adaptive DDPG energy management strategy incorporating TCN-KELM condition feedback is constructed.