# DLLT: A dual-layer LSTM-transformer model for real-time energy and dynamics prediction in plug-in hybrid electric vehicles

## TL;DR

This paper introduces DLLT, a new model that predicts energy use and driving dynamics in plug-in hybrid vehicles using real-world data and driver behavior.

## Contribution

The novel dual-layer LSTM-Transformer model improves real-time energy prediction and mode identification in PHEVs.

## Key findings

- DLLT achieves 93% accuracy in predicting vehicle operation modes.
- The model outperforms existing models with R2 values of 0.99 for fuel consumption and 0.86 for acceleration.
- DLLT shows strong generalization under unseen driving conditions.

## Abstract

Plug-in Hybrid Electric Vehicles (PHEVs) are increasingly favored for their low emissions and freedom from range anxiety, combining electric efficiency with the extended range of a gasoline engine. While previous research on PHEV energy consumption has predominantly focused on powertrain design and energy management strategies, there is growing recognition of the critical role played by driver behavior in determining real-world energy efficiency. Based on multi-mode vehicle data collected from real-world driving scenarios, we propose a novel dual-layer LSTM-Transformer model, named DLLT, for real-time prediction of energy consumption and driving dynamics in multi-mode PHEVs. The first layer employs an LSTM network to perform mode clustering, while the second layer conducts energy consumption regression using a Transformer model with integrated mode information. This hierarchical architecture enables adaptation to diverse driving and braking modes, significantly enhancing the model’s ability to accurately identify vehicle operation modes and precisely predict energy consumption. To more accurately validate the effectiveness of DLLT in modeling eco-driving behavior for PHEVs, we collected a large amount of multidimensional time-series data from real-world PHEVs. Experimental results demonstrate that the model achieves a 93% accuracy rate in vehicle mode prediction. Under unseen driving conditions, it attains R2 values of 0.99 for fuel consumption, 0.86 for acceleration, and 0.81 for electric power, outperforming existing models across all evaluation metrics. With its high prediction accuracy and robust generalization capability, DLLT shows great potential for applications in PHEV eco-driving behavior analysis, intelligent energy management systems, and future autonomous driving control strategies.