Multi-head learning models for power consumption prediction of unmanned ground vehicles

Predicting the power performance of unmanned vehicles is an emerging topic, and has gained widespread acceptance. However, there are inevitable tradeoffs between feature diversity and interference, and between feature engineering and prediction efficiency. This paper proposes Ensemble Multi-Head learning (EMH) and 2-Stage Multi-Head learning (2SMH) models for power consumption prediction. These models obtain multi-head learning to reduce feature interference and multi-task learning to enhance prediction. Experimental results show that the proposed 2SMH significantly outperforms the benchmark, with a maximum Mean Absolute Error (MAE) of 30.4%. The 2SMH also exhibits significant improvement in few-shot and zero-shot prediction, demonstrating its outstanding generality and robustness. Moreover, the developed homogeneous transfer mechanisms show MAE reductions of 18.2% and 13.7%, respectively. In summary, the 2SMH achieves excellent MAEs and demonstrates remarkable transferability and predictability.