@article{LV2027129151,
title = {Physics-constrained multi-task energy consumption prediction for the safe operation of embodied mobile robots},
journal = {Applied Energy},
volume = {427},
pages = {129151},
year = {2027},
issn = {0306-2619},
doi = {https://doi.org/10.1016/j.apenergy.2026.129151},
url = {https://www.sciencedirect.com/science/article/pii/S0306261926018076},
author = {Zhilong Lv and Shiqi (Shawn) Ou and Shaoqin He and Yi Jiang and Hao Jing and Chen Zhang},
keywords = {Energy prediction, Mobile robot, Operational decision support, Physics-constrained multi-task learning, Task available energy},
abstract = {Reliable task-level energy assessment is essential for sustained mission execution and safe operation of battery-powered mobile robots under dynamic conditions. State of charge (SOC) and fixed-capacity estimates provide limited information about the effective energy reserve under changing electrical, thermal, and locomotion conditions. This study proposes a two-level indicator system comprising task-available energy (TAE) and a predictive relative risk margin. TAE characterizes the energy reserve available for a specific task under current operating conditions, extending conventional SOC-based assessment toward task-conditioned energy availability. An 85-day operational dataset covering five gait modes is segmented into task fragments. The proposed physics-constrained multi-task model combines categorical embeddings, stacked long short-term memory encoding, temporal attention, and current-state fusion to jointly predict TAE, future energy consumption, future minimum voltage, and future maximum battery temperature. On the independent test set, normalized root-mean-square error (NRMSE) values of 1.34%, 1.60%, 0.68%, and 0.59% are obtained for the four targets, respectively. Compared with the strongest energy-related baselines, the NRMSE values for TAE and future energy are reduced by 72.4% and 65.6%, respectively. The predictive relative risk margin integrates energy, voltage, and thermal constraints, achieving a coefficient of determination of 0.993, a risk-level agreement of 94.01%, and a dominant-risk-factor accuracy of 93.30%. This work reframes battery management as a task-aware predictive capability that integrates energy state assessment, operational risk evaluation, and autonomous decision-making, enabling safer and more reliable operation of mobile robots in dynamic environments.}
}