11 February 2026 in SSRN Electronic Journal
Accurate prediction of battery state of charge (SOC) and energy consumption rate (ECR) is crucial for effective energy management in electric vehicles, yet remains challenging due to their strong coupling and uncertainty under real-world driving conditions. This study proposes a multi-task deep learning framework to jointly predict SOC and ECR using real-world operational data collected from ten electric vehicles over a twelve-week period in Xi’an. Model epistemic uncertainty is quantified via Monte Carlo (MC) methods, enabling the construction of high-confidence prediction intervals. Experimental results demonstrate that the proposed approach reduces the root mean square error (RMSE) of SOC and ECR predictions by 43.2% and 42.3%, respectively, compared with conventional baseline models. Compared with an independent prediction scheme, the joint multi-task framework further reduces inference time by 42.9–51.1% and lowers memory usage by approximately 41%, indicating improved computational efficiency. Moreover, the expected calibration error (ECE) is reduced by approximately 60% relative to the independent scheme, reflecting enhanced uncertainty calibration. These results indicate that the proposed framework provides accurate, efficient, and reliable joint state estimation for risk-aware electric vehicle energy management.
1 January 2026 in IEEE Transactions on Intelligent Transportation Systems