A Unified Machine Learning Framework for Remaining Discharge Time Prediction in Electric Vehicle Batteries Using Pseudo-Target Learning

Abstract

Accurate Remaining Discharge Time (RDT) prediction is critical for mitigating range anxiety in Battery Electric Vehicles, particularly under irregular off-road operating conditions. However, existing approaches rely heavily on complete time-series data, which is often unavailable in industrial environments. This work proposes a unified machine learning framework for joint estimation of State of Charge (SoC), State of Health (SoH), and Remaining Useful Life (RUL) using heterogeneous battery datasets. A key contribution is a domain-informed pseudo-target generation methodology that reconstructs training labels from incomplete and non-time-series industrial data. The framework integrates XGBoost for SoC estimation, a Kalman Filter-based Neural Network for robust SoH tracking, and Gaussian Process Regression for RUL prediction with uncertainty quantification. Experimental evaluation on NASA PCoE, CALCE, and LG 18650 datasets demonstrates strong performance, achieving SoC MAE of 0.89213 and SoH MAE of 0.00022. While RUL prediction from snapshot data remains challenging $\left(R^{2}=0.24\right)$, the proposed approach provides uncertainty-aware estimates suitable for safety-critical applications. The proposed system enables scalable battery health monitoring in real-world industrial settings, particularly where high-resolution telemetry is unavailable.