Abstract
Estimating the long-term degradation of lithium-ion batteries is a fundamental requirement for optimizing fleet maintenance, managing warranty risks, improving user experience, and ensuring reliable energy storage in modern electric vehicles (EVs). Although recent deep learning approaches have achieved promising state-of-health (SoH) prediction performance, most studies estimate future battery degradation only under observed operating conditions and provide limited insight into the long-term battery health implications of alternative future driving behaviors. This paper presents a persona-driven framework that integrates long-term operational simulation, battery aging modeling, and a direct multi-horizon convolutional neural network (CNN)–long short-term memory (LSTM) hybrid network structure for long-term SoH forecasting and paired behavioral counterfactual analysis. Representative driver personas are used to generate diverse multi-year operational scenarios by considering driving behavior, vehicle operating modes, trip characteristics, and seasonal ambient temperature variations. A paired evaluation strategy is then adopted, in which identical first-year operating histories are preserved while only subsequent driving behavior is modified, enabling a direct assessment of the long-term battery health impact of behavioral changes. A matched Day-365 benchmark evaluating 13 forecasting and ablation configurations with three random seeds revealed that the CNN–LSTM structure achieved mean absolute errors of 0.177 at day 730 and 0.238 at day 1095, expressed as SoH percentage points. Random forest produced the lowest corresponding errors of 0.163 and 0.077 percentage points. Removing persona encoding increased the CNN–LSTM structure’s errors to 1.053 and 1.896 percentage points, identifying operational context as the most influential tested input. In a separate ten-run persona-level robustness evaluation, the MAE of the CNN–LSTM structure ranged from 0.29 to 1.05 SoH percentage points. The paired scenario analysis produced Year-3 differences of 1.84–5.39 percentage points, all of which remain conditional on the adopted synthetic aging and scenario assumptions. Together, these evaluations demonstrate long-horizon forecasting and controlled comparison of alternative usage scenarios within the adopted simulation framework.

This publication has 22 references indexed in Scilit: