Battery Management System in Electric Vehicles Using Deep Learning: A Hybrid LSTM-CNN Framework for Enhanced State Estimation and Predictive Maintenance

Electric vehicles (EVs) have emerged as a key pillar of sustainable mobility, driven by the need to reduce carbon emissions from transportation and address climate change more broadly.With the global shift toward sustainable mobility, EVs have become a cornerstone of decarbonized transportation, as the need to lower carbon emissions from transportation and tackle climate change more broadly grows. The Battery Management System (BMS) is at the heart of EV performance, safety and durability, and the accuracy of these key battery states directly affects vehicle range, reliability and user confidence. Most traditional BMS methods such as equivalent-circuit model, extended Kalman filter (EKF), and heuristic algorithms are not effective in tracking the temperature-dependent, history-dependent and highly nonlinear degradation process of modern Lithium-ion batteries in real driving conditions. This paper introduces a novel hybrid deep learning approach which combines both Convolutional Neural Network (CNN) for learning spatial features from multi-dimensional voltage-current-temperature (V-I-T) matrices and Long Short-Term Memory (LSTM) network with attention mechanism for capturing long-range temporal dependency of battery aging trajectories. The architecture utilizes the multi task learning method to predict both State of Charge (SoC), State of Health (SoH), and Remaining Useful Life (RUL). Edge optimized deployment of TensorRT/ONNX supports real-time inference for resource constrained microcontrollers in the automotive space. Under UDDS driving cycles and US06 driving cycles, the state-of-the-art performance is demonstrated by obtaining an error reduction of 71% under RMSE and MAE criteria when estimating state of charge (SoC) on NASA Ames Prognostics Center and CALCE benchmark datasets, and by maintaining an RMSE below 1.15% for SoH prediction across 1,150+ cycles, with RUL predictions showing an accuracy of +-18.7 cycles. With 38 ms inference latency and 0.9 mJ energy consumption, Edge deployment meets tough automotive real-time and thermal requirements. A proposed framework brings science into the real world, with a pathway to future deployable intelligence for predictive, adaptive and safety-critical battery management in next-generation EVs.