A Comparative Machine Learning-based State-of-Charge Estimation of Lithium-Ion Batteries using OCV-Corrected Reference Modeling and Multi-Domain Feature Engineering
Jul 2026· 2026 International Conference on Intelligent and Sustainable AI Systems (ICOSAAS)· pp. 363-370· 0 citations· 15 references
Abstract
State of Charge (SoC) estimation is an indispensable feature in Battery Management Systems (BMS) and is an important function to be implemented for Electric Vehicles (EVs), Renewable Energy Storage (RES) and Portable electronics. For the nonlinearity of lithium-ion Batteries, the current model-based ones generally suffer from uncertain parameters and lose accuracy under various operating conditions. This paper discusses four traditional machine learning algorithms: Extreme Gradient Boosting (XGBoost), Random Forest (RF), Extra Trees (ET), Support Vector Regression (SVR) and explores them in the context of a common evaluation methodology with regard to SoC estimation. In order to make a fair comparison between the machine learning algorithms, a set of common methodology, which involves preprocessing, feature engineering, splitting into train and test set, and hyperparameters tuning are evaluated in all these machine learning techniques. Various performance parameters have been taken into account for assessing the performance of the actual prediction and to decide the extent to which the selected models are appropriate to be utilized in actual BMS applications.
State of charge (SoC) is a key indicator that represents the proportion of energy remaining in a battery relative to its full capacity. This parameter is continuously tracked by the battery management system (BMS) to maintain safe and efficient operation. Precise determination of SoC is essential, as errors in estimation can negatively affect device performance, accelerate battery wear, and in severe cases lead to system malfunction. As a result, achieving reliable and accurate SoC estimation is a critical aspect that requires careful attention and advanced monitoring techniques. This study aims to design a method for estimating state of charge (SoC) levels through the application of time series forecasting techniques. In this paper five models are used: Deep Neural Networks (DNN), Long Short Time Memory (LSTM), Convolutional Neural Network (CNN), Combination of CNN with LSTM and autoregressive LSTM. For this time-series analysis, battery voltage, battery discharge current, and the core temperature are used as the inputs and the battery SoC used as the output. The investigation covers multi-sample single-output time series forecasting method. The models are evaluated based on performance indices and training time with forecast diagrams graphically represented for each of the models under investigation. Dataset used for this study is based on the public LG® 18650HG2 lithium ion (Li-ion) battery dataset covering series of tests under six operating temperatures. This paper concludes that combining CNN with LSTM is the most accurate model with its lowest mean average error (MAE) on both the validation and test sets. While DNN model has the shortest training time out of the five models.
Basil AlMukhtar, Patrick Denny· 2026 6th International Confe...· 0 citations
An experimental/synthetic hybrid, data-driven FDI framework that leverages supervised machine learning (ML) integrated with an experimentally validated second-order electro-thermal battery model to generate a mixed experimental–synthetic dataset enables fast, real-time diagnosis of complex multi-fault scenarios at the cell or module level in series–parallel LIB pack architectures.
Taha Mohamed Abdelatif Maaradji, Saïd Alem, Emanuele Gravante et al.· Transactions of the Institut...· 0 citations
Accurate prediction of lithium-ion battery state of charge (SOC) is critical for ensuring safety, durability, and efficient energy management in electric vehicles (EVs). This study presents a systematic comparative analysis of supervised machine learning (ML) models for SOC prediction using the publicly available SiCWell battery dataset. Five regression-based algorithms—Generalized Linear Model (GLM), Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), Distributed Random Forest (DRF), and Extremely Randomized Trees (XRT)—are implemented within the H2O framework. The analysis considers six physically measurable battery features derived from voltage, resistance, charge throughput, and current direction. Model performance is evaluated using multiple regression metrics, including MAE, MSE, RMSE, RMSLE, residual deviance, and R², under a consistent train–test evaluation protocol. Results indicate that GLM achieves low prediction errors across both training and testing datasets, suggesting that the selected feature set exhibits a strong linear relationship with SOC within the evaluated operating range. However, ensemble models demonstrate higher variance and sensitivity to hyperparameter configurations, leading to reduced generalization performance. Rather than asserting a universally superior model, this study highlights how dataset characteristics, feature linearity, and model complexity influence SOC prediction accuracy. The findings provide insights into selecting appropriate ML models for battery management systems while emphasizing the importance of dataset transparency, interpretability, and generalization analysis.
M. S. Reddy, M. Monisha· Scientific Reports· 0 citations
Accurate State of Charge (SoC) estimation is essential for safe and efficient lithium-ion battery operation in electric mobility. Although machine learning methods achieve high predictive capability, many studies rely on laboratory cycling data and overlook deployment constraints such as latency and embedded hardware. This paper presents a deployment-oriented evaluation of data-driven SoC estimation using large-scale operational data from a battery electric bus. The dataset includes more than 4.3 million field measurements of current, voltage, temperature, and SoC collected under realistic driving conditions. Five models are compared: Random Forest, LightGBM, XGBoost, Temporal Convolutional Networks (TCN), and Gated Recurrent Unit (GRU) networks. Models are evaluated across sampling intervals from 30 s to 210 s and multiple training-data fractions, considering accuracy, training time, and inference latency. Results show that temporal resolution strongly affects the accuracy-efficiency trade-off. The best configuration, LightGBM at 180 s using 75 % of training data, achieved a test MAE of 6.338, RMSE of 8.495, and median inference latency of 1.10 ms. Compared with the Random Forest baseline at 30 s, it reduced MAE by 34.4 %, RMSE by 35.6 %, and latency by 97.6 %, supporting efficient real-time battery management system deployment.
J. Orellana-Iñiguez, W. L. Gallo, Madon C. de Almeida· Eletrônica de Potência· 0 citations
Online battery state-of-health (SoH) assessment is paramount to improving battery electric vehicles (BEVs) competitiveness. By continuously acquiring data from the on-board battery, it enables to implement strategies aiming to reduce degradation rate and enhance BEVs safety. This work presents the application and validation of two novel patented methodologies for lithium-ion battery SoH estimation, evaluated under realistic BEVs fast-charging operating profiles and benchmarked against experimental measurements. The proposed methods, i.e., the “OCV–SoC-based method” relying on the open-circuit voltage (OCV)–state-of-charge (SoC) characteristic curve and the “discrete wavelet transform (DWT)-based method” applying DWT analysis to in operando voltage signals, are compared with the widely adopted reference technique incremental capacity analysis (ICA). Performances of all applied methods are assessed with reference to experimental capacity fading data obtained from a dedicated cycling aging campaign performed on commercial nickel manganese cobalt (NMC) lithium-ion cells. Both proposed methods exhibit a high accuracy (0.63% and 1.37% root-mean-square-error values for “DWT-based method” and “OCV–SoC-based method,” respectively) comparable with the ICA one (0.98%). Pros and cons of the proposed methods are discussed, together with future activities needed to extend their application domain.
D. Pelosi, C. Scarpelli, F. Quilici et al.· IEEE Open Journal of Industr...· 0 citations
State of charge (SoC) estimation for an electric vehicle (EV) is critical for range estimation, preventing overcharging or undercharging, energy management, and battery lifespan. The current studies are typically based on single-instance learning, focus on a specific battery chemistry or a single vehicle type, rely on controlled Lab conditions, and often lack explainability mechanisms. To overcome all these limitations, this paper proposes a more practically applicable and explainable SoC estimation framework that jointly addresses vehicle diversity (bus and passenger EVs), battery chemistry variability (Nickel Cobalt Manganese and Lithium Iron Phosphate), collective battery dynamics, and real-world on-road operational uncertainties within a unified learning architecture. This study introduces MIL-LGBM, a specialized method that integrates Multiple Instance Learning with Light Gradient Boosting Machine to effectively capture the complex nonlinear behavior of battery SoC while maintaining a low computational footprint. An experimental study conducted on a real-world dataset demonstrated that the proposed framework achieved a 0.799 MAE for passenger vehicles across a wide range of on-road driving conditions.