Digital twin-based prediction of battery parameters from limited initial data using an optimised time-series multi-layer perceptron

Digital twin (DT) technology is increasingly leveraged for real-time monitoring and predictive modelling of battery systems. However, existing machine learning (ML)-based approaches for battery parameter estimation often rely on large historical datasets, limiting their applicability during the early stages of operation. In this paper, we propose a novel hybrid framework that enables accurate battery parameter prediction and early anomaly detection with minimal initial data. The approach integrates a time-series multi-layer perceptron (TS-MLP) model optimised by a genetic algorithm (GA) to dynamically select input intervals that minimise prediction error. To enhance anomaly detection, an optimised one-class support vector machine (OC-SVM) is trained using both real and synthetically generated anomalies via Gaussian process regression (GPR), ensuring robust performance from early cycles. The proposed method is validated on NASA’s 18650 lithium-ion battery dataset. Results show that the model accurately predicts next-cycle cell temperature with a minimum R 2 of 0.9875 and a maximum RMSE of 0.34 °C. This framework provides a reliable and data-efficient solution for early-stage battery diagnostics in DT environments.