Real-world analysis of machine learning and Pi-model techniques for SoC prediction in urban electric buses

Accurate estimation of the State of Charge (SoC) of electric vehicle (EV) batteries is important for efficient energy management, battery safety, and reliable range prediction. This is particularly challenging for urban electric buses, where battery operation varies continuously with changing driving and operating conditions. Conventional model-based methods can face difficulties in such conditions because of parameter variations and limited adaptability to real-world operation. In this study, six months of real-world operational data from PMI electric buses running on the Meerut–Sardhana route in India are used to evaluate different approaches for SoC estimation. A Battery Equivalent Circuit (BEC) Pi-model is developed and its performance is compared with six machine learning approaches: Random Forest, XG Boost, Light GBM, Support Vector Regression, Linear Regression, and a hybrid K-Means-assisted regression approach. The models are evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), coefficient of determination (R 2 ), and correlation coefficient. The results show that the Pi-model and the ensemble-based ML models provide accurate SoC estimates under real-world operating conditions. Among the methods studied, the Pi-model provides the best overall performance while also offering better physical interpretability and a closer connection to the actual battery behaviour. The study also identifies challenges related to auxiliary load estimation and the practical implementation of SoC estimation in real-time Battery Management Systems (BMS). The results demonstrate the potential of combining physics-based and data-driven methods to develop reliable SoC estimation techniques for electric bus applications.