Appositeness of automated machine learning libraries on prediction of energy consumption for electric two-wheelers based on micro-trip approach

The use of conventional machine learning (ML) models in predicting energy consumption (EC) for electric vehicles (EVs) requires significant human effort and time for their selection and tuning. Consequently, automated ML (autoML) libraries may reliably predict the EC with a wide range of ML models and minimal domain knowledge. This study predicts the EC of electric two-wheelers (E2Ws) using the base and ensembled ML models from H2O, PyCaret, and Auto-Sklearn autoML libraries. For EC prediction, the feature variables (vehicle, geographical, and ambient) and target variable (EC/km) are calculated for 1815 micro-trips segmented from 42 driving trips collected at a predefined driving route of Vellore city using E2W. Then, the EC is predicted using selected autoML libraries and their suitability is evaluated based on different performance indices. Subsequently, the feature variables' significance on the target variable is assessed. The results indicate that an ensembled ML model from each autoML library improves the prediction accuracy. Particularly, the PyCaret library's stacked ensembled ML model has better R2, Tweedie and pinball scores of 0.7549, 0.7405 and 0.4634 than other autoML libraries' best ML models. Also, it is found that average acceleration, maximum trip speed and deceleration time have significantly influenced the EC of E2Ws.