Predictive modeling of electric vehicle supply chain demand in India with a focus on three-wheeler segments

Abstract

The sudden surge of electric three wheelers (3Ws) in India has posed enormous challenge in demand forecasting’s plannings and coordinating infrastructure. As vehicle registrations rises across states the0re is an immediate need for robust and accurate forecasting models to inform data driven policy interventions. The objectives of the current study is to develop an effective machine learning model which forecast monthly 3W electric vehicle registrations using past registration data while considering temporal pattern seasonality and nonlinear pattern of demand change. To this end a time series dataset from January 2018 to December 2024 were used, and three supervised learning models Decision Tree (DT) Random Forest (RF) and Gradient Boosted Trees (GBT) were applied. Some of key methodology included feature engineering using lagged values and moving average temporal validation and hyperparameter tuning through grid search in RapidMiner. Model performance were assessed in terms of RMSE and R² scores. Among the models tried the best performance was given by the GBT with an RMSE of 1837.58 and an R² of 0.991 which show high predictive accuracy. Its ability to sequentially correct errors and learn from engineered temporal features made it the most suitable model for long-term forecasting in this context. This study demonstrates the effectiveness of machine learning, and in particular boosting techniques, in modeling real-world EV adoption behavior. It offer the reproducible frameworks for a extrapolating similar forecasting approach to other vehicle classes. Future studies can involve the incorporation of policy variables, macroeconomic indicators and geospatial data to enhance model generalization and allow more widespread electric mobility planning initiatives in India.