Optimizing electric vehicle driving range prediction using deep learning: A deep neural network (DNN) approach
Shahid A. Hasib, Muhammad Majid Gulzar, Adnan Shakoor, Salman Habib, Ali Faisal Murtaza
Abstract
The rapid growth in the popularity of Electric Vehicles (EVs) requires accurate driving range predictions to minimize range anxiety and optimize trip planning, especially in real-world driving conditions where diverse factors affect range. This study addresses the challenges of EV range prediction by presenting a novel deep learning technique that uses a Deep Neural Network (DNN) model optimized with the RMSProp optimizer. This approach leverages a unique real-world dataset that reflects varied driving environments, leading to superior performance. The model achieves an R2 score of 0.99, a Mean Absolute Percentage Error (MAPE) of 2.01%, a Mean Absolute Error (MAE) of 6.81 km, and a Root Mean Squared Error (RMSE) of 9.32 km. These results significantly outperform conventional machine learning techniques like support vector machines and linear regression, demonstrating the practicality and reliability of the proposed model for reducing EV range anxiety and improving trip planning in real-world scenarios. This enhancement supports the broader adoption of EVs, ultimately contributing to a sustainable transportation ecosystem • Initially, an Artificial Neural Network (ANN) model is implemented to predict the EV range. This model is trained and tested using a real-world publicly available dataset, incorporating various influencing factors such as EV weight, battery performance, top speed, acceleration, first charge, and efficiency. • Additionally, different optimizers are implemented into the ANN model to make a comparison of their impact on prediction performance. Optimizers play an important role in the model's parameter tuning, and a comprehensive comparison is conducted to analyze the effectiveness of each optimizer in improving the errors of range prediction. • Finally, many Machine Learning models are implemented on the same dataset for further evaluation. Those models are rigorously evaluated and compared to provide a comprehensive understanding of their respective capabilities in predicting EV range.