REMAINING DRIVING RANGE AND TIME ESTIMATION IN E-CLEAVE ELECTRIC BICYCLE BY MACHINE LEARNING METHOD

• Main Author: Dewi Kharisma, Meilisa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/47968
• Institution: Institut Teknologi Bandung

The development of electric bicycles (e-bikes) in Indonesia has been stagnating due to the lack of infrastructure and research on the electric vehicle’s support system. Therefore, users may experience range anxiety and furthermore reserve 30% of the remaining energy, decreasing the performance of e-bikes. An e-bike energy efficacy is accounted by the remaining driving range and time value which are somewhat influenced by the driving profile and the driving condition. This final project goal is to give an estimation on e-bike’s remaining driving range and time by considering the road’s elevation and the driving mode of e-bikes. Machine learning methods is used to address the dynamic relation of the measured variables in hope to improve the performance and reliability of the estimated value. Three steps are done to achieve this goal, which includes, designing and implementing a new battery management system on the Energy Managements’s Laboratorium e-bike (e-cleave), estimating the battery’s state-of-charge, and finally executing the machine learning program to estimate the remaining driving range and time of e-bike. The design and implementation of a new battery management system is done by integrating ESP32 microcontroller and Strava, a smartphone’s application to pinpoint the location and speed of e-bike. Battery’s state-of-charge estimation is done by integrating coulomb counting’s method and recursive least square methods. Lastly, machine learning procedures is done, which includes feature selection, hyperparameter tuning, and model selection between polynomial regression, support vector machine, K-nearest neighbors, and decision tree algorithm. A total of 14,730 data are acquired from a four-cycle attempt and are used for further analysis. First, battery state-of-charge estimation is done by filtering the raw data, estimating the open circuit voltage of battery by recursive least square methods, and comparing the results to the state-of-charge to open-circuit-voltage graph acquired from the coulomb counting method. The estimated state-of-charge have a 2.536% root mean squared error (RMSE) compared to the real value. Next, feature selection is done by analyzing the pearson correlation value between feature and by scikitlearn’s algorithm, SelectKBest. While hyperparameter tuning is done by the GridSearchCV algorithm, the model selection is done by doing cross-validation and comparing the RMSE of estimated to the real value. The result shows that decision tree model with a maximum depth of 12 and minumum sample leaf of 2 gives the best performance, with a RMSE/RMSPE of 0.17 km/2.67% and 139.79 seconds/1.89% in estimating the e-bike’s remaining driving range and time, respectively. This accuracy is also complemented by the robust nature of decision tree’s algorithm and the low-cost computing needed by the model.