Date of Award

January 2023

Document Type


Degree Name

Master of Science (MS)


Electrical Engineering

First Advisor

Prakash Ranganathan


As the Electric Vehicles (EVs) market continues to expand, ensuring the access to charging stations remains a significant concern. This work focuses on addressing multiple challenges related to EV charging behavior and Vehicle-to-Grid (V2G) services. Firstly, it focuses on accurate minute-ahead (20 minute \& 30 minute intervals) load forecasts for an EV charging station by using four years of historical data, from 2018-2021. This data is recorded from a university campus garage charging station. Machine Learning (ML) models such as Seasonal Auto-Regressive Integrated Moving Average (SARIMA), Random Forest (RF), and Neural Networks (NN) are employed for load forecasts in terms of Kilowatt hour (kWh) delivered from 54 charging stations. Preliminary results indicate that RF method performed better compared to other ML approaches, achieving a average Mean Absolute Error (MAE) of 7.26 on historical weekdays data.

Secondly, it focuses on estimating the probability of aggregated available capacity of users for V2G connections, which could be sold back to the grid through V2G system. To achieve this, an Idle Time (IT) parameter was tracked from the time spent by the EV users at the charging station after being fully charged. ML classification methods such as Logistic Regression (LR) and Linear Support Vector Classifier (SVC) were employed to estimate the IT variable. The SVC model performed better in estimating IT variable with an accuracy of 85% over LR 81%.

This work also analyzes the aggregated excess kWh available from the charging stations for V2G services, which offer benefits to both EV owners through incentives and the grid by balancing the load. ML models, including Support Vector Regressor (SVR), Gradient Boosting Regressor (GBR), Long-Short Term Memory (LSTM), and Random Forest (RF), are employed. LSTM performs better for this prediction problem with a Mean Absolute Percentage Error (MAPE) of 3.12, and RF as second best with lowest 3.59, when considering historical data on weekdays.

Furthermore, this work estimated the number of users available for V2G services corresponding to 15\% and 30\% of excess kWh, by using ML classification models such as Decision Tree (DT) and K Nearest Neighbor (KNN). Among these models, DT performed better, with highest 89% and 84% accuracy respectively.

This work also investigated the impact of the COVID-19 pandemic on EV users' charging behavior. This study analyzes the behavior modelled as before, after, and during COVID-19, employing data visualization using K-means and hierarchical clustering methods to identify common charging pattern with connection and disconnection time of the vehicles. K-means clustering proves to be more effective in all three scenarios modeled with a high silhouette index. Furthermore, prediction of collective charging session duration is achieved using ML Models, RF and XgBoost which achieved a MAPE of 14.6% and 15.1% respectively.