A Data-Driven Framework for Forecasting EV Charging Infrastructure Using Large Language Models, By Rahman Dashti

Research

Title	A Data-Driven Framework for Forecasting EV Charging Infrastructure Using Large Language Models
Type	Presentation
Keywords	Electric Vehicle Charging Stations (EVCS), Large Language Model (LLM), Location Prediction, Machine Learning, K-Nearest Neighbors (KNN), Population Density, Points of Interest (POI)
Researchers	Mohamadreza mansouri (First researcher) , Reza Gharibi (Second researcher) , ali darvishi (Third researcher) , Behnam Ranjber (Fourth researcher) , Rahman Dashti (Fifth researcher)

Abstract

With increasing electric vehicle (EV) adoption, optimal charging station (EVCS) installation is vital for urban planning. Determining suitable locations remains a significant challenge. Previous Dubai studies used traditional machine learning like K-Nearest Neighbors (KNN), Logistic Regression, Neural Networks, and Support Vector Machines. Trained on 80% of the dataset, these achieved up to 89% KNN accuracy, using features like geographic coordinates, population density, Points of Interest (POI), and security cameras. This paper introduces a novel Large Language Model (LLM) approach. Structured data is converted to text; Llama 3 8B is fine-tuned via QLoRA for EVCS classification. Results show substantial accuracy and data efficiency: 98% (60% data) and 97% (50% data). This contrasts significantly with the 89% accuracy from 80% training data in prior methods. Fine-tuned LLMs prove a powerful, highly dataefficient tool for optimizing EVCS distribution and advancing smart city infrastructure.