Machine learning identification of Electric Vehicles from charging session data
Federico Ferretti, Antonio De Paola
Abstract
Alternating Current (AC) charging is currently the most cost-effective and widely adopted solution for charging of Electric Vehicles (EVs). However, the existing AC charging infrastructure generally exhibits limited communication capabilities with the connected EVs, as information about the vehicle can only be collected through external logging systems that operate independently of the charger itself. A straightforward and interoperable method for extracting information from charging vehicles (e.g., vehicle model, battery capacity, and State of Charge) could significantly enhance the implementation of advanced smart charging strategies, unlocking the flexibility of connected EVs, enabling cost reductions and supporting the provision of ancillary services to the grid. This article implements a novel machine-learning approach to estimate relevant information on AC charging vehicles in a real-world experimental setting designed and implemented by the authors. The proposed approach does not require any hardware adjustment and is capable of predicting several features of the connected EVs (e.g., brand, model, year, battery capacity, End-of-Charge status) by exclusively considering their charging profile in response to specific prescribed current setpoints. Possible applications of the model range from the design of smart charging facilities capable of identifying regular users and forecasting their charging patterns to the real-time estimation of the aggregate flexibility of connected EVs, an essential component in vehicle-to-grid (V2G) applications. Extensive practical demonstrations based on experimental data are provided to validate the identification procedure. An example of flexibility envelope estimation of charging EVs is also included to demonstrate the potential applications of the proposed method for ancillary services provision. • Electric vehicles are a key flexibility resource for future power system operation. • Brand and model of charging EVs can be identified with machine learning techniques. • EV identification requires only standard measurements of their charging session. • Experimental validation shows accurate and robust predictions. • EV identification can support ancillary services provision by charging stations.