4 présentations

• 
  15h45 - 16h10

  Développement d'un modèle ML pour la surveillance de la santé des batteries de véhicules électriques basé sur des données open source

  • Arnaud ABIAG, prés., Polytechnique Montréal
  • Ashkan Amirnia, Polytechnique Montréal
  • Ahmad shahnejat-bushehri, Polytechnique Montréal
  • samira keivanpour, Polytechnique Montréal

  La surveillance précise d’état de santé (SOH) des batteries des véhicules électriques (VE) est cruciale pour garantir la sécurité, optimiser les performances et prolonger la durée de vie de la batterie. Ce travail propose le développement d'un modèle d'apprentissage automatique (ML) pour estimer le SOH des batteries de VE en utilisant des données open source. En exploitant des ensembles de données accessibles au public contenant des données de fonctionnement des batteries dans diverses conditions, nous visons à construire un modèle robuste et précis capable de suivre la dégradation de la batterie au fil du temps. Ce modèle utilisera des techniques de ML telles que CatBoost, Random Forest, LSTM, SVM et RNN, pour analyser les données de tension, de courant, de température et d'autres paramètres pertinents afin de prédire le SOH. La validation du modèle sera effectuée à l'aide de données open source indépendantes, et son potentiel pour des applications de surveillance en temps réel sera exploré. Ce projet contribuera à l'avancement des méthodes de gestion de la santé des batteries de VE en utilisant des ressources de données accessibles.

• 
  16h10 - 16h35

  Sustainable EV Charging with Battery Energy Storage System: A Game-Theoretic Approach using Cooperative Reinforcement Learning

  • Mehdi Mehryar, Islamic Azad University
  • Arian Shah Kamrani, prés., Mr.
  • Samira Keivanpour, Polytechnique Montréal
  • Hanane Dagdougui, Ecole Polytechnique de Montréal

  Achieving sustainable and profitable operation of multiple electric vehicle charging stations (EVCS) equipped with battery energy storage systems (BESS) is a complex real-time challenge due to factors like dynamically fluctuating grid electricity prices, variable charging demands of users, the longevity of the battery, and the need to maximize profit. A cooperative method can be advantageous for sites with multiple nearby EVCS, but a key challenge lies in determining how to distribute incomes equitably. This work presents a real-time multi-objective function model designed to maximize profit, user satisfaction, and battery longevity, solved using a deep Q-learning-based cooperative reinforcement learning (CRL) approach, modeled as a cooperative game with a transferable utility method. This study examines and analyzes CRL in the context of multiple EVCS located at a single site using the Shapley value. All coalitions based on cooperative game theory (CGT) are thoroughly investigated to understand their formation, stability, and potential benefits. A comprehensive sensitivity analysis was conducted by varying the coefficients of the objective function’s components. The insights gained from this analysis, combined with strategic cooperation, led to a significant increase in profit potential. Finally, the model’s performance on a three EVCS scenario demonstrated substantial improvements, with the owners experiencing net profit increases of 30.94%, 58.75%, and 71.74%, respectively, due to strategic cooperation. Additionally,
  the longevity of the batteries was extended from 11 years to 14 years, and queue
  times were effectively reduced to zero.

• 
  16h35 - 17h00

  Electric Vehicle Optimization in Dynamic Wireless Charging Lanes Using Deep Reinforcement Learning

  • Ahmed ramzi Houalef, prés., UBE
  • Florian Delavernhe, UBE
  • Sidi-Mohammed Senouci, UBE
  • El-Hassane Aglzim, UBE

  As electric vehicles (EVs) become more common, new challenges appear in managing their energy use, charging time, and travel efficiency. Dynamic Wireless Charging Lanes (DWCLs) offer a helpful solution by letting EVs charge while driving, which reduces the need for regular charging stops and saves time. However, it's still difficult to figure out the best times and speeds for a vehicle to enter or leave a DWCL, while also driving in an efficient way.

  In this work, we use a Deep Q-Network (DQN), a type of reinforcement learning, to control both lane changes and speed on roads with multiple lanes. The goal is to help the vehicle make smart choices that save energy, reduce travel time, and get the most out of wireless charging. The system also takes into account real-world factors like road shape, accurate energy use, and how well the car lines up with the charging coils. We use the CARLA simulator to test and train our model in a realistic environment. The system can decide when to change lanes, adjust speed, and charge, all on its own. Our results show that the method improves charging performance, avoids unnecessary lane changes, and supports smooth and energy-efficient driving in real time.

• 
  17h00 - 17h25

  Reinforcement Learning for Electric Vehicle Charging Scheduling: A Survey and Future Directions

  • Shamim Mahmoudzadeh Vaziri, prés.,
  • Samira Keivanpour, Polytechnique Montréal
  • Martin Trépanier, Polytechnique Montréal, CIRRELT

  As the number of electric vehicles (EVs) rapidly increases, efficient charging scheduling becomes critical. This study addresses the challenges of dynamic EV charging, including long charging times, infrastructure limitations, and inherent uncertainties, by focusing on reinforcement learning (RL) approaches. We survey existing literature on the application of RL to optimize EV charging station assignment, considering factors such as EV demand and station availability. Our analysis highlights the advantages of RL in handling the dynamic and complex nature of the EV charging scheduling problem, demonstrating superior performance compared to traditional methods. Furthermore, we identify key research gaps and propose future directions for advancing RL-based EV charging scheduling, including the development of robust, scalable, and real-time capable algorithms.