Identifying the shortest charging duration without compromising cell integrity is still one of the biggest issues facing the Li-ion battery industry. Currently, accurate electrochemical models are used to solve a complex optimal control problem. However, these models are computationally intensive and suffer from observability and identifiability issues. Alternatives include more straightforward algorithms that provide output feedback, but they require trial and error tuning to achieve optimal performance. Further techniques are required to navigate safety constraints adeptly, we present an optimal charging strategy driven by Deep Reinforcement Learning (DRL). Our emphasis lies in the utilization of a policy gradient methodology for managing continuous sets of states with actions, predicated on the presumption of access to comprehensive state measurements derived from the Doyle-Fuller-Newman (DFN) model, but we later relax this assumption and consider limited output measurements as agent observations. Additionally, we demonstrate how the proposed policy can adapt to changing environmental factors. We undertake a comparative analysis of our approach against established methodologies in the scholarly literature, such as the proportional-integral-derivative approach and the reference governor.
Enhancing the Dynamic Control of Fast-Charging Strategies Driven by Reinforcement Learning for Li-Ion Battery
31.07.2024
1473736 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Reinforcement Learning Based Fast Charging of Electric Vehicle Battery Packs
SAE Technical Papers | 2023
|REINFORCEMENT LEARNING FOR CONTINUED LEARNING OF OPTIMAL BATTERY CHARGING
Europäisches Patentamt | 2024
|Bus fast charging station ordered charging method considering battery dynamic charging loss
Europäisches Patentamt | 2024
|