Using machine learning for the classification of the remaining useful cycles in Lithium-ion batteries

In order to keep up with the increasing focus on renewable energy, the demand for new battery technology and peripherals has likewise increased greatly. Given the relatively slow rate of change of new battery chemistry and technology, it is the peripherals to the batteries that are often relied upon to provide this necessary increase in performance. The 18650 battery with Lithium-Ion internal chemistry is one of the most widely used batteries and is depended upon in many industries to provide power portability and storage. Using an extensive freely available dataset compromising of the charge cycles of 121 18650 batteries, this paper evaluates multiple algorithms’ effectiveness at predicting the remaining useful cycles of a battery from a single discharge curve. Upon evaluation of the algorithms, ’Weighted K Nearest Neighbours’ was shown to be the most accurate model and was further improved to ensure that the maximum accuracy was acquired. Finally, a user interface was created to allow for the demonstration of a potential use case for the model. This model and user interface show the potential for easy testing of batteries to determine the number of remaining useful cycles. This makes the possibility of repurposing or extending the initial purpose of these batteries much greater, which is preferable from both an economic standpoint and an ecological one