Rapid discharge failure prediction model for solar charged lithium-ion batteries

Abstract

Lithium-ion batteries are continually being deployed in many appliances. This is due to their high energy density and cost effectiveness. Most of these have been around for years in portable devices such as mobile phones. With the onset of smartphones, there is an ever increasing need to have batteries with superior performance. This can be viewed from the context of the need for fast charging and an ability to support a fully multitasked smartphone. Lithium-ion batteries have become the defacto battery type in many of these and similar applications due to their inherent characteristics. They have found use in not just mobile phones but also in innovative products designed to light homes as well provide for mobile phone charging in rural Africa. These products include a battery pack of Lithium-ion batteries cells charged by solar panels. There are a number of challenges facing the companies dealing with such products. There is a need to provide a superior product while at the same time ensure efficiency in the production line so as to bring down costs. All these need to be done while maintaining the elusive customer loyalty. One of the major issues faced is accelerated degradation which cannot be noticed using conventional approaches. Currently the main mode of triage for failure is visualization of graphs from data collected from the sensors attached to the batteries and observing for irregularities in the charge and discharging patterns. Existing literature talks about models used on linear data for forecasting in various fields of research. It also proposes an approach to predict battery life in batteries used on various applications such as hybrid electric vehicles. The proposed method will take advantage of predictive analytics in time series analysis to predict failure based on data from the batteries. Data from the batteries spanning 30 days was used to generate gradients of daily charging gradients. These were used as the training data with a binary class of faulty and good. We are able to train a model using the nearest neighbor algorithm to obtain over 80% accuracy with only a sample of 200 batteries data.