Abstract
Wide use of Lithium-Ion Batteries (LIB) in applications that energy storage is critical has led to the development of battery management systems. 2 main aspects are prognosis and live health monitoring. Firstly, predicting Remaining Useful Life (RUL) allows LIB’s to be replaced early to prevent any downtime of the system. In the era of big data, supervised machine learning using Long Short-Term Memory (LSTM) cells was shown to perform relatively well with low wastage cycles (<3) even with low amounts of data. Selection of features and trimming of training RUL was shown to improve noisy data with limited generalizability. Using a single feature that captures the overall degradation trend, paired with high trimming for training was found to have lowest errors. Batteries with different RUL curves should be trained separately as the network is unable to generalize due to lack of training samples. Trimming was also shown to focus the network towards the End of Life predictions, with relatively low wastage cycles.Secondly, live health monitoring was done by predicting capacity at any given cycle. Feature extraction of voltage, current and temperature of charging cycles are shown to be capable of mapping to capacity. A variety of LSTM frameworks was explored to depict the different ways to utilize historical and current information to predict capacity. A dynamic hybrid LSTM approach shows best results with high accuracy due to a synthesized method to utilize as much information as possible at the current time step; and is highly recommended for LIB live monitoring.