Prediction of individual cell performance in a long-string lead/acid peak-shaving battery: application of artificial neural networks

This work represents the culmination of several years of study of an operating large energy storage battery with the purpose of determining if computerized pattern recognition of maintenance data (and/or available fabrication data) could be used for the early detection of poorly performing cells. Also investigated was the possible identification of cells with predicted high performance. Previous studies using k-nearest neighbor pattern recognition have been augmented with the investigation of artificial neural network analysis. Both methods have achieved practical levels of prediction, but the neural network prediction results are somewhat better. It was possible to select 70% of the high-performing cells, without any false selections from the low-performing cells; it was possible to identify nearly 96% of the poor-performance cells, with none of the high-performance cells mis-selected. These results suggest the feasibility of the routine application of neural networks for performance prediction as part of a maintenance strategy for long-string energy storage systems.