Battery Voltage Prediction Using Neural Networks

Abstract

The battery voltage prediction is critical to model predictive controls for the safe and efficient operation of battery systems. This paper presents a comprehensive study using a long-short-term-memory-based method to predict the battery voltage with past voltage and forecasted current and SOC information. Unlike prior art using many-to-one architecture, a many-to-many architecture was used with test data representing three temperatures. Battery-controller-accessible inputs were also selected. Further, the effectiveness of normalization for voltage prediction was investigated. The results show the temperature has no noticeable impact on the prediction accuracy. The lowest RMSE obtained from the 0 °C case is 0.0997. With having both inputs and output already on a similar scale, applying data normalization didn't provide any consistent accuracy improvement across the three selected temperatures.