Unsupervised data-preprocessing for Long Short-Term Memory based battery model under electric vehicle operation

The high voltage battery is the most valuable component inside an electric vehicle (EV). To safe time and costs during EV development, simulations with digital twins of automotive batteries can help to efficiently reduce and replace expensive laboratory testing. Data-driven methods, such as Long Short-Term Memory (LSTM) neural networks, show great potential in battery modeling. To properly learn battery behavior, LSTMs require a large variety of time series data. Automotive applications generate sensor data continuously, but cannot process this large amount of data due to limited computational capabilities. Either we have high data availability on-board with insufficient computational power or limited data transfer capacities into the cloud for high computational power. To overcome these limitations, the data must be efficiently reduced. We propose a novel selection approach to create a well balanced data set by compressing very large automotive time series data and use it for neural network based battery electric modeling. The in-vehicle battery data is separated into recurrent load situations by applying an unsupervised pattern discovery algorithm. Selecting only representative load patterns, the data set can be compressed by over 95% without losing corner case information. The compressed data set covers over 80% of all different load situations inside the original data set and is well balanced through all seen battery states. Additionally, the accuracy and performance of the LSTM battery model can be even improved by 16% compared to network training without preprocessed data sets.

Automated summary

To efficiently model batteries in electric vehicles, a large amount of time series data is required but must be compressed due to limited computational capabilities. By separating in-vehicle battery data into recurrent load situations and selecting representative load patterns, the data set can be compressed by over 95%, covering over 80% of all different load situations and maintaining balance through all seen battery states. This compression approach improves the accuracy and performance of LSTM battery models by up to 16% compared to training with unprocessed data sets.