Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

State-Of-Charge Estimation Including Uncertainty Analysis

Combining the Information from Charging Profiles and Voltage Relaxation During Pauses Using Neural Networks
: Augustin, Beatrix
: Schmid, Ute; Lehmann, Thomas

Bamberg, 2020, XI, 81 pp.
Bamberg, Univ., Master Thesis, 2020
Master Thesis
Fraunhofer IVI ()
electromobility; battery monitoring; neural networks; state of charge (SOC); uncertainty

The electrification of transport is on the rise to fight the climate crises. Therefore, the fuel tank is replaced by batteries as energy storage device. In order to plan routes the remaining capacity in the battery is necessary but challenging to derive. The simple method of Coulomb Counting (CC, integrating the current over time) suffers from increasing inaccuracy. Other methods like the open circuit voltage (OCV) by comparison, are more accurate but need a long resting period up to several hours until they produce reliable estimates. Advanced methods like Kalman filters or neural networks are used to approximate the State-of-Charge (SoC). However, the literature focuses on cars by using data simulating typical driving conditions. In this thesis we target the use case of buses in public transport. Different from cars, buses are on the road for a long period of time with only multiple short stops. For our approach we use a neural network. In addition to other SoC-estimations we also consider the uncertainty due to the measuring error. We analyse the influence of the relaxation and pre-pause measurements through two separate models with a reduced number of features. The full model performed best, especially at the beginning of the pause it outperforms model with features of the voltage decrease during the pause. In a further analysis we examine the uncertainty of the model using a network with a Bayesian output layer. This network predicts the SoC with less accuracy and with a high uncertainty of the estimate. Our proposed model reduces the necessary time for an OCV comparison from 30 minutes to 1 to 2 minutes. We present proof-of-concept implementation for updating the CC during pauses. Furthermore, our approach is able to predict reliable SoC values of real world data while only being trained with data produced on a test bench.