Authors
Xuetian Shi, Mert Alagözlü, Muhammet USLU @ Università della Svizzera italiana (USI)
Spacial thanks to Prof. Illia Horenko, Edoardo Vecchi and Steffen Albrecht for their guidance and interest in our questions!

Battery_SoC_Estimation

Nowadays most of the BEVs manufacturers use Li-Ion based batteries which takes the biggest portion of vehicle architecture. Those Electronic components need to be controlled, diagnosed and monitored by some sophisticated systems so called Battery Management Systems (BMS). Our research on this topic showed that prognostics is one of the most important function of BMS and we found out that some of scientific works proposed that Random/Stochastic Methods have promising results from such as Random Forests comparison to other Machine Learning models.

Here we implemented several Stochastic Methods and Machine Learning algorithms on Li-ion battery dataset from NASA for comparison.

Results

	COARSE TREE	MEDIUM TREE	FINE TREE	BAGGED TREE	BOOSTED TREE	LINEAR REGRESSION	LINEAR REGRESSION STEPWISE	LSVM	QSVM	CSVM	RANDOM FOREST	NEURAL NETWORKS
RMSE	0.0728	0.0260	0.0168	0.0176	0.0674	0.0231	0.0166	0.0241	0.0232	0.0253	0.0188	0.0707
Training Time (sec)	0.613	0.629	0.643	1.584	1.089	4.464	8.725	8.484	8.306	7.952	6.743	5.60
Prediction Speed	19000	22000	18000	4900	4900	2100	2300	9500	9400	9100	6400	6250

Future Work

As the results of our current work showed some Stochastic methods outperform on Battery dataset; we would like to try different methods such as eSPA+, LSTM on different data sets and/or augment current dataset from NASA. Since the application can be useful for edge computing, for instance Electronic Control Unit (ECU) of a vehicle, we're also interested in machine learning model efficiency in limited resource applications (memory, energy). Finally; in safety critical applicaitons the explainability is key we would like to compare the Stochastic Methods in terms of accuracy reproducibility, trustfullability.