Data-driven algorithms, such as the neural network ones, seem very appealing and accurate solutions to estimate the lithium-ion battery’s State of Charge. Their accuracy is strongly related to the amount of data used in their training phase. Therefore, huge experimental campaigns are needed to effectively train the neural network used to State of Charge estimation. The main idea behind this paper is to mitigate this drawback by training the algorithm with synthetic datasets generated from simulations of a model of the battery, instead of experimentally collected data. Two instances of the same Long-Short-Term-Memory neural network architecture designed for battery State of Charge estimation are trained, one with an experimental dataset, and the other with a synthetic one. The two neural network instances are then evaluated with the same test dataset derived from experimental data and their estimation accuracies are compared. Results show that the performances of the two networks are comparable. The experimental trained neural network scored a RMSE of only 0.3% lower than the RMSE of the synthetic trained one. These results suggest the possibility of fruitfully using a synthetic training dataset to speed up and reduce the complexity and cost of the training phase of neural network algorithm for battery state of charge estimation.
Comparison of Lithium-Ion Battery SoC Estimation Accuracy of LSTM Neural Network Trained with Experimental and Synthetic Datasets
Hattouti L. A.;Di Rienzo R.;Nicodemo N.;Verani A.;Baronti F.;Roncella R.;Saletti R.
2024-01-01
Abstract
Data-driven algorithms, such as the neural network ones, seem very appealing and accurate solutions to estimate the lithium-ion battery’s State of Charge. Their accuracy is strongly related to the amount of data used in their training phase. Therefore, huge experimental campaigns are needed to effectively train the neural network used to State of Charge estimation. The main idea behind this paper is to mitigate this drawback by training the algorithm with synthetic datasets generated from simulations of a model of the battery, instead of experimentally collected data. Two instances of the same Long-Short-Term-Memory neural network architecture designed for battery State of Charge estimation are trained, one with an experimental dataset, and the other with a synthetic one. The two neural network instances are then evaluated with the same test dataset derived from experimental data and their estimation accuracies are compared. Results show that the performances of the two networks are comparable. The experimental trained neural network scored a RMSE of only 0.3% lower than the RMSE of the synthetic trained one. These results suggest the possibility of fruitfully using a synthetic training dataset to speed up and reduce the complexity and cost of the training phase of neural network algorithm for battery state of charge estimation.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.