Knowledge transfer-oriented deep neural network framework for estimation and forecasting the state of health of the Lithium-ion batteries

• Published in: Journal of energy storage 2022-09, Vol.53, p.105183, Article 105183
• Main Authors: Maleki, Sajjad, Mahmoudi, Amin, Yazdani, Amirmehdi
• Format: Article
• Language: English
• Subjects: Autoregressive integrated moving average (ARIMA); Deep neural networks; Estimation method; Forecasting; Lithium-ion (Li-ion) battery; State of health; Transfer learning
• ISSN: 2352-152X; 2352-1538
• DOI: 10.1016/j.est.2022.105183

This paper proposes an efficient data-driven framework for estimating and forecasting the state of health (SOH) of Lithium-ion (Li-ion) batteries. The proposed framework is established upon a deep neural network (DNN) model, knowledge transfer asset, and autoregressive integrated moving average (ARIMA) forecasting model. The knowledge transfer property reduces the required data for training the model and hence the approach becomes fast and good fit for forecasting the SOH of Li-ion batteries. Among various possibilities, the most efficient training features are picked by Pearson correlation coefficient and least absolute shrinkage and selection operator (LASSO) regression. To suppress existing noises, Savitzky-Golay filter is applied to the signals. The proposed framework allows to use a limited portion of the dataset (about 25 %) for training phase and guarantees high accuracy (almost 96 %) of estimation according to coefficient of determination. Mean squared error (MSE) of the estimations is 0.00075 which is small enough to trust on results. MSE of the model not only during training via 25 % of data is measured, but also after training by 20 % and 30 % of dataset is calculated as well. Training by 20 % of dataset results in a great downfall in the model performance with a 26.6 % rise in the MSE value. Surprisingly, training the model with 30 % portion of the dataset does not add any noticeable accuracy to the model. This study confirms that the transfer learning property and DNN model combination could achieve a dramatic reduction of the dataset portion for training purpose. •Selecting the proper training features of battery during reference charging/discharging cycles to obtain accurate results.•Presenting a deep learning method which uses only 25% of dataset at training step.•Filtering noises of the dataset using Savitzky-Golay (S-G) filter for increasing the accuracy of estimation..•Presenting a hybrid framework that not only estimates the SOH but also can forecast the SOH during a few coming days.•Incorporating cross-validation to mitigate the uncertainties associated with the deep learning model calculations.