Battery management system; Gated recurrent unit; Lithium-ion battery; Long-short term memory; Online state prediction; Recurrent neural networks; State of health

Online state of health {SOH) prediction of lithium-ion batteries remains a very important problem in assessing the safety and reliability of batterypowered systems. Deep learning techniques based on recurrent neural networks with memory, such as the long short-term memory {LSTM) and gated recurrent unit {GRU), have very promising advantages, when compared to other SOH estimation algorithms. This work addresses the battery SOH prediction based on GRU. A complete BMS is presented along with the internal structure and configuration parameters. The neural network was highly optimized by adaptive moment estimation {Adam) algorithm. Experimental data show very good estimation results for different temperature values, not only at room value. Comparisons performed against other relevant estimation methods highlight the performance of the recursive neural network algorithms such as GRU and LSTM, with the exception ofthe battery regeneration points. Compared to LSTM, the GRU algorithm gives slightly higher estimation errors, but within similar prediction error range, while needing significantly fewer parameters {about 25% fewer), thus making it a very suitable candidate for embedded implementations.