State-of-charge estimation hybrid method for lithium-ion batteries using BiGRU and AM co-modified Seq2Seq network and H-infinity filter

Abstract

Accurate battery state of charge (SOC) estimation can provide guarantee for safety and guide the use and maintenance of power battery. A novel SOC estimation hybrid method which combines deep learning network with filter optimization is proposed. Firstly, a sequence-to-sequence (Seq2Seq) neural network is used to fit the nonlinear relationship between SOC and measured signals. Then, the H-infinity (H∞) filter is used to reduce the noise of neural network. This method avoids the complex mechanism analysis process by establishing a high accurate battery data driven model and greatly improve the estimation accuracy by combining with H∞ filter. The advantage of proposed Seq2Seq network lies in that the model has a front and rear direction information of entire input sequence and larger receptive field by applying the bidirectional gated recurrent units (BiGRU) and the attention mechanism (AM). The proposed Seq2Seq network reduces the MAXE of estimation results by 4.55 % and 2.59 %. The estimation accuracy of the proposed hybrid algorithm is verified with the BJDST test set at 5°C–25 °C. The results show that the MAE, RMSE and MAXE of SOC estimation errors for the hybrid algorithm are less than 0.35 %, 0.4 % and about 0.75 %, respectively.