Lithium Battery SOC Estimation Based On Neural Networ

The battery management system(BMS)is an important bridge connecting electric vehicles and lithium batteries.The state of charge(SOC)of lithium batteries is a key parameter in BMS and plays a role that cannot be ignored in the safe and stable operation of lithium batteries.However,the value of SOC cannot be obtained by direct measurement.Therefore,accurate estimation of the SOC of lithium batteries has become an urgent problem to be solved.This paper mainly studies the SOC estimation methods of lithium batteries based bidirectional long short-term memory(Bi LSTM),bidirectional gated recurrent unit(Bi GRU),and regularized extreme learning machine(RELM).The research content can be summarized as follows:First,in order to collect the discharge data of lithium batteries under different working conditions,the lithium battery charge and discharge test platform based on Xinwei battery tester is used.Based on this platform,the lithium battery was discharged under different driving cycles,including the urban dynamometer driving schedule(UDDS)and the unified cycle driving schedule(UCDS).The terminal voltage and current of the battery were collected in real-time during the discharge tests and determined as the inputs of the neural network.Second,in order to accurately estimate the SOC of lithium batteries,an adaptive momentum(Adam)algorithm based on the hybrid model SG-Bi LSTM,which is constructed based on Bi LSTM neural network and Savitzky-Golay filter,is proposed for the optimization of network parameters.In this hybrid method,Bi LSTM,which can make full use of the battery data at different times is used as the main structure to estimate SOC;Savitzky-Golay filter is used to optimize the output of the network to further improve the estimation accuracy.The estimated SOC after optimization is considered as the final estimated result of the model.Third,in order to accurately estimate the SOC of lithium batteries and reduce model parameters,an Adam algorithm based on Attention mechanism and Bi GRU is proposed to estimate SOC.In this hybrid model,Bi GRU,which can make full use of the battery data at different times and has a simpler structure than LSTM with fewer network parameters,is used as the main structure to estimate SOC.The Attention mechanism is introduced to the model to further improve the estimation accuracy.Fourth,in order to accurately estimate the SOC of lithium batteries and reduce the training time,a conjugate gradient(CG)algorithm based on the RELM network is proposed.In this hybrid model,to simplify the network structure and reduce the network parameters,the RELM structure with fixed input weights and hidden layer biases is used to model the SOC,which reduces the computational complexity of the algorithm.The CG algorithm is adopted to optimize the network parameters,which avoids the calculation of the inverse matrix and guarantee fast convergence.Fifth,in order to improve the intelligence and the generalization ability of the model,a RELM,which is jointly optimized by the spectral Fletcher-Reeves(SFR)algorithm and the beetle antennae search(BAS)algorithm,is proposed.In this hybrid model,the RELM network is used as the main structure to estimate the SOC;in order to achieve the best generalization ability of RELM,the BAS algorithm is used to optimize the regularization coefficient;in order to further optimize the training process of the model,the SFR algorithm is used to calculate the network parameters.