Research On On-line SOC Estimation Of Lead-acid Battery Based On Data-driven And FPGA Implementation

More than 160 years have passed since the invention of lead-acid battery.Although various new batteries have been invented in succession,lead-acid battery still occupy a significant share in the field of battery due to their rich raw materials,mature technology,low price,high resource recycling rate,high voltage stability,convenient and simple maintenance and other characteristics.SOC is an important parameter when the battery management system manages the battery.Although many methods have been proposed,it is still a challenge to accurately estimate SOC due to many factors affecting battery SOC.In this paper,the SOC of lead-acid battery is estimated by deep learning technology,and the CNN-GRU model combining one-dimensional Convolutional Neural Network and Gated Recurrent Unit is trained.The model is tested and evaluated on the constant power discharge and constant current discharge data sets in this paper.The mean value of estimation error under various working conditions is counted.The root mean square error is 0.67%,the average absolute error is 0.54%,and the maximum error is 1.64%.The research focus of this paper is how to apply the SOC estimation model based on deep learning technology in practice.A limiting factor of the practical application of deep learning technology is that it has high requirements for hardware performance,and it is difficult to meet the requirements of real-time when running in the equipment with limited performance.FPGA has become a research hotspot of accelerated deep learning model in recent years because of its inherent parallel computing ability and hardware reconfigurable characteristics.In this paper,FPGA is selected as the final implementation platform of SOC estimation model.Aiming at the SOC estimation model CNN-GRU,a variety of basic calculation modules including matrix multiplication,matrix transpose,sigmoid activation function,tanh activation function and relu activation function are realized by using Hardware Description Language Verilog HDL.Based on these,the basic deep learning network structure,convolution layer,gated recurrent unit and full connection layer are realized.Finally,the CNN-GRU model is implemented in Xilinx ZYNQ7020 FPGA in the embedded field.The average estimation error under various working conditions is 2.44% of the root mean square error,2.14% of the average absolute error and 4.61% of the maximum error.When the running clock frequency is 156 MHz,it takes 1.8ms to execute one reasoning,which is greatly improved compared with the execution time of 30.6ms of the model deployed in the ZYNQ7020 FPGA PS.It can meet the real-time requirements of SOC estimation in lead-acid battery management system.